Add Marathi

README.md

@@ -1,10 +1,10 @@
 ---
 title: Vakyansh Marathi TTS
-emoji: 🏃
+emoji: 🦀
 colorFrom: green
 colorTo: yellow
 sdk: gradio
-sdk_version: 2.9.1
+sdk_version: 2.8.13
 app_file: app.py
 pinned: false
 license: apache-2.0

app.py

@@ -0,0 +1,28 @@
+import os
+os.system('wget -q https://storage.googleapis.com/vakyansh-open-models/tts/marathi/mr-IN/female_voice_0/glow.zip && unzip -q glow.zip -d ttsv/checkpoints/female')
+os.system('wget -q https://storage.googleapis.com/vakyansh-open-models/tts/marathi/mr-IN/female_voice_0/hifi.zip && unzip -q hifi.zip -d ttsv/checkpoints/female')
+os.system('rm glow.zip && rm hifi.zip')
+# os.system('wget -q https://storage.googleapis.com/vakyansh-open-models/tts/hindi/hi-IN/male_voice_1/glow.zip && unzip -q glow.zip -d ttsv/checkpoints/male')
+# os.system('wget -q https://storage.googleapis.com/vakyansh-open-models/tts/hindi/hi-IN/male_voice_1/hifi.zip && unzip -q hifi.zip -d ttsv/checkpoints/male')
+os.system('wget -q https://storage.googleapis.com/vakyansh-open-models/translit_models.zip -P ttsv/checkpoints/ && unzip -q ttsv/checkpoints/translit_models.zip -d ttsv/checkpoints/')
+
+
+for path, subdirs, files in os.walk('ttsv/checkpoints/'):
+    print(subdirs)
+    for name in files:
+        print(os.path.join(path, name))
+
+from ttsv.utils.inference.run_gradio import *
+from argparse import Namespace
+
+#os.system('python ttsv/utils/inference/run_gradio.py -a ttsv/checkpoints/glow/male -v ttsv/checkpoints/hifi/male -d cpu -L hi')
+
+
+args = {
+    'acoustic':'/home/user/app/ttsv/checkpoints/female/glow_ckp',
+    'vocoder':'/home/user/app/ttsv/checkpoints/female/hifi_ckp',
+    'device':'cpu',
+    'lang':'mr'
+}
+
+build_gradio(Namespace(**args))

packages.txt

@@ -0,0 +1 @@
+libsndfile1

requirements.txt

@@ -0,0 +1,19 @@
+Cython==0.29.24
+layers==0.1.5
+librosa==0.8.1
+matplotlib==3.3.4
+numpy==1.21.0
+scipy==1.5.4
+tensorboardX==2.4
+tensorboard==2.7.0
+tqdm==4.62.3
+fastapi==0.70.0
+uvicorn==0.15.0
+gradio==2.5.2
+wavio==0.0.4
+mosestokenizer==1.2.1
+indic-nlp-library==0.81
+inflect==5.3.0
+Unidecode==1.3.2
+torch
+pydload

ttsv/.gitignore

@@ -0,0 +1,132 @@
+# Byte-compiled / optimized / DLL files
+.DS_Store
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+.idea/

ttsv/LICENSE.md

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Open-Speech-EkStep
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

ttsv/README.md

@@ -0,0 +1,220 @@
+# vakyansh-tts
+Text to Speech for Indic languages
+
+## 1. Installation and Setup for training
+
+Clone repo
+Note : for multspeaker glow-tts training use branch [multispeaker](https://github.com/Open-Speech-EkStep/vakyansh-tts/tree/multispeaker)
+```
+git clone https://github.com/Open-Speech-EkStep/vakyansh-tts
+```
+Build conda virtual environment
+```
+cd ./vakyansh-tts
+conda create --name <env_name> python=3.7
+conda activate <env_name>
+pip install -r requirements.txt
+```
+Install [apex](https://github.com/NVIDIA/apex); commit: 37cdaf4 for Mixed-precision training
+
+Note : used only for glow-tts
+```
+cd ..
+git clone https://github.com/NVIDIA/apex
+cd apex
+git checkout 37cdaf4
+pip install -v --disable-pip-version-check --no-cache-dir ./
+cd ../vakyansh-tts
+```
+Build Monotonic Alignment Search Code (Cython)
+
+Note : used only for glow-tts
+```
+bash install.sh
+```
+
+## 2. Data Resampling
+
+The data format should have a folder containing all the .wav files for glow-tts and a text file containing filenames with their sentences.
+
+Directory structure: 
+
+langauge_folder_name
+```
+language_folder_name
+|-- ./wav/*.wav
+|-- ./text_file_name.txt
+```
+The format for text_file_name.txt (Text file is only needed for glow-tts training)
+
+```
+( audio1.wav "Sentence1." )
+( audio2.wav "Sentence2." )
+```
+
+To resample the .wav files to 22050 sample rate, change the following parameters in the vakyansh-tts/scripts/data/resample.sh
+
+```
+input_wav_path : absolute path to wav file folder in vakyansh_tts/data/
+output_wav_path : absolute path to vakyansh_tts/data/resampled_wav_folder_name
+output_sample_rate : 22050 (or any other desired sample rate)
+```
+
+To run:  
+```bash
+cd scripts/data/
+bash resample.sh
+```
+
+
+## 3. Spectogram Training (glow-tts)
+
+### 3.1 Data Preparation
+
+
+To prepare the data edit the vakyansh-tts/scripts/glow/prepare_data.sh file and change the following parameters
+```
+input_text_path : absolute path to vakyansh_tts/data/text_file_name.txt
+input_wav_path : absolute path to vakyansh_tts/data/resampled_wav_folder_name
+gender : female or male voice
+```
+To run:  
+```bash
+cd scripts/glow/
+bash prepare_data.sh
+```
+### 3.2 Training glow-tts
+
+To start the spectogram-training edit the vakyansh-tts/scripts/glow/train_glow.sh file and change the following parameter:
+```
+gender : female or male voice
+```
+Make sure that the gender is same as that of the prepare_data.sh file
+
+To start the training, run:  
+```bash
+cd scripts/glow/
+bash train_glow.sh
+```
+## 4. Vocoder Training (hifi-gan)
+
+### 4.1 Data Preparation
+
+To prepare the data edit the vakyansh-tts/scripts/hifi/prepare_data.sh file and change the following parameters
+```
+input_wav_path : absolute path to vakyansh_tts/data/resampled_wav_folder_name
+gender : female or male voice
+```
+To run:  
+```bash
+cd scripts/hifi/
+bash prepare_data.sh
+```
+### 4.2 Training hifi-gan
+
+To start the spectogram-training edit the vakyansh-tts/scripts/hifi/train_hifi.sh file and change the following parameter:
+```
+gender : female or male voice
+```
+Make sure that the gender is same as that of the prepare_data.sh file
+
+To start the training, run:  
+```bash
+cd scripts/hifi/
+bash train_hifi.sh
+```
+
+## 5. Inference
+
+### 5.1 Using Gradio 
+
+To use the gradio link edit the following parameters in the vakyansh-tts/scripts/inference/gradio.sh file:
+```
+gender : female or male voice
+device : cpu or cuda
+lang : langauge code
+```
+
+To run:  
+```bash
+cd scripts/inference/
+bash gradio.sh
+```
+### 5.2 Using fast API 
+To use the fast api link edit the parameters in the vakyansh-tts/scripts/inference/api.sh file similar to section 5.1
+
+To run:  
+```bash
+cd scripts/inference/
+bash api.sh
+```
+
+### 5.3 Direct Inference using text  
+To infer, edit the parameters in the vakyansh-tts/scripts/inference/infer.sh file similar to section 5.1 and set the text to the text variable
+
+To run:  
+```bash
+cd scripts/inference/
+bash infer.sh
+```
+
+To configure other parameters there is a version that runs the advanced inference as well. Additional Parameters:
+```
+noise_scale : can vary from 0 to 1 for noise factor
+length_scale : can vary from 0 to 2 for changing the speed of the generated audio 
+transliteration : whether to switch on/off transliteration. 1: ON, 0: OFF
+number_conversion : whether to switch on/off number to words conversion. 1: ON, 0: OFF
+split_sentences : whether to switch on/off splitting of sentences. 1: ON, 0: OFF
+```
+To run:
+```
+cd scripts/inference/
+bash advanced_infer.sh
+```
+
+### 5.4 Installation of tts_infer package
+
+In tts_infer package, we currently have two components:
+    
+    1. Transliteration (AI4bharat's open sourced models) (Languages supported: {'hi', 'gu', 'mr', 'bn', 'te', 'ta', 'kn', 'pa', 'gom', 'mai', 'ml', 'sd', 'si', 'ur'} )
+    
+    2. Num to Word (Languages supported: {'en', 'hi', 'gu', 'mr', 'bn', 'te', 'ta', 'kn', 'or', 'pa'} )
+```
+git clone https://github.com/Open-Speech-EkStep/vakyansh-tts
+cd vakyansh-tts
+bash install.sh
+python setup.py bdist_wheel
+pip install -e .
+cd tts_infer
+gsutil -m cp -r gs://vakyaansh-open-models/translit_models .
+```
+
+Usage: Refer to example file in tts_infer/
+```
+from tts_infer.tts import TextToMel, MelToWav
+from tts_infer.transliterate import XlitEngine
+from tts_infer.num_to_word_on_sent import normalize_nums
+
+import re
+from scipy.io.wavfile import write
+
+text_to_mel = TextToMel(glow_model_dir='/path/to/glow-tts/checkpoint/dir', device='cuda')
+mel_to_wav = MelToWav(hifi_model_dir='/path/to/hifi/checkpoint/dir', device='cuda')
+
+def translit(text, lang):
+    reg = re.compile(r'[a-zA-Z]')
+    engine = XlitEngine(lang)
+    words = [engine.translit_word(word, topk=1)[lang][0] if reg.match(word) else word for word in text.split()]
+    updated_sent = ' '.join(words)
+    return updated_sent
+    
+def run_tts(text, lang):
+    text = text.replace('।', '.') # only for hindi models
+    text_num_to_word = normalize_nums(text, lang) # converting numbers to words in lang
+    text_num_to_word_and_transliterated = translit(text_num_to_word, lang) # transliterating english words to lang
+    
+    mel = text_to_mel.generate_mel(text_num_to_word_and_transliterated)
+    audio, sr = mel_to_wav.generate_wav(mel)
+    write(filename='temp.wav', rate=sr, data=audio) # for saving wav file, if needed
+    return (sr, audio)
+```

