Modelpublic

pierreguillou/whisper-medium-french

sourceHugging Faceapache-2.0updated 4y agoView on Hugging Face

9likes48downloads

Model Card

French Medium Whisper

This model is a fine-tuned version of openai/whisper-medium on the commonvoice11_0 dataset. It achieves the following results on the evaluation set:

—Loss: 0.2664
—Wer (without normalization): 15.8969
—Wer (with normalization): 11.1406

Blog post

All information about this model in this blog post: Speech-to-Text & IA | Transcreva qualquer áudio para o português com o Whisper (OpenAI)... sem nenhum custo!.

New SOTA

The Normalized WER in the OpenAI Whisper article with the Common Voice 9.0 test dataset is 16.0.

As this test dataset is similar to the Common Voice 11.0 test dataset used to evaluate our model (WER and WER Norm), it means that our French Medium Whisper is better than the [Medium Whisper](https://huggingface.co/openai/whisper-medium) model at transcribing audios French in text.

OpenAI results with Whisper Medium and Test dataset of Commons Voice 9.0

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

—learning_rate: 1e-05
—trainbatchsize: 32
—evalbatchsize: 16
—seed: 42
—optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
—lrschedulertype: linear
—lrschedulerwarmup_steps: 500
—training_steps: 5000
—mixedprecisiontraining: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Wer Norm
0.2695	0.2	1000	0.3080	17.8083	12.9791
0.2099	0.4	2000	0.2981	17.4792	12.4242
0.1978	0.6	3000	0.2864	16.7767	12.0913
0.1455	0.8	4000	0.2752	16.4597	11.8966
0.1712	1.0	5000	0.2664	15.8969	11.1406

Framework versions

—Transformers 4.26.0.dev0
—Pytorch 1.13.0+cu117
—Datasets 2.7.1.dev0
—Tokenizers 0.13.2