Hugging Face's logo

Plim
/
xls-r-300m-cv_8-fr

Automatic Speech Recognition Transformers PyTorch French wav2vec2 mozilla-foundation/common_voice_8_0 generated_from_trainer Eval Results Inference Endpoints
Model card Files Community
xls-r-300m-cv_8-fr / README.md
Plim's picture
Plim
Update README.md
799632e
preview code
|
raw history blame contribute delete
No virus
4.14 kB
metadata
language:
  - fr
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
model-index:
  - name: XLS-R-300m - French
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: fr
        metrics:
          - name: Test WER
            type: wer
            value: to recompute with STEP 24000
          - name: Test CER
            type: cer
            value: to recompute with STEP 24000
      - 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: 35.29
          - name: Test CER
            type: cer
            value: 13.94

Model description

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - FR dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 7.5e-05
  • train_batch_size: 16
  • eval_batch_size: 16
  • 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_steps: 2000
  • num_epochs: 5.0 (extended to 7.0 with training with checkpoint)
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
2.9114 0.29 1000 inf 0.9997
1.2436 0.57 2000 inf 0.4310
1.0552 0.86 3000 inf 0.3144
1.0044 1.15 4000 inf 0.2814
0.9718 1.43 5000 inf 0.2658
0.9502 1.72 6000 inf 0.2566
0.9418 2.01 7000 inf 0.2476
0.9215 2.29 8000 inf 0.2420
0.9236 2.58 9000 inf 0.2388
0.9014 2.87 10000 inf 0.2354
0.8814 3.15 11000 inf 0.2312
0.8809 3.44 12000 inf 0.2285
0.8717 3.73 13000 inf 0.2263
0.8787 4.01 14000 inf 0.2218
0.8567 4.3 15000 inf 0.2193
0.8488 4.59 16000 inf 0.2187
0.8359 4.87 17000 inf 0.2172

Training continued with checkpoint from STEP 17000:

Training Loss Epoch Step Validation Loss Wer
/ 5.16 18000 inf 0.2176
/ 5.45 19000 inf 0.2181
/ 5.73 20000 inf 0.2155
/ 6.02 21000 inf 0.2140
/ 6.31 22000 inf 0.2124
/ 6.59 23000 inf 0.2117
/ 6.88 24000 inf 0.2116

It achieves the best result on the validation set on Step 24000:

  • Wer: 0.2116

Got some issue with validation loss calculation.

Framework versions

  • Transformers 4.17.0.dev0
  • Pytorch 1.10.2+cu102
  • Datasets 1.18.3.dev0
  • Tokenizers 0.11.0

Evaluation Commands

  1. To evaluate on mozilla-foundation/common_voice_8 with split test
python eval.py --model_id Plim/xls-r-300m-cv_8-fr --dataset mozilla-foundation/common_voice_8_0 --config fr --split test
  1. To evaluate on speech-recognition-community-v2/dev_data
python eval.py --model_id Plim/xls-r-300m-cv_8-fr --dataset speech-recognition-community-v2/dev_data --config fr --split validation --chunk_length_s 5.0 --stride_length_s 1.0