ttsv/config/glow/base.json

@@ -0,0 +1,54 @@
+{
+  "train": {
+    "use_cuda": true,
+    "log_interval": 20,
+    "seed": 1234,
+    "epochs": 10000,
+    "learning_rate": 1e0,
+    "betas": [0.9, 0.98],
+    "eps": 1e-9,
+    "warmup_steps": 4000,
+    "scheduler": "noam",
+    "batch_size": 16,
+    "ddi": true,
+    "fp16_run": true,
+    "save_epoch": 1
+  },
+  "data": {
+    "load_mel_from_disk": false,
+    "training_files":"../data/training/train.txt", 
+    "validation_files":"../data/training/valid.txt",
+    "chars":"",
+    "punc":"",
+    "text_cleaners":["basic_indic_cleaners"],
+    "max_wav_value": 32768.0,
+    "sampling_rate": 22050,
+    "filter_length": 1024,
+    "hop_length": 256,
+    "win_length": 1024,
+    "n_mel_channels": 80,
+    "mel_fmin": 80.0,
+    "mel_fmax": 7600.0,
+    "add_noise": true
+  },
+  "model": {
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "filter_channels_dp": 256,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "n_blocks_dec": 12,
+    "n_layers_enc": 6,
+    "n_heads": 2,
+    "p_dropout_dec": 0.05,
+    "dilation_rate": 1,
+    "kernel_size_dec": 5,
+    "n_block_layers": 4,
+    "n_sqz": 2,
+    "prenet": true,
+    "mean_only": true,
+    "hidden_channels_enc": 192,
+    "hidden_channels_dec": 192,
+    "window_size": 4
+  }
+}

ttsv/config/glow/base_blank.json

@@ -0,0 +1,55 @@
+{
+  "train": {
+    "use_cuda": true,
+    "log_interval": 20,
+    "seed": 1234,
+    "epochs": 10000,
+    "learning_rate": 1e0,
+    "betas": [0.9, 0.98],
+    "eps": 1e-9,
+    "warmup_steps": 4000,
+    "scheduler": "noam",
+    "batch_size": 16,
+    "ddi": true,
+    "fp16_run": true,
+    "save_epoch": 1
+  },
+  "data": {
+    "load_mel_from_disk": false,
+    "training_files":"../data/training/train.txt", 
+    "validation_files":"../data/training/valid.txt",
+    "chars":"",
+    "punc":"",
+    "text_cleaners":["basic_indic_cleaners"],
+    "max_wav_value": 32768.0,
+    "sampling_rate": 22050,
+    "filter_length": 1024,
+    "hop_length": 256,
+    "win_length": 1024,
+    "n_mel_channels": 80,
+    "mel_fmin": 80.0,
+    "mel_fmax": 7600.0,
+    "add_noise": true,
+    "add_blank": true
+  },
+  "model": {
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "filter_channels_dp": 256,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "n_blocks_dec": 12,
+    "n_layers_enc": 6,
+    "n_heads": 2,
+    "p_dropout_dec": 0.05,
+    "dilation_rate": 1,
+    "kernel_size_dec": 5,
+    "n_block_layers": 4,
+    "n_sqz": 2,
+    "prenet": true,
+    "mean_only": true,
+    "hidden_channels_enc": 192,
+    "hidden_channels_dec": 192,
+    "window_size": 4
+  }
+}

ttsv/config/hifi/config_v1.json

@@ -0,0 +1,37 @@
+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 24,
+    "learning_rate": 0.0002,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.999,
+    "seed": 1234,
+
+    "upsample_rates": [8,8,2,2],
+    "upsample_kernel_sizes": [16,16,4,4],
+    "upsample_initial_channel": 512,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "segment_size": 8192,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 80,
+    "fmax": 7600,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

ttsv/config/hifi/config_v2.json

@@ -0,0 +1,37 @@
+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 24,
+    "learning_rate": 0.0002,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.999,
+    "seed": 1234,
+
+    "upsample_rates": [8,8,2,2],
+    "upsample_kernel_sizes": [16,16,4,4],
+    "upsample_initial_channel": 128,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "segment_size": 8192,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 80,
+    "fmax": 7600,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

ttsv/config/hifi/config_v3.json

@@ -0,0 +1,37 @@
+{
+    "resblock": "2",
+    "num_gpus": 0,
+    "batch_size": 24,
+    "learning_rate": 0.0002,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.999,
+    "seed": 1234,
+
+    "upsample_rates": [8,8,4],
+    "upsample_kernel_sizes": [16,16,8],
+    "upsample_initial_channel": 256,
+    "resblock_kernel_sizes": [3,5,7],
+    "resblock_dilation_sizes": [[1,2], [2,6], [3,12]],
+
+    "segment_size": 8192,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 80,
+    "fmax": 7600,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

ttsv/install.sh

@@ -0,0 +1,6 @@
+cd src/glow_tts/monotonic_align/
+pip install .
+cd ../../../
+
+# torch
+pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

ttsv/scripts/data/duration.sh

@@ -0,0 +1,9 @@
+wav_path='/home/harveen/en/iitm_data/english/wav_22k'
+#######################
+
+dir=$PWD
+parentdir="$(dirname "$dir")"
+parentdir="$(dirname "$parentdir")"
+
+
+python $parentdir/utils/data/duration.py $wav_path

ttsv/scripts/data/resample.sh

@@ -0,0 +1,14 @@
+input_wav_path='/home/harveen/en/iitm_data/english/wav/'
+output_wav_path='/home/harveen/en/iitm_data/english/wav_22k/'
+output_sample_rate=22050
+
+#######################
+
+dir=$PWD
+parentdir="$(dirname "$dir")"
+parentdir="$(dirname "$parentdir")"
+
+mkdir -p $output_wav_path
+python $parentdir/utils/data/resample.py -i $input_wav_path -o $output_wav_path -s $output_sample_rate
+
+python $parentdir/utils/data/duration.py $output_wav_path

ttsv/scripts/glow/prepare_data.sh

@@ -0,0 +1,12 @@
+input_text_path='/home/harveen/en/iitm_data/english/txt.done.data'
+input_wav_path='/home/harveen/en/iitm_data/english/wav_22k'
+gender='male'
+
+
+output_data_path='../../data/glow/'$gender
+
+valid_samples=100
+test_samples=10
+
+mkdir -p $output_data_path
+python ../../utils/glow/prepare_iitm_data_glow_en.py -i $input_text_path -o $output_data_path -w $input_wav_path -v $valid_samples -t $test_samples

ttsv/scripts/glow/train_glow.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+
+gender='male'
+
+config='../../config/glow/'$gender'.json'
+modeldir='../../checkpoints/glow/'$gender
+logdir='../../logs/glow/'$gender
+init=1  # 1 if start from scratch. 0 if start from last checkpoint
+
+
+####################################################
+
+if [[ $init -eq 1 ]]
+then
+  python ../../src/glow_tts/init.py -c $config -m $modeldir -l $logdir
+fi
+python ../../src/glow_tts/train.py -c $config -m $modeldir -l $logdir

ttsv/scripts/hifi/prepare_data.sh

@@ -0,0 +1,10 @@
+input_wav_path='/home/harveen/en/iitm_data/english/wav_22k' #give multiple folders separated by comma(,)
+gender='male'
+
+output_data_path='../../data/hifi/'$gender
+
+valid_samples=100
+test_samples=10
+
+mkdir -p $output_data_path
+python ../../utils/hifi/prepare_iitm_data_hifi.py -i $input_wav_path -v $valid_samples -t $test_samples -d $output_data_path

ttsv/scripts/hifi/train_hifi.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+gender='male'
+
+config='../../config/hifi/config_v1.json'
+modeldir='../../checkpoints/hifi/'$gender
+logdir='../../logs/hifi/'$gender
+
+
+####################################################
+
+
+
+python ../../src/hifi_gan/train.py \
+    --config $config \
+    --input_training_file '../../data/hifi/'$gender'/train.txt' \
+    --input_validation_file '../../data/hifi/'$gender'/valid.txt' \
+    --checkpoint_path $modeldir \
+    --logs_path $logdir \
+    --checkpoint_interval 10000 \
+    --stdout_interval 50

ttsv/scripts/inference/advanced_infer.sh

@@ -0,0 +1,22 @@
+gender='male'
+glowdir='../../checkpoints/glow/'$gender'/'
+hifidir='../../checkpoints/hifi/'$gender'/'
+device='cpu'
+text='Hey mr. I am testing this one. Now on multiple sentences. Just want to see the flow.'
+noise_scale='0.667'
+length_scale='1.0'
+transliteration=1
+number_conversion=1
+split_sentences=1
+lang='en'
+
+
+timestamp=$(date +%s)
+wav='../../results/'$gender'/'
+wav_file=$wav/$timestamp'.wav'
+
+
+mkdir -p $wav
+
+python ../../utils/inference/advanced_tts.py -a $glowdir -v $hifidir -d $device -t "$text" -w $wav_file -L $lang -n $noise_scale -l $length_scale -T $transliteration -N $number_conversion -S $split_sentences
+echo "File saved at: "$wav_file

ttsv/scripts/inference/api.sh

@@ -0,0 +1,8 @@
+gender='male'
+glowdir='../../checkpoints/glow/'$gender'/'
+hifidir='../../checkpoints/hifi/'$gender'/'
+device='cpu'
+lang='en'
+
+
+python ../../utils/inference/api.py -a $glowdir -v $hifidir -d $device -L $lang

