metadata

language:
  - fr
license: apache-2.0
tags:
  - automatic-speech-recognition
  - polinaeterna/voxpopuli
  - generated_from_trainer
  - hf-asr-leaderboard
  - robust-speech-event
datasets:
  - polinaeterna/voxpopuli
model-index:
  - name: Fine-tuned Wav2Vec2 XLS-R 1B model for ASR in French
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Voxpopuli
          type: polinaeterna/voxpopuli
          args: fr
        metrics:
          - name: Test WER
            type: wer
            value: 11.7
          - name: Test CER
            type: cer
            value: 5.8
          - name: Test WER (+LM)
            type: wer
            value: 10.01
          - name: Test CER (+LM)
            type: cer
            value: 5.63
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 9
          type: mozilla-foundation/common_voice_9_0
          args: fr
        metrics:
          - name: Test WER
            type: wer
            value: 45.74
          - name: Test CER
            type: cer
            value: 22.99
          - name: Test WER (+LM)
            type: wer
            value: 38.81
          - name: Test CER (+LM)
            type: cer
            value: 23.25
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: fr
        metrics:
          - name: Test WER
            type: wer
            value: 27.86
          - name: Test CER
            type: cer
            value: 13.2
          - name: Test WER (+LM)
            type: wer
            value: 22.53
          - name: Test CER (+LM)
            type: cer
            value: 12.82

Fine-tuned Wav2Vec2 XLS-R 1B model for ASR in French

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the POLINAETERNA/VOXPOPULI - FR dataset. It achieves the following results on the evaluation set:

Loss: 0.2906
Wer: 0.1093

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_epochs: 12.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.4628	0.93	500	0.3834	0.1625
0.3577	1.85	1000	0.3231	0.1367
0.3103	2.78	1500	0.2918	0.1287
0.2884	3.7	2000	0.2845	0.1227
0.2615	4.63	2500	0.2819	0.1189
0.242	5.56	3000	0.2915	0.1165
0.2268	6.48	3500	0.2768	0.1187
0.2188	7.41	4000	0.2719	0.1128
0.1979	8.33	4500	0.2741	0.1134
0.1834	9.26	5000	0.2827	0.1096
0.1719	10.19	5500	0.2906	0.1093
0.1723	11.11	6000	0.2868	0.1104

Framework versions

Transformers 4.23.0.dev0
Pytorch 1.12.0+cu113
Datasets 2.4.0
Tokenizers 0.12.1

Evaluation

To evaluate on mozilla-foundation/common_voice_9_0

python eval.py \
  --model_id "bhuang/wav2vec2-xls-r-1b-voxpopuli-fr" \
  --dataset "polinaeterna/voxpopuli" \
  --config "fr" \
  --split "test" \
  --log_outputs \
  --outdir "outputs/results_polinaeterna_voxpopuli_with_lm"

To evaluate on mozilla-foundation/common_voice_9_0

python eval.py \
  --model_id "bhuang/wav2vec2-xls-r-1b-voxpopuli-fr" \
  --dataset "mozilla-foundation/common_voice_9_0" \
  --config "fr" \
  --split "test" \
  --log_outputs \
  --outdir "outputs/results_mozilla-foundatio_common_voice_9_0_with_lm"

To evaluate on speech-recognition-community-v2/dev_data

python eval.py \
  --model_id "bhuang/wav2vec2-xls-r-1b-voxpopuli-fr" \
  --dataset "speech-recognition-community-v2/dev_data" \
  --config "fr" \
  --split "validation" \
  --chunk_length_s 5.0 \
  --stride_length_s 1.0 \
  --log_outputs \
  --outdir "outputs/results_speech-recognition-community-v2_dev_data_with_lm"