proxectonos/Nos_TTS-sabela-vits-phonemes

Sabela: Nos Project's Galician TTS Model

Model description

This model was trained from scratch using the Coqui TTS Python library on the Sabela corpus of the dataset CRPIH_UVigo-GL-Voices.

A live inference demo can be found in our official page, here.

The model needs the Cotovia tool to work correctly. For installation and deployment please consult the Cotovía Preprocessor section.

Intended uses and limitations

You can use this model to generate synthetic speech in Galician.

How to use

Usage

Cotovía preprocessor

To generate fonectic transcriptions, the Cotovía tool is needed. The tool can be downloaded from the SourceForge website. The required debian packages are cotovia_0.5_amd64.deb and cotovia-lang-gl_0.5_all.deb, that can be installed with the following commands:

sudo dpkg -i cotovia_0.5_amd64.deb
sudo dpkg -i cotovia-lang-gl_0.5_all.deb

The tool can be used to generate the phonetic transcription of the text. The following command can be used to generate the phonetic transcription of a text string:

echo "Era unha avioneta... O piloto era pequeno, que se chega a ser dos grande, tómbate!" | cotovia -t -n -S | iconv -f iso88591 -t utf8

The output of the command is the phonetic transcription of the input text. This string may be used in the inference part, as shown next.

Required libraries:

pip install TTS

Synthesize a speech using python:

import tempfile
import numpy as np
import os
import json

from typing import Optional
from TTS.config import load_config
from TTS.utils.manage import ModelManager
from TTS.utils.synthesizer import Synthesizer
model_path = # Absolute path to the model checkpoint.pth
config_path = # Absolute path to the model config.json
text = "Text to synthetize"
synthesizer = Synthesizer(
    model_path, config_path, None, None, None, None,
)
wavs = synthesizer.tts(text)

Training

Training Procedure

Data preparation

Hyperparameter

The model is based on VITS proposed by Kim et al. The following hyperparameters were set in the coqui framework.

Hyperparameter	Value
Model	vits
Batch Size	48
Eval Batch Size	16
Mixed Precision	true
Window Length	1024
Hop Length	256
FTT size	1024
Num Mels	80
Phonemizer	null
Phoneme Lenguage	null
Text Cleaners	null
Formatter	nos_fonemas
Optimizer	adam
Adam betas	(0.8, 0.99)
Adam eps	1e-09
Adam weight decay	0.01
Learning Rate Gen	0.0002
Lr. schedurer Gen	ExponentialLR
Lr. schedurer Gamma Gen	0.999875
Learning Rate Disc	0.0002
Lr. schedurer Disc	ExponentialLR
Lr. schedurer Gamma Disc	0.999875

The model was trained for 256275 steps.

The nos_fonemas formatter is a modification of the LJSpeech formatter with one extra column for the normalized input (extended numbers and acronyms).

Additional information

Authors

Alp Öktem, Carmen Magariños and Antonio Moscoso.

Contact information

For further information, send an email to [email protected]

Licensing Information

Apache License, Version 2.0

Funding

This research was funded by “The Nós project: Galician in the society and economy of Artificial Intelligence”, resulting from the agreement 2021-CP080 between the Xunta de Galicia and the University of Santiago de Compostela, and thanks to the Investigo program, within the National Recovery, Transformation and Resilience Plan, within the framework of the European Recovery Fund (NextGenerationEU).

Citation information

If you use this model, please cite as follows:

Öktem, Alp; Magariños, Carmen; Moscoso, Antonio. 2024. Nos_TTS-sabela-vits-phonemes. URL: https://huggingface.co/proxectonos/Nos_TTS-sabela-vits-phonemes