ttsv/scripts/inference/gradio.sh

@@ -0,0 +1,8 @@
+gender='male'
+glowdir='../../checkpoints/glow/'$gender'/'
+hifidir='../../checkpoints/hifi/'$gender'/'
+device='cpu'
+lang='en'
+
+
+python ../../utils/inference/run_gradio.py -a $glowdir -v $hifidir -d $device -L $lang

ttsv/scripts/inference/infer.sh

@@ -0,0 +1,15 @@
+gender='male'
+glowdir='../../checkpoints/glow/'$gender'/'
+hifidir='../../checkpoints/hifi/'$gender'/'
+device='cpu'
+text='testing this one'
+
+
+timestamp=$(date +%s)
+wav='../../results/'$gender'/'
+wav_file=$wav/$timestamp'.wav'
+
+
+mkdir -p $wav
+python ../../utils/inference/tts.py -a $glowdir -v $hifidir -d $device -t "$text" -w $wav_file 
+echo "File saved at: "$wav_file

ttsv/setup.py

@@ -0,0 +1,55 @@
+from setuptools import setup, find_packages
+
+with open("README.md", "r") as f:
+    long_description = f.read()
+
+setup(
+    name="vakyansh-tts",
+    version="0.0.5",
+    description="Text to speech for Indic languages",
+    long_description=long_description,
+    long_description_content_type="text/markdown",
+    url="https://github.com/Open-Speech-EkStep/vakyansh-tts.git",
+    keywords="nlp, tts, Indic languages, deep learning, text to speech",
+    # package_dir={'': 'src'},
+    # packages=find_packages(where='src'),
+    packages=["tts_infer"],
+    python_requires=">=3.7, <4",
+    install_requires=[
+        "Cython==0.29.24",
+        "layers==0.1.5",
+        "librosa==0.8.1",
+        "matplotlib==3.3.4",
+        "numpy==1.20.2",
+        "scipy==1.5.4",
+        "tensorboardX==2.4",
+        "tensorboard==2.7.0",
+        "tqdm==4.62.3",
+        "fastapi==0.70.0",
+        "uvicorn==0.15.0",
+        "gradio==2.5.2",
+        "wavio==0.0.4",
+        "pydload==1.0.9",
+        "mosestokenizer==1.2.1",
+        "indic-nlp-library==0.81"
+    ],
+    classifiers=[
+        # How mature is this project? Common values are
+        #   3 - Alpha
+        #   4 - Beta
+        #   5 - Production/Stable
+        "Development Status :: 3 - Alpha",
+        # Indicate who your project is intended for
+        "Intended Audience :: Developers",
+        "Intended Audience :: Education",
+        "Intended Audience :: Science/Research",
+        "Topic :: Scientific/Engineering :: Artificial Intelligence",
+        "Topic :: Text Processing :: Linguistic",
+        # Pick your license as you wish (should match "license" above)
+        "License :: OSI Approved :: MIT License",
+        # Specify the Python versions you support here. In particular, ensure
+        # that you indicate whether you support Python 2, Python 3 or both.
+        "Programming Language :: Python :: 3.7",
+    ],
+    include_package_data=True,
+)

ttsv/src/glow_tts/attentions.py

@@ -0,0 +1,378 @@
+import copy
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import commons
+import modules
+from modules import LayerNorm
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=None,
+        block_length=None,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.block_length = block_length
+
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    window_size=window_size,
+                    p_dropout=p_dropout,
+                    block_length=block_length,
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        for i in range(self.n_layers):
+            x = x * x_mask
+            y = self.attn_layers[i](x, x, attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class CouplingBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+        sigmoid_scale=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = p_dropout
+        self.sigmoid_scale = sigmoid_scale
+
+        start = torch.nn.Conv1d(in_channels // 2, hidden_channels, 1)
+        start = torch.nn.utils.weight_norm(start)
+        self.start = start
+        # Initializing last layer to 0 makes the affine coupling layers
+        # do nothing at first.  It helps to stabilze training.
+        end = torch.nn.Conv1d(hidden_channels, in_channels, 1)
+        end.weight.data.zero_()
+        end.bias.data.zero_()
+        self.end = end
+
+        self.wn = modules.WN(
+            in_channels,
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels,
+            p_dropout,
+        )
+
+    def forward(self, x, x_mask=None, reverse=False, g=None, **kwargs):
+        b, c, t = x.size()
+        if x_mask is None:
+            x_mask = 1
+        x_0, x_1 = x[:, : self.in_channels // 2], x[:, self.in_channels // 2 :]
+
+        x = self.start(x_0) * x_mask
+        x = self.wn(x, x_mask, g)
+        out = self.end(x)
+
+        z_0 = x_0
+        m = out[:, : self.in_channels // 2, :]
+        logs = out[:, self.in_channels // 2 :, :]
+        if self.sigmoid_scale:
+            logs = torch.log(1e-6 + torch.sigmoid(logs + 2))
+
+        if reverse:
+            z_1 = (x_1 - m) * torch.exp(-logs) * x_mask
+            logdet = None
+        else:
+            z_1 = (m + torch.exp(logs) * x_1) * x_mask
+            logdet = torch.sum(logs * x_mask, [1, 2])
+
+        z = torch.cat([z_0, z_1], 1)
+        return z, logdet
+
+    def store_inverse(self):
+        self.wn.remove_weight_norm()
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        window_size=None,
+        heads_share=True,
+        p_dropout=0.0,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super().__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.p_dropout = p_dropout
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels ** -0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        if proximal_init:
+            self.conv_k.weight.data.copy_(self.conv_q.weight.data)
+            self.conv_k.bias.data.copy_(self.conv_q.bias.data)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+
+    def forward(self, x, c, attn_mask=None):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(self, query, key, value, mask=None):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s, t_t = (*key.size(), query.size(2))
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(self.k_channels)
+        if self.window_size is not None:
+            assert (
+                t_s == t_t
+            ), "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(query, key_relative_embeddings)
+            rel_logits = self._relative_position_to_absolute_position(rel_logits)
+            scores_local = rel_logits / math.sqrt(self.k_channels)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores * block_mask + -1e4 * (1 - block_mask)
+        p_attn = F.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        """
+        x: [b, h, l, m]
+        y: [h or 1, m, d]
+        ret: [b, h, l, d]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        """
+        x: [b, h, l, d]
+        y: [h or 1, m, d]
+        ret: [b, h, l, m]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        max_relative_position = 2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = F.pad(
+                relative_embeddings,
+                commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+            )
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        """
+        x: [b, h, l, 2*l-1]
+        ret: [b, h, l, l]
+        """
+        batch, heads, length, _ = x.size()
+        # Concat columns of pad to shift from relative to absolute indexing.
+        x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
+
+        # Concat extra elements so to add up to shape (len+1, 2*len-1).
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])
+        )
+
+        # Reshape and slice out the padded elements.
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        """
+        x: [b, h, l, l]
+        ret: [b, h, l, 2*l-1]
+        """
+        batch, heads, length, _ = x.size()
+        # padd along column
+        x = F.pad(
+            x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])
+        )
+        x_flat = x.view([batch, heads, length ** 2 + length * (length - 1)])
+        # add 0's in the beginning that will skew the elements after reshape
+        x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length):
+        """Bias for self-attention to encourage attention to close positions.
+        Args:
+          length: an integer scalar.
+        Returns:
+          a Tensor with shape [1, 1, length, length]
+        """
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation=None,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+
+        self.conv_1 = nn.Conv1d(
+            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.conv_2 = nn.Conv1d(
+            filter_channels, out_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.drop = nn.Dropout(p_dropout)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(x * x_mask)
+        if self.activation == "gelu":
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        return x * x_mask

ttsv/src/glow_tts/audio_processing.py

@@ -0,0 +1,100 @@
+import torch
+import numpy as np
+from scipy.signal import get_window
+import librosa.util as librosa_util
+
+
+def window_sumsquare(
+    window,
+    n_frames,
+    hop_length=200,
+    win_length=800,
+    n_fft=800,
+    dtype=np.float32,
+    norm=None,
+):
+    """
+    # from librosa 0.6
+    Compute the sum-square envelope of a window function at a given hop length.
+
+    This is used to estimate modulation effects induced by windowing
+    observations in short-time fourier transforms.
+
+    Parameters
+    ----------
+    window : string, tuple, number, callable, or list-like
+        Window specification, as in `get_window`
+
+    n_frames : int > 0
+        The number of analysis frames
+
+    hop_length : int > 0
+        The number of samples to advance between frames
+
+    win_length : [optional]
+        The length of the window function.  By default, this matches `n_fft`.
+
+    n_fft : int > 0
+        The length of each analysis frame.
+
+    dtype : np.dtype
+        The data type of the output
+
+    Returns
+    -------
+    wss : np.ndarray, shape=`(n_fft + hop_length * (n_frames - 1))`
+        The sum-squared envelope of the window function
+    """
+    if win_length is None:
+        win_length = n_fft
+
+    n = n_fft + hop_length * (n_frames - 1)
+    x = np.zeros(n, dtype=dtype)
+
+    # Compute the squared window at the desired length
+    win_sq = get_window(window, win_length, fftbins=True)
+    win_sq = librosa_util.normalize(win_sq, norm=norm) ** 2
+    win_sq = librosa_util.pad_center(win_sq, n_fft)
+
+    # Fill the envelope
+    for i in range(n_frames):
+        sample = i * hop_length
+        x[sample : min(n, sample + n_fft)] += win_sq[: max(0, min(n_fft, n - sample))]
+    return x
+
+
+def griffin_lim(magnitudes, stft_fn, n_iters=30):
+    """
+    PARAMS
+    ------
+    magnitudes: spectrogram magnitudes
+    stft_fn: STFT class with transform (STFT) and inverse (ISTFT) methods
+    """
+
+    angles = np.angle(np.exp(2j * np.pi * np.random.rand(*magnitudes.size())))
+    angles = angles.astype(np.float32)
+    angles = torch.autograd.Variable(torch.from_numpy(angles))
+    signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+
+    for i in range(n_iters):
+        _, angles = stft_fn.transform(signal)
+        signal = stft_fn.inverse(magnitudes, angles).squeeze(1)
+    return signal
+
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return torch.exp(x) / C

