whisper-medium-he[WIP]

This model is a fine-tuned version of imvladikon/whisper-medium-he on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2061
• Wer: 13.4020

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0983	0.1	1000	0.3072	16.4362
0.1219	0.2	2000	0.2923	15.6642
0.134	0.3	3000	0.2345	13.7450
0.2113	0.39	4000	0.2061	13.4020

Inference

HF

from transformers import pipeline
pipe = pipeline("automatic-speech-recognition", model="imvladikon/whisper-medium-he", device_map="auto") # requires `pip install accelerate`
print(recognize("sample.mp3"))

whisper.cpp

Prepared : https://huggingface.co/imvladikon/whisper-medium-he/blob/main/ggml-hebrew.bin

But if need to convert:

git clone https://github.com/openai/whisper
git clone https://github.com/ggerganov/whisper.cpp
git clone https://huggingface.co/imvladikon/whisper-medium-he
python3 ./whisper.cpp/models/convert-h5-to-ggml.py ./whisper-medium-he/ ./whisper .

Then possible to check (if produced model is ggml-model.bin):

cd whisper.cpp && ./main -m ../ggml-model.bin -f ../sample.wav

Framework versions

• Transformers 4.36.0.dev0
• Pytorch 2.1.0+cu118
• Datasets 2.15.0
• Tokenizers 0.15.0