deepdml/whisper-medium-mix-fr

This model is a fine-tuned version of deepdml/whisper-medium-mix-fr on the mozilla-foundation/common_voice_11_0, google/fleurs, facebook/multilingual_librispeech and facebook/voxpopuli datasets. It achieves the following results on the evaluation set:

Loss: 0.2599
Wer: 11.2278

Using the evalutaion script provided in the Whisper Sprint the model achieves these results on the test sets (WER):

google/fleurs: 9.3526 %
(python run_eval_whisper_streaming.py --model_id="deepdml/whisper-medium-mix-fr" --dataset="google/fleurs" --config="fr_fr" --device=0 --language="fr")
facebook/multilingual_librispeech: 6.3468 %
(python run_eval_whisper_streaming.py --model_id="deepdml/whisper-medium-mix-fr" --dataset="facebook/multilingual_librispeech" --config="french" --device=0 --language="fr")
facebook/voxpopuli: 10.0653 %
(python run_eval_whisper_streaming.py --model_id="deepdml/whisper-medium-mix-fr" --dataset="facebook/voxpopuli" --config="fr" --device=0 --language="fr")

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

Training data used:

mozilla-foundation/common_voice_11_0: fr, train+validation
google/fleurs: fr_fr, train
facebook/multilingual_librispeech: french, train
facebook/voxpopuli: fr, train

Evaluating over test split from mozilla-foundation/common_voice_11_0 dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 32
eval_batch_size: 16
seed: 42
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
0.0855	0.25	1000	0.2826	12.4230
0.0569	0.5	2000	0.2768	11.9577
0.0724	0.75	3000	0.2670	11.6106
0.069	1.0	4000	0.2599	11.2278

Framework versions

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