ttsv/src/glow_tts/commons.py

@@ -0,0 +1,273 @@
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from librosa.filters import mel as librosa_mel_fn
+from audio_processing import dynamic_range_compression
+from audio_processing import dynamic_range_decompression
+from stft import STFT
+
+
+def intersperse(lst, item):
+    result = [item] * (len(lst) * 2 + 1)
+    result[1::2] = lst
+    return result
+
+
+def mle_loss(z, m, logs, logdet, mask):
+    l = torch.sum(logs) + 0.5 * torch.sum(
+        torch.exp(-2 * logs) * ((z - m) ** 2)
+    )  # neg normal likelihood w/o the constant term
+    l = l - torch.sum(logdet)  # log jacobian determinant
+    l = l / torch.sum(
+        torch.ones_like(z) * mask
+    )  # averaging across batch, channel and time axes
+    l = l + 0.5 * math.log(2 * math.pi)  # add the remaining constant term
+    return l
+
+
+def duration_loss(logw, logw_, lengths):
+    l = torch.sum((logw - logw_) ** 2) / torch.sum(lengths)
+    return l
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+def convert_pad_shape(pad_shape):
+    l = pad_shape[::-1]
+    pad_shape = [item for sublist in l for item in sublist]
+    return pad_shape
+
+
+def shift_1d(x):
+    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def sequence_mask(length, max_length=None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def maximum_path(value, mask, max_neg_val=-np.inf):
+    """Numpy-friendly version. It's about 4 times faster than torch version.
+    value: [b, t_x, t_y]
+    mask: [b, t_x, t_y]
+    """
+    value = value * mask
+
+    device = value.device
+    dtype = value.dtype
+    value = value.cpu().detach().numpy()
+    mask = mask.cpu().detach().numpy().astype(np.bool)
+
+    b, t_x, t_y = value.shape
+    direction = np.zeros(value.shape, dtype=np.int64)
+    v = np.zeros((b, t_x), dtype=np.float32)
+    x_range = np.arange(t_x, dtype=np.float32).reshape(1, -1)
+    for j in range(t_y):
+        v0 = np.pad(v, [[0, 0], [1, 0]], mode="constant", constant_values=max_neg_val)[
+            :, :-1
+        ]
+        v1 = v
+        max_mask = v1 >= v0
+        v_max = np.where(max_mask, v1, v0)
+        direction[:, :, j] = max_mask
+
+        index_mask = x_range <= j
+        v = np.where(index_mask, v_max + value[:, :, j], max_neg_val)
+    direction = np.where(mask, direction, 1)
+
+    path = np.zeros(value.shape, dtype=np.float32)
+    index = mask[:, :, 0].sum(1).astype(np.int64) - 1
+    index_range = np.arange(b)
+    for j in reversed(range(t_y)):
+        path[index_range, index, j] = 1
+        index = index + direction[index_range, index, j] - 1
+    path = path * mask.astype(np.float32)
+    path = torch.from_numpy(path).to(device=device, dtype=dtype)
+    return path
+
+
+def generate_path(duration, mask):
+    """
+    duration: [b, t_x]
+    mask: [b, t_x, t_y]
+    """
+    device = duration.device
+
+    b, t_x, t_y = mask.shape
+    cum_duration = torch.cumsum(duration, 1)
+    path = torch.zeros(b, t_x, t_y, dtype=mask.dtype).to(device=device)
+
+    cum_duration_flat = cum_duration.view(b * t_x)
+    path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
+    path = path.view(b, t_x, t_y)
+    path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+    path = path * mask
+    return path
+
+
+class Adam:
+    def __init__(
+        self,
+        params,
+        scheduler,
+        dim_model,
+        warmup_steps=4000,
+        lr=1e0,
+        betas=(0.9, 0.98),
+        eps=1e-9,
+    ):
+        self.params = params
+        self.scheduler = scheduler
+        self.dim_model = dim_model
+        self.warmup_steps = warmup_steps
+        self.lr = lr
+        self.betas = betas
+        self.eps = eps
+
+        self.step_num = 1
+        self.cur_lr = lr * self._get_lr_scale()
+
+        self._optim = torch.optim.Adam(params, lr=self.cur_lr, betas=betas, eps=eps)
+
+    def _get_lr_scale(self):
+        if self.scheduler == "noam":
+            return np.power(self.dim_model, -0.5) * np.min(
+                [
+                    np.power(self.step_num, -0.5),
+                    self.step_num * np.power(self.warmup_steps, -1.5),
+                ]
+            )
+        else:
+            return 1
+
+    def _update_learning_rate(self):
+        self.step_num += 1
+        if self.scheduler == "noam":
+            self.cur_lr = self.lr * self._get_lr_scale()
+            for param_group in self._optim.param_groups:
+                param_group["lr"] = self.cur_lr
+
+    def get_lr(self):
+        return self.cur_lr
+
+    def step(self):
+        self._optim.step()
+        self._update_learning_rate()
+
+    def zero_grad(self):
+        self._optim.zero_grad()
+
+    def load_state_dict(self, d):
+        self._optim.load_state_dict(d)
+
+    def state_dict(self):
+        return self._optim.state_dict()
+
+
+class TacotronSTFT(nn.Module):
+    def __init__(
+        self,
+        filter_length=1024,
+        hop_length=256,
+        win_length=1024,
+        n_mel_channels=80,
+        sampling_rate=22050,
+        mel_fmin=0.0,
+        mel_fmax=8000.0,
+    ):
+        super(TacotronSTFT, self).__init__()
+        self.n_mel_channels = n_mel_channels
+        self.sampling_rate = sampling_rate
+        self.stft_fn = STFT(filter_length, hop_length, win_length)
+        mel_basis = librosa_mel_fn(
+            sampling_rate, filter_length, n_mel_channels, mel_fmin, mel_fmax
+        )
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+
+    def spectral_normalize(self, magnitudes):
+        output = dynamic_range_compression(magnitudes)
+        return output
+
+    def spectral_de_normalize(self, magnitudes):
+        output = dynamic_range_decompression(magnitudes)
+        return output
+
+    def mel_spectrogram(self, y):
+        """Computes mel-spectrograms from a batch of waves
+        PARAMS
+        ------
+        y: Variable(torch.FloatTensor) with shape (B, T) in range [-1, 1]
+
+        RETURNS
+        -------
+        mel_output: torch.FloatTensor of shape (B, n_mel_channels, T)
+        """
+        assert torch.min(y.data) >= -1
+        assert torch.max(y.data) <= 1
+
+        magnitudes, phases = self.stft_fn.transform(y)
+        magnitudes = magnitudes.data
+        mel_output = torch.matmul(self.mel_basis, magnitudes)
+        mel_output = self.spectral_normalize(mel_output)
+        return mel_output
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+    clip_value = float(clip_value)
+
+    total_norm = 0
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+
+        p.grad.data.clamp_(min=-clip_value, max=clip_value)
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm
+
+
+def squeeze(x, x_mask=None, n_sqz=2):
+    b, c, t = x.size()
+
+    t = (t // n_sqz) * n_sqz
+    x = x[:, :, :t]
+    x_sqz = x.view(b, c, t // n_sqz, n_sqz)
+    x_sqz = x_sqz.permute(0, 3, 1, 2).contiguous().view(b, c * n_sqz, t // n_sqz)
+
+    if x_mask is not None:
+        x_mask = x_mask[:, :, n_sqz - 1 :: n_sqz]
+    else:
+        x_mask = torch.ones(b, 1, t // n_sqz).to(device=x.device, dtype=x.dtype)
+    return x_sqz * x_mask, x_mask
+
+
+def unsqueeze(x, x_mask=None, n_sqz=2):
+    b, c, t = x.size()
+
+    x_unsqz = x.view(b, n_sqz, c // n_sqz, t)
+    x_unsqz = x_unsqz.permute(0, 2, 3, 1).contiguous().view(b, c // n_sqz, t * n_sqz)
+
+    if x_mask is not None:
+        x_mask = x_mask.unsqueeze(-1).repeat(1, 1, 1, n_sqz).view(b, 1, t * n_sqz)
+    else:
+        x_mask = torch.ones(b, 1, t * n_sqz).to(device=x.device, dtype=x.dtype)
+    return x_unsqz * x_mask, x_mask

ttsv/src/glow_tts/data_utils.py

@@ -0,0 +1,274 @@
+import random
+import numpy as np
+import torch
+import torch.utils.data
+
+import commons
+from utils import load_wav_to_torch, load_filepaths_and_text
+from text import text_to_sequence
+
+class TextMelLoader(torch.utils.data.Dataset):
+    """
+    1) loads audio,text pairs
+    2) normalizes text and converts them to sequences of one-hot vectors
+    3) computes mel-spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths_and_text, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+        self.text_cleaners = hparams.text_cleaners
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.load_mel_from_disk = hparams.load_mel_from_disk
+        self.add_noise = hparams.add_noise
+        self.symbols = hparams.punc + hparams.chars
+        self.add_blank = getattr(hparams, "add_blank", False)  # improved version
+        self.stft = commons.TacotronSTFT(
+            hparams.filter_length,
+            hparams.hop_length,
+            hparams.win_length,
+            hparams.n_mel_channels,
+            hparams.sampling_rate,
+            hparams.mel_fmin,
+            hparams.mel_fmax,
+        )
+        random.seed(1234)
+        random.shuffle(self.audiopaths_and_text)
+
+    def get_mel_text_pair(self, audiopath_and_text):
+        # separate filename and text
+        audiopath, text = audiopath_and_text[0], audiopath_and_text[1]
+        text = self.get_text(text)
+        mel = self.get_mel(audiopath)
+        return (text, mel)
+
+    def get_mel(self, filename):
+        if not self.load_mel_from_disk:
+            audio, sampling_rate = load_wav_to_torch(filename)
+            if sampling_rate != self.stft.sampling_rate:
+                raise ValueError(
+                    "{} {} SR doesn't match target {} SR".format(
+                        sampling_rate, self.stft.sampling_rate
+                    )
+                )
+            if self.add_noise:
+                audio = audio + torch.rand_like(audio)
+            audio_norm = audio / self.max_wav_value
+            audio_norm = audio_norm.unsqueeze(0)
+            melspec = self.stft.mel_spectrogram(audio_norm)
+            melspec = torch.squeeze(melspec, 0)
+        else:
+            melspec = torch.from_numpy(np.load(filename))
+            assert (
+                melspec.size(0) == self.stft.n_mel_channels
+            ), "Mel dimension mismatch: given {}, expected {}".format(
+                melspec.size(0), self.stft.n_mel_channels
+            )
+
+        return melspec
+
+    def get_text(self, text):
+        text_norm = text_to_sequence(text, self.symbols, self.text_cleaners)
+        if self.add_blank:
+            text_norm = commons.intersperse(
+                text_norm, len(self.symbols)
+            )  # add a blank token, whose id number is len(symbols)
+        text_norm = torch.IntTensor(text_norm)
+        return text_norm
+
+    def __getitem__(self, index):
+        return self.get_mel_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextMelCollate:
+    """Zero-pads model inputs and targets based on number of frames per step"""
+
+    def __init__(self, n_frames_per_step=1):
+        self.n_frames_per_step = n_frames_per_step
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and mel-spectrogram
+        PARAMS
+        ------
+        batch: [text_normalized, mel_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        input_lengths, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([len(x[0]) for x in batch]), dim=0, descending=True
+        )
+        max_input_len = input_lengths[0]
+
+        text_padded = torch.LongTensor(len(batch), max_input_len)
+        text_padded.zero_()
+        for i in range(len(ids_sorted_decreasing)):
+            text = batch[ids_sorted_decreasing[i]][0]
+            text_padded[i, : text.size(0)] = text
+
+        # Right zero-pad mel-spec
+        num_mels = batch[0][1].size(0)
+        max_target_len = max([x[1].size(1) for x in batch])
+        if max_target_len % self.n_frames_per_step != 0:
+            max_target_len += (
+                self.n_frames_per_step - max_target_len % self.n_frames_per_step
+            )
+            assert max_target_len % self.n_frames_per_step == 0
+
+        # include mel padded
+        mel_padded = torch.FloatTensor(len(batch), num_mels, max_target_len)
+        mel_padded.zero_()
+        output_lengths = torch.LongTensor(len(batch))
+        for i in range(len(ids_sorted_decreasing)):
+            mel = batch[ids_sorted_decreasing[i]][1]
+            mel_padded[i, :, : mel.size(1)] = mel
+            output_lengths[i] = mel.size(1)
+
+        return text_padded, input_lengths, mel_padded, output_lengths
+
+
+"""Multi speaker version"""
+
+
+class TextMelSpeakerLoader(torch.utils.data.Dataset):
+    """
+    1) loads audio, speaker_id, text pairs
+    2) normalizes text and converts them to sequences of one-hot vectors
+    3) computes mel-spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths_sid_text, hparams):
+        self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
+        self.text_cleaners = hparams.text_cleaners
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.load_mel_from_disk = hparams.load_mel_from_disk
+        self.add_noise = hparams.add_noise
+        self.symbols = hparams.punc + hparams.chars
+        self.add_blank = getattr(hparams, "add_blank", False)  # improved version
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 190)
+        self.stft = commons.TacotronSTFT(
+            hparams.filter_length,
+            hparams.hop_length,
+            hparams.win_length,
+            hparams.n_mel_channels,
+            hparams.sampling_rate,
+            hparams.mel_fmin,
+            hparams.mel_fmax,
+        )
+
+        self._filter_text_len()
+        random.seed(1234)
+        random.shuffle(self.audiopaths_sid_text)
+
+    def _filter_text_len(self):
+        audiopaths_sid_text_new = []
+        for audiopath, sid, text in self.audiopaths_sid_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_sid_text_new.append([audiopath, sid, text])
+        self.audiopaths_sid_text = audiopaths_sid_text_new
+
+    def get_mel_text_speaker_pair(self, audiopath_sid_text):
+        # separate filename, speaker_id and text
+        audiopath, sid, text = (
+            audiopath_sid_text[0],
+            audiopath_sid_text[1],
+            audiopath_sid_text[2],
+        )
+        text = self.get_text(text)
+        mel = self.get_mel(audiopath)
+        sid = self.get_sid(sid)
+        return (text, mel, sid)
+
+    def get_mel(self, filename):
+        if not self.load_mel_from_disk:
+            audio, sampling_rate = load_wav_to_torch(filename)
+            if sampling_rate != self.stft.sampling_rate:
+                raise ValueError(
+                    "{} {} SR doesn't match target {} SR".format(
+                        sampling_rate, self.stft.sampling_rate
+                    )
+                )
+            if self.add_noise:
+                audio = audio + torch.rand_like(audio)
+            audio_norm = audio / self.max_wav_value
+            audio_norm = audio_norm.unsqueeze(0)
+            melspec = self.stft.mel_spectrogram(audio_norm)
+            melspec = torch.squeeze(melspec, 0)
+        else:
+            melspec = torch.from_numpy(np.load(filename))
+            assert (
+                melspec.size(0) == self.stft.n_mel_channels
+            ), "Mel dimension mismatch: given {}, expected {}".format(
+                melspec.size(0), self.stft.n_mel_channels
+            )
+
+        return melspec
+
+    def get_text(self, text):
+        text_norm = text_to_sequence(text, self.symbols, self.text_cleaners)
+        if self.add_blank:
+            text_norm = commons.intersperse(
+                text_norm, len(self.symbols)
+            )  # add a blank token, whose id number is len(symbols)
+        text_norm = torch.IntTensor(text_norm)
+        return text_norm
+
+    def get_sid(self, sid):
+        sid = torch.IntTensor([int(sid)])
+        return sid
+
+    def __getitem__(self, index):
+        return self.get_mel_text_speaker_pair(self.audiopaths_sid_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_sid_text)
+
+
+class TextMelSpeakerCollate:
+    """Zero-pads model inputs and targets based on number of frames per step"""
+
+    def __init__(self, n_frames_per_step=1):
+        self.n_frames_per_step = n_frames_per_step
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and mel-spectrogram
+        PARAMS
+        ------
+        batch: [text_normalized, mel_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        input_lengths, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([len(x[0]) for x in batch]), dim=0, descending=True
+        )
+        max_input_len = input_lengths[0]
+
+        text_padded = torch.LongTensor(len(batch), max_input_len)
+        text_padded.zero_()
+        for i in range(len(ids_sorted_decreasing)):
+            text = batch[ids_sorted_decreasing[i]][0]
+            text_padded[i, : text.size(0)] = text
+
+        # Right zero-pad mel-spec
+        num_mels = batch[0][1].size(0)
+        max_target_len = max([x[1].size(1) for x in batch])
+        if max_target_len % self.n_frames_per_step != 0:
+            max_target_len += (
+                self.n_frames_per_step - max_target_len % self.n_frames_per_step
+            )
+            assert max_target_len % self.n_frames_per_step == 0
+
+        # include mel padded & sid
+        mel_padded = torch.FloatTensor(len(batch), num_mels, max_target_len)
+        mel_padded.zero_()
+        output_lengths = torch.LongTensor(len(batch))
+        sid = torch.LongTensor(len(batch))
+        for i in range(len(ids_sorted_decreasing)):
+            mel = batch[ids_sorted_decreasing[i]][1]
+            mel_padded[i, :, : mel.size(1)] = mel
+            output_lengths[i] = mel.size(1)
+            sid[i] = batch[ids_sorted_decreasing[i]][2]
+
+        return text_padded, input_lengths, mel_padded, output_lengths, sid

ttsv/src/glow_tts/generate_mels.py

@@ -0,0 +1,70 @@
+import numpy as np
+import os
+import torch
+import commons
+
+import models
+import utils
+from argparse import ArgumentParser
+from tqdm import tqdm
+from text import text_to_sequence
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("-m", "--model_dir", required=True, type=str)
+    parser.add_argument("-s", "--mels_dir", required=True, type=str)
+    args = parser.parse_args()
+    MODEL_DIR = args.model_dir  # path to model dir
+    SAVE_MELS_DIR = args.mels_dir  # path to save generated mels
+
+    if not os.path.exists(SAVE_MELS_DIR):
+        os.makedirs(SAVE_MELS_DIR)
+
+    hps = utils.get_hparams_from_dir(MODEL_DIR)
+    symbols = list(hps.data.punc) + list(hps.data.chars)
+    checkpoint_path = utils.latest_checkpoint_path(MODEL_DIR)
+    cleaner = hps.data.text_cleaners
+
+    model = models.FlowGenerator(
+        len(symbols) + getattr(hps.data, "add_blank", False),
+        out_channels=hps.data.n_mel_channels,
+        **hps.model
+    ).to("cuda")
+
+    utils.load_checkpoint(checkpoint_path, model)
+    model.decoder.store_inverse()  # do not calcuate jacobians for fast decoding
+    _ = model.eval()
+
+    def get_mel(text, fpath):
+        if getattr(hps.data, "add_blank", False):
+            text_norm = text_to_sequence(text, symbols, cleaner)
+            text_norm = commons.intersperse(text_norm, len(symbols))
+        else:  # If not using "add_blank" option during training, adding spaces at the beginning and the end of utterance improves quality
+            text = " " + text.strip() + " "
+            text_norm = text_to_sequence(text, symbols, cleaner)
+
+        sequence = np.array(text_norm)[None, :]
+
+        x_tst = torch.autograd.Variable(torch.from_numpy(sequence)).cuda().long()
+        x_tst_lengths = torch.tensor([x_tst.shape[1]]).cuda()
+
+        with torch.no_grad():
+            noise_scale = 0.667
+            length_scale = 1.0
+            (y_gen_tst, *_), *_, (attn_gen, *_) = model(
+                x_tst,
+                x_tst_lengths,
+                gen=True,
+                noise_scale=noise_scale,
+                length_scale=length_scale,
+            )
+
+        np.save(os.path.join(SAVE_MELS_DIR, fpath), y_gen_tst.cpu().detach().numpy())
+
+    for f in [hps.data.training_files, hps.data.validation_files]:
+        file_lines = open(f).read().splitlines()
+
+        for line in tqdm(file_lines):
+            fname, text = line.split("|")
+            fname = os.path.basename(fname).replace(".wav", ".npy")
+            get_mel(text, fname)

ttsv/src/glow_tts/hifi/__init__.py

@@ -0,0 +1,5 @@
+from .env import AttrDict
+from .models import Generator
+
+if __name__ == "__main__":
+    pass

ttsv/src/glow_tts/hifi/env.py

@@ -0,0 +1,15 @@
+import os
+import shutil
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+def build_env(config, config_name, path):
+    t_path = os.path.join(path, config_name)
+    if config != t_path:
+        os.makedirs(path, exist_ok=True)
+        shutil.copyfile(config, os.path.join(path, config_name))

ttsv/src/glow_tts/hifi/models.py

@@ -0,0 +1,403 @@
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+from .utils import init_weights, get_padding
+
+LRELU_SLOPE = 0.1
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.h = h
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.h = h
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(init_weights)
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Generator(torch.nn.Module):
+    def __init__(self, h):
+        super(Generator, self).__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        self.conv_pre = weight_norm(
+            Conv1d(80, h.upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = ResBlock1 if h.resblock == "1" else ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        h.upsample_initial_channel // (2 ** i),
+                        h.upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        print("Removing weight norm...")
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self):
+        super(MultiPeriodDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList(
+            [
+                DiscriminatorP(2),
+                DiscriminatorP(3),
+                DiscriminatorP(5),
+                DiscriminatorP(7),
+                DiscriminatorP(11),
+            ]
+        )
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 128, 15, 1, padding=7)),
+                norm_f(Conv1d(128, 128, 41, 2, groups=4, padding=20)),
+                norm_f(Conv1d(128, 256, 41, 2, groups=16, padding=20)),
+                norm_f(Conv1d(256, 512, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(512, 1024, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 1, groups=16, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiScaleDiscriminator(torch.nn.Module):
+    def __init__(self):
+        super(MultiScaleDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList(
+            [
+                DiscriminatorS(use_spectral_norm=True),
+                DiscriminatorS(),
+                DiscriminatorS(),
+            ]
+        )
+        self.meanpools = nn.ModuleList(
+            [AvgPool1d(4, 2, padding=2), AvgPool1d(4, 2, padding=2)]
+        )
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            if i != 0:
+                y = self.meanpools[i - 1](y)
+                y_hat = self.meanpools[i - 1](y_hat)
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+def feature_loss(fmap_r, fmap_g):
+    loss = 0
+    for dr, dg in zip(fmap_r, fmap_g):
+        for rl, gl in zip(dr, dg):
+            loss += torch.mean(torch.abs(rl - gl))
+
+    return loss * 2
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+    loss = 0
+    r_losses = []
+    g_losses = []
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        r_loss = torch.mean((1 - dr) ** 2)
+        g_loss = torch.mean(dg ** 2)
+        loss += r_loss + g_loss
+        r_losses.append(r_loss.item())
+        g_losses.append(g_loss.item())
+
+    return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+    loss = 0
+    gen_losses = []
+    for dg in disc_outputs:
+        l = torch.mean((1 - dg) ** 2)
+        gen_losses.append(l)
+        loss += l
+
+    return loss, gen_losses

ttsv/src/glow_tts/hifi/utils.py

@@ -0,0 +1,57 @@
+import glob
+import os
+import matplotlib
+import torch
+from torch.nn.utils import weight_norm
+
+matplotlib.use("Agg")
+import matplotlib.pylab as plt
+
+
+def plot_spectrogram(spectrogram):
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+
+    fig.canvas.draw()
+    plt.close()
+
+    return fig
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def apply_weight_norm(m):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        weight_norm(m)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict
+
+
+def save_checkpoint(filepath, obj):
+    print("Saving checkpoint to {}".format(filepath))
+    torch.save(obj, filepath)
+    print("Complete.")
+
+
+def scan_checkpoint(cp_dir, prefix):
+    pattern = os.path.join(cp_dir, prefix + "????????")
+    cp_list = glob.glob(pattern)
+    if len(cp_list) == 0:
+        return None
+    return sorted(cp_list)[-1]

ttsv/src/glow_tts/init.py

@@ -0,0 +1,79 @@
+import os
+import json
+import argparse
+import math
+import torch
+from torch import nn, optim
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+
+from data_utils import TextMelLoader, TextMelCollate
+import models
+import commons
+import utils
+
+
+class FlowGenerator_DDI(models.FlowGenerator):
+    """A helper for Data-dependent Initialization"""
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        for f in self.decoder.flows:
+            if getattr(f, "set_ddi", False):
+                f.set_ddi(True)
+
+
+def main():
+    hps = utils.get_hparams()
+    logger = utils.get_logger(hps.log_dir)
+    logger.info(hps)
+    utils.check_git_hash(hps.log_dir)
+
+    torch.manual_seed(hps.train.seed)
+
+    train_dataset = TextMelLoader(hps.data.training_files, hps.data)
+    collate_fn = TextMelCollate(1)
+    train_loader = DataLoader(
+        train_dataset,
+        num_workers=8,
+        shuffle=True,
+        batch_size=hps.train.batch_size,
+        pin_memory=True,
+        drop_last=True,
+        collate_fn=collate_fn,
+    )
+    symbols = hps.data.punc + hps.data.chars
+    generator = FlowGenerator_DDI(
+        len(symbols) + getattr(hps.data, "add_blank", False),
+        out_channels=hps.data.n_mel_channels,
+        **hps.model
+    ).cuda()
+    optimizer_g = commons.Adam(
+        generator.parameters(),
+        scheduler=hps.train.scheduler,
+        dim_model=hps.model.hidden_channels,
+        warmup_steps=hps.train.warmup_steps,
+        lr=hps.train.learning_rate,
+        betas=hps.train.betas,
+        eps=hps.train.eps,
+    )
+
+    generator.train()
+    for batch_idx, (x, x_lengths, y, y_lengths) in enumerate(train_loader):
+        x, x_lengths = x.cuda(), x_lengths.cuda()
+        y, y_lengths = y.cuda(), y_lengths.cuda()
+
+        _ = generator(x, x_lengths, y, y_lengths, gen=False)
+        break
+
+    utils.save_checkpoint(
+        generator,
+        optimizer_g,
+        hps.train.learning_rate,
+        0,
+        os.path.join(hps.model_dir, "ddi_G.pth"),
+    )
+
+
+if __name__ == "__main__":
+    main()

ttsv/src/glow_tts/models.py

@@ -0,0 +1,403 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import modules
+import commons
+import attentions
+import monotonic_align
+
+
+class DurationPredictor(nn.Module):
+    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+
+        self.drop = nn.Dropout(p_dropout)
+        self.conv_1 = nn.Conv1d(
+            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_1 = attentions.LayerNorm(filter_channels)
+        self.conv_2 = nn.Conv1d(
+            filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_2 = attentions.LayerNorm(filter_channels)
+        self.proj = nn.Conv1d(filter_channels, 1, 1)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+        x = self.proj(x * x_mask)
+        return x * x_mask
+
+
+class TextEncoder(nn.Module):
+    def __init__(
+        self,
+        n_vocab,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        filter_channels_dp,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        window_size=None,
+        block_length=None,
+        mean_only=False,
+        prenet=False,
+        gin_channels=0,
+    ):
+
+        super().__init__()
+
+        self.n_vocab = n_vocab
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.filter_channels_dp = filter_channels_dp
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.block_length = block_length
+        self.mean_only = mean_only
+        self.prenet = prenet
+        self.gin_channels = gin_channels
+
+        self.emb = nn.Embedding(n_vocab, hidden_channels)
+        nn.init.normal_(self.emb.weight, 0.0, hidden_channels ** -0.5)
+
+        if prenet:
+            self.pre = modules.ConvReluNorm(
+                hidden_channels,
+                hidden_channels,
+                hidden_channels,
+                kernel_size=5,
+                n_layers=3,
+                p_dropout=0.5,
+            )
+        self.encoder = attentions.Encoder(
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            window_size=window_size,
+            block_length=block_length,
+        )
+
+        self.proj_m = nn.Conv1d(hidden_channels, out_channels, 1)
+        if not mean_only:
+            self.proj_s = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj_w = DurationPredictor(
+            hidden_channels + gin_channels, filter_channels_dp, kernel_size, p_dropout
+        )
+
+    def forward(self, x, x_lengths, g=None):
+        x = self.emb(x) * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+
+        if self.prenet:
+            x = self.pre(x, x_mask)
+        x = self.encoder(x, x_mask)
+
+        if g is not None:
+            g_exp = g.expand(-1, -1, x.size(-1))
+            x_dp = torch.cat([torch.detach(x), g_exp], 1)
+        else:
+            x_dp = torch.detach(x)
+
+        x_m = self.proj_m(x) * x_mask
+        if not self.mean_only:
+            x_logs = self.proj_s(x) * x_mask
+        else:
+            x_logs = torch.zeros_like(x_m)
+
+        logw = self.proj_w(x_dp, x_mask)
+        return x_m, x_logs, logw, x_mask
+
+
+class FlowSpecDecoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_blocks,
+        n_layers,
+        p_dropout=0.0,
+        n_split=4,
+        n_sqz=2,
+        sigmoid_scale=False,
+        gin_channels=0,
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_blocks = n_blocks
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+        self.n_split = n_split
+        self.n_sqz = n_sqz
+        self.sigmoid_scale = sigmoid_scale
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for b in range(n_blocks):
+            self.flows.append(modules.ActNorm(channels=in_channels * n_sqz))
+            self.flows.append(
+                modules.InvConvNear(channels=in_channels * n_sqz, n_split=n_split)
+            )
+            self.flows.append(
+                attentions.CouplingBlock(
+                    in_channels * n_sqz,
+                    hidden_channels,
+                    kernel_size=kernel_size,
+                    dilation_rate=dilation_rate,
+                    n_layers=n_layers,
+                    gin_channels=gin_channels,
+                    p_dropout=p_dropout,
+                    sigmoid_scale=sigmoid_scale,
+                )
+            )
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            flows = self.flows
+            logdet_tot = 0
+        else:
+            flows = reversed(self.flows)
+            logdet_tot = None
+
+        if self.n_sqz > 1:
+            x, x_mask = commons.squeeze(x, x_mask, self.n_sqz)
+        for f in flows:
+            if not reverse:
+                x, logdet = f(x, x_mask, g=g, reverse=reverse)
+                logdet_tot += logdet
+            else:
+                x, logdet = f(x, x_mask, g=g, reverse=reverse)
+        if self.n_sqz > 1:
+            x, x_mask = commons.unsqueeze(x, x_mask, self.n_sqz)
+        return x, logdet_tot
+
+    def store_inverse(self):
+        for f in self.flows:
+            f.store_inverse()
+
+
+class FlowGenerator(nn.Module):
+    def __init__(
+        self,
+        n_vocab,
+        hidden_channels,
+        filter_channels,
+        filter_channels_dp,
+        out_channels,
+        kernel_size=3,
+        n_heads=2,
+        n_layers_enc=6,
+        p_dropout=0.0,
+        n_blocks_dec=12,
+        kernel_size_dec=5,
+        dilation_rate=5,
+        n_block_layers=4,
+        p_dropout_dec=0.0,
+        n_speakers=0,
+        gin_channels=0,
+        n_split=4,
+        n_sqz=1,
+        sigmoid_scale=False,
+        window_size=None,
+        block_length=None,
+        mean_only=False,
+        hidden_channels_enc=None,
+        hidden_channels_dec=None,
+        prenet=False,
+        **kwargs
+    ):
+
+        super().__init__()
+        self.n_vocab = n_vocab
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.filter_channels_dp = filter_channels_dp
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_heads = n_heads
+        self.n_layers_enc = n_layers_enc
+        self.p_dropout = p_dropout
+        self.n_blocks_dec = n_blocks_dec
+        self.kernel_size_dec = kernel_size_dec
+        self.dilation_rate = dilation_rate
+        self.n_block_layers = n_block_layers
+        self.p_dropout_dec = p_dropout_dec
+        self.n_speakers = n_speakers
+        self.gin_channels = gin_channels
+        self.n_split = n_split
+        self.n_sqz = n_sqz
+        self.sigmoid_scale = sigmoid_scale
+        self.window_size = window_size
+        self.block_length = block_length
+        self.mean_only = mean_only
+        self.hidden_channels_enc = hidden_channels_enc
+        self.hidden_channels_dec = hidden_channels_dec
+        self.prenet = prenet
+
+        self.encoder = TextEncoder(
+            n_vocab,
+            out_channels,
+            hidden_channels_enc or hidden_channels,
+            filter_channels,
+            filter_channels_dp,
+            n_heads,
+            n_layers_enc,
+            kernel_size,
+            p_dropout,
+            window_size=window_size,
+            block_length=block_length,
+            mean_only=mean_only,
+            prenet=prenet,
+            gin_channels=gin_channels,
+        )
+
+        self.decoder = FlowSpecDecoder(
+            out_channels,
+            hidden_channels_dec or hidden_channels,
+            kernel_size_dec,
+            dilation_rate,
+            n_blocks_dec,
+            n_block_layers,
+            p_dropout=p_dropout_dec,
+            n_split=n_split,
+            n_sqz=n_sqz,
+            sigmoid_scale=sigmoid_scale,
+            gin_channels=gin_channels,
+        )
+
+        if n_speakers > 1:
+            self.emb_g = nn.Embedding(n_speakers, gin_channels)
+            nn.init.uniform_(self.emb_g.weight, -0.1, 0.1)
+
+    def forward(
+        self,
+        x,
+        x_lengths,
+        y=None,
+        y_lengths=None,
+        g=None,
+        gen=False,
+        noise_scale=1.0,
+        length_scale=1.0,
+    ):
+        if g is not None:
+            g = F.normalize(self.emb_g(g)).unsqueeze(-1)  # [b, h]
+        x_m, x_logs, logw, x_mask = self.encoder(x, x_lengths, g=g)
+
+        if gen:
+            w = torch.exp(logw) * x_mask * length_scale
+            w_ceil = torch.ceil(w)
+            y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
+            y_max_length = None
+        else:
+            y_max_length = y.size(2)
+        y, y_lengths, y_max_length = self.preprocess(y, y_lengths, y_max_length)
+        z_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y_max_length), 1).to(
+            x_mask.dtype
+        )
+        attn_mask = torch.unsqueeze(x_mask, -1) * torch.unsqueeze(z_mask, 2)
+
+        if gen:
+            attn = commons.generate_path(
+                w_ceil.squeeze(1), attn_mask.squeeze(1)
+            ).unsqueeze(1)
+            z_m = torch.matmul(
+                attn.squeeze(1).transpose(1, 2), x_m.transpose(1, 2)
+            ).transpose(
+                1, 2
+            )  # [b, t', t], [b, t, d] -> [b, d, t']
+            z_logs = torch.matmul(
+                attn.squeeze(1).transpose(1, 2), x_logs.transpose(1, 2)
+            ).transpose(
+                1, 2
+            )  # [b, t', t], [b, t, d] -> [b, d, t']
+            logw_ = torch.log(1e-8 + torch.sum(attn, -1)) * x_mask
+
+            z = (z_m + torch.exp(z_logs) * torch.randn_like(z_m) * noise_scale) * z_mask
+            y, logdet = self.decoder(z, z_mask, g=g, reverse=True)
+            return (
+                (y, z_m, z_logs, logdet, z_mask),
+                (x_m, x_logs, x_mask),
+                (attn, logw, logw_),
+            )
+        else:
+            z, logdet = self.decoder(y, z_mask, g=g, reverse=False)
+            with torch.no_grad():
+                x_s_sq_r = torch.exp(-2 * x_logs)
+                logp1 = torch.sum(-0.5 * math.log(2 * math.pi) - x_logs, [1]).unsqueeze(
+                    -1
+                )  # [b, t, 1]
+                logp2 = torch.matmul(
+                    x_s_sq_r.transpose(1, 2), -0.5 * (z ** 2)
+                )  # [b, t, d] x [b, d, t'] = [b, t, t']
+                logp3 = torch.matmul(
+                    (x_m * x_s_sq_r).transpose(1, 2), z
+                )  # [b, t, d] x [b, d, t'] = [b, t, t']
+                logp4 = torch.sum(-0.5 * (x_m ** 2) * x_s_sq_r, [1]).unsqueeze(
+                    -1
+                )  # [b, t, 1]
+                logp = logp1 + logp2 + logp3 + logp4  # [b, t, t']
+
+                attn = (
+                    monotonic_align.maximum_path(logp, attn_mask.squeeze(1))
+                    .unsqueeze(1)
+                    .detach()
+                )
+            z_m = torch.matmul(
+                attn.squeeze(1).transpose(1, 2), x_m.transpose(1, 2)
+            ).transpose(
+                1, 2
+            )  # [b, t', t], [b, t, d] -> [b, d, t']
+            z_logs = torch.matmul(
+                attn.squeeze(1).transpose(1, 2), x_logs.transpose(1, 2)
+            ).transpose(
+                1, 2
+            )  # [b, t', t], [b, t, d] -> [b, d, t']
+            logw_ = torch.log(1e-8 + torch.sum(attn, -1)) * x_mask
+            return (
+                (z, z_m, z_logs, logdet, z_mask),
+                (x_m, x_logs, x_mask),
+                (attn, logw, logw_),
+            )
+
+    def preprocess(self, y, y_lengths, y_max_length):
+        if y_max_length is not None:
+            y_max_length = (y_max_length // self.n_sqz) * self.n_sqz
+            y = y[:, :, :y_max_length]
+        y_lengths = (y_lengths // self.n_sqz) * self.n_sqz
+        return y, y_lengths, y_max_length
+
+    def store_inverse(self):
+        self.decoder.store_inverse()

ttsv/src/glow_tts/modules.py

@@ -0,0 +1,276 @@
+import copy
+import math
+import numpy as np
+import scipy
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import commons
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-4):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        n_dims = len(x.shape)
+        mean = torch.mean(x, 1, keepdim=True)
+        variance = torch.mean((x - mean) ** 2, 1, keepdim=True)
+
+        x = (x - mean) * torch.rsqrt(variance + self.eps)
+
+        shape = [1, -1] + [1] * (n_dims - 2)
+        x = x * self.gamma.view(*shape) + self.beta.view(*shape)
+        return x
+
+
+class ConvReluNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        out_channels,
+        kernel_size,
+        n_layers,
+        p_dropout,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+        assert n_layers > 1, "Number of layers should be larger than 0."
+
+        self.conv_layers = nn.ModuleList()
+        self.norm_layers = nn.ModuleList()
+        self.conv_layers.append(
+            nn.Conv1d(
+                in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
+            )
+        )
+        self.norm_layers.append(LayerNorm(hidden_channels))
+        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
+        for _ in range(n_layers - 1):
+            self.conv_layers.append(
+                nn.Conv1d(
+                    hidden_channels,
+                    hidden_channels,
+                    kernel_size,
+                    padding=kernel_size // 2,
+                )
+            )
+            self.norm_layers.append(LayerNorm(hidden_channels))
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask):
+        x_org = x
+        for i in range(self.n_layers):
+            x = self.conv_layers[i](x * x_mask)
+            x = self.norm_layers[i](x)
+            x = self.relu_drop(x)
+        x = x_org + self.proj(x)
+        return x * x_mask
+
+
+class WN(torch.nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+    ):
+        super(WN, self).__init__()
+        assert kernel_size % 2 == 1
+        assert hidden_channels % 2 == 0
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = p_dropout
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(p_dropout)
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(
+                gin_channels, 2 * hidden_channels * n_layers, 1
+            )
+            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+        for i in range(n_layers):
+            dilation = dilation_rate ** i
+            padding = int((kernel_size * dilation - dilation) / 2)
+            in_layer = torch.nn.Conv1d(
+                hidden_channels,
+                2 * hidden_channels,
+                kernel_size,
+                dilation=dilation,
+                padding=padding,
+            )
+            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+            self.in_layers.append(in_layer)
+
+            # last one is not necessary
+            if i < n_layers - 1:
+                res_skip_channels = 2 * hidden_channels
+            else:
+                res_skip_channels = hidden_channels
+
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+            self.res_skip_layers.append(res_skip_layer)
+
+    def forward(self, x, x_mask=None, g=None, **kwargs):
+        output = torch.zeros_like(x)
+        n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+        if g is not None:
+            g = self.cond_layer(g)
+
+        for i in range(self.n_layers):
+            x_in = self.in_layers[i](x)
+            x_in = self.drop(x_in)
+            if g is not None:
+                cond_offset = i * 2 * self.hidden_channels
+                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+            else:
+                g_l = torch.zeros_like(x_in)
+
+            acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+
+            res_skip_acts = self.res_skip_layers[i](acts)
+            if i < self.n_layers - 1:
+                x = (x + res_skip_acts[:, : self.hidden_channels, :]) * x_mask
+                output = output + res_skip_acts[:, self.hidden_channels :, :]
+            else:
+                output = output + res_skip_acts
+        return output * x_mask
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0:
+            torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+
+class ActNorm(nn.Module):
+    def __init__(self, channels, ddi=False, **kwargs):
+        super().__init__()
+        self.channels = channels
+        self.initialized = not ddi
+
+        self.logs = nn.Parameter(torch.zeros(1, channels, 1))
+        self.bias = nn.Parameter(torch.zeros(1, channels, 1))
+
+    def forward(self, x, x_mask=None, reverse=False, **kwargs):
+        if x_mask is None:
+            x_mask = torch.ones(x.size(0), 1, x.size(2)).to(
+                device=x.device, dtype=x.dtype
+            )
+        x_len = torch.sum(x_mask, [1, 2])
+        if not self.initialized:
+            self.initialize(x, x_mask)
+            self.initialized = True
+
+        if reverse:
+            z = (x - self.bias) * torch.exp(-self.logs) * x_mask
+            logdet = None
+        else:
+            z = (self.bias + torch.exp(self.logs) * x) * x_mask
+            logdet = torch.sum(self.logs) * x_len  # [b]
+
+        return z, logdet
+
+    def store_inverse(self):
+        pass
+
+    def set_ddi(self, ddi):
+        self.initialized = not ddi
+
+    def initialize(self, x, x_mask):
+        with torch.no_grad():
+            denom = torch.sum(x_mask, [0, 2])
+            m = torch.sum(x * x_mask, [0, 2]) / denom
+            m_sq = torch.sum(x * x * x_mask, [0, 2]) / denom
+            v = m_sq - (m ** 2)
+            logs = 0.5 * torch.log(torch.clamp_min(v, 1e-6))
+
+            bias_init = (
+                (-m * torch.exp(-logs)).view(*self.bias.shape).to(dtype=self.bias.dtype)
+            )
+            logs_init = (-logs).view(*self.logs.shape).to(dtype=self.logs.dtype)
+
+            self.bias.data.copy_(bias_init)
+            self.logs.data.copy_(logs_init)
+
+
+class InvConvNear(nn.Module):
+    def __init__(self, channels, n_split=4, no_jacobian=False, **kwargs):
+        super().__init__()
+        assert n_split % 2 == 0
+        self.channels = channels
+        self.n_split = n_split
+        self.no_jacobian = no_jacobian
+
+        w_init = torch.qr(torch.FloatTensor(self.n_split, self.n_split).normal_())[0]
+        if torch.det(w_init) < 0:
+            w_init[:, 0] = -1 * w_init[:, 0]
+        self.weight = nn.Parameter(w_init)
+
+    def forward(self, x, x_mask=None, reverse=False, **kwargs):
+        b, c, t = x.size()
+        assert c % self.n_split == 0
+        if x_mask is None:
+            x_mask = 1
+            x_len = torch.ones((b,), dtype=x.dtype, device=x.device) * t
+        else:
+            x_len = torch.sum(x_mask, [1, 2])
+
+        x = x.view(b, 2, c // self.n_split, self.n_split // 2, t)
+        x = (
+            x.permute(0, 1, 3, 2, 4)
+            .contiguous()
+            .view(b, self.n_split, c // self.n_split, t)
+        )
+
+        if reverse:
+            if hasattr(self, "weight_inv"):
+                weight = self.weight_inv
+            else:
+                weight = torch.inverse(self.weight.float()).to(dtype=self.weight.dtype)
+            logdet = None
+        else:
+            weight = self.weight
+            if self.no_jacobian:
+                logdet = 0
+            else:
+                logdet = torch.logdet(self.weight) * (c / self.n_split) * x_len  # [b]
+
+        weight = weight.view(self.n_split, self.n_split, 1, 1)
+        z = F.conv2d(x, weight)
+
+        z = z.view(b, 2, self.n_split // 2, c // self.n_split, t)
+        z = z.permute(0, 1, 3, 2, 4).contiguous().view(b, c, t) * x_mask
+        return z, logdet
+
+    def store_inverse(self):
+        self.weight_inv = torch.inverse(self.weight.float()).to(dtype=self.weight.dtype)

ttsv/src/glow_tts/monotonic_align/monotonic_align/__init__.py

@@ -0,0 +1,5 @@
+import pkg_resources
+
+__version__ = pkg_resources.get_distribution("monotonic_align").version
+
+from monotonic_align.mas import *

ttsv/src/glow_tts/monotonic_align/monotonic_align/core.pyx

@@ -0,0 +1,45 @@
+import numpy as np
+cimport numpy as np
+cimport cython
+from cython.parallel import prange
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_x, int t_y, float max_neg_val) nogil:
+  cdef int x
+  cdef int y
+  cdef float v_prev
+  cdef float v_cur
+  cdef float tmp
+  cdef int index = t_x - 1
+
+  for y in range(t_y):
+    for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+      if x == y:
+        v_cur = max_neg_val
+      else:
+        v_cur = value[x, y-1]
+      if x == 0:
+        if y == 0:
+          v_prev = 0.
+        else:
+          v_prev = max_neg_val
+      else:
+        v_prev = value[x-1, y-1]
+      value[x, y] = max(v_cur, v_prev) + value[x, y]
+
+  for y in range(t_y - 1, -1, -1):
+    path[index, y] = 1
+    if index != 0 and (index == y or value[index, y-1] < value[index-1, y-1]):
+      index = index - 1
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_xs, int[::1] t_ys, float max_neg_val=-1e9) nogil:
+  cdef int b = values.shape[0]
+
+  cdef int i
+  for i in prange(b, nogil=True):
+    maximum_path_each(paths[i], values[i], t_xs[i], t_ys[i], max_neg_val)

ttsv/src/glow_tts/monotonic_align/monotonic_align/mas.py

@@ -0,0 +1,57 @@
+from typing import overload
+import numpy as np
+import torch
+from monotonic_align.core import maximum_path_c
+
+
+def mask_from_len(lens: torch.Tensor, max_len=None):
+    """
+    Make a `mask` from lens.
+
+    :param inputs: (B, T, D)
+    :param lens: (B)
+
+    :return:
+        `mask`: (B, T)
+    """
+    if max_len is None:
+        max_len = lens.max()
+    index = torch.arange(max_len).to(lens).view(1, -1)
+    return index < lens.unsqueeze(1)  # (B, T)
+
+
+def mask_from_lens(
+    similarity: torch.Tensor,
+    symbol_lens: torch.Tensor,
+    mel_lens: torch.Tensor,
+):
+    """
+    :param similarity: (B, S, T)
+    :param symbol_lens: (B,)
+    :param mel_lens: (B,)
+    """
+    _, S, T = similarity.size()
+    mask_S = mask_from_len(symbol_lens, S)
+    mask_T = mask_from_len(mel_lens, T)
+    mask_ST = mask_S.unsqueeze(2) * mask_T.unsqueeze(1)
+    return mask_ST.to(similarity)
+
+
+def maximum_path(value, mask=None):
+    """Cython optimised version.
+    value: [b, t_x, t_y]
+    mask: [b, t_x, t_y]
+    """
+    if mask is None:
+        mask = torch.zeros_like(value)
+
+    value = value * mask
+    device = value.device
+    dtype = value.dtype
+    value = value.data.cpu().numpy().astype(np.float32)
+    path = np.zeros_like(value).astype(np.int32)
+    mask = mask.data.cpu().numpy()
+    t_x_max = mask.sum(1)[:, 0].astype(np.int32)
+    t_y_max = mask.sum(2)[:, 0].astype(np.int32)
+    maximum_path_c(path, value, t_x_max, t_y_max)
+    return torch.from_numpy(path).to(device=device, dtype=dtype)

ttsv/src/glow_tts/monotonic_align/pyproject.toml

@@ -0,0 +1,7 @@
+[build-system]
+requires = [
+    "wheel",
+    "setuptools",
+    "cython>=0.24.0",
+    "numpy<v1.20.0",
+]

ttsv/src/glow_tts/monotonic_align/setup.py

@@ -0,0 +1,23 @@
+import numpy
+from setuptools import Extension, find_packages
+from distutils.core import setup
+from Cython.Build import cythonize
+
+
+_VERSION = "1.1"
+
+
+ext_modules = cythonize(
+    "monotonic_align/core.pyx",
+    compiler_directives={"language_level": "3"},
+)
+
+setup(
+    name="monotonic_align",
+    ext_modules=ext_modules,
+    include_dirs=[numpy.get_include(), "monotonic_align"],
+    packages=find_packages(),
+    setup_requires=["numpy", "cython"],
+    install_requires=["numpy"],
+    version=_VERSION,
+)