wav2vec2-xlsr-czech / README.md
anton-l's picture
anton-l HF staff
Upload README.md
7a6989f
metadata
language:
  - cs
license: apache-2.0
tags:
  - automatic-speech-recognition
  - cs
  - generated_from_trainer
  - hf-asr-leaderboard
  - model_for_talk
  - mozilla-foundation/common_voice_8_0
  - robust-speech-event
datasets:
  - mozilla-foundation/common_voice_8_0
model-index:
  - name: sammy786/wav2vec2-xlsr-czech
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: cs
        metrics:
          - name: Test WER
            type: wer
            value: 11.22
          - name: Test CER
            type: cer
            value: 2.52
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: cs
        metrics:
          - name: Test WER
            type: wer
            value: 97.02
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Test Data
          type: speech-recognition-community-v2/eval_data
          args: cs
        metrics:
          - name: Test WER
            type: wer
            value: 69.7

sammy786/wav2vec2-xlsr-czech

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - cs dataset. It achieves the following results on evaluation set (which is 10 percent of train data set merged with other and dev datasets):

  • Loss: 7.26
  • Wer: 19.32

Model description

"facebook/wav2vec2-xls-r-1b" was finetuned.

Intended uses & limitations

More information needed

Training and evaluation data

Training data - Common voice Finnish train.tsv, dev.tsv, invalidated.tsv and other.tsv

Training procedure

For creating the train dataset, all possible datasets were appended and 90-10 split was used.

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.000045637994662983496
  • train_batch_size: 8
  • eval_batch_size: 16
  • seed: 13
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine_with_restarts
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 7
  • mixed_precision_training: Native AMP

Training results

Step Training Loss Validation Loss Wer
200 6.654600 3.329486 1.000000
400 1.700600 0.317266 0.409446
600 0.767400 0.211371 0.313981
800 0.718600 0.167771 0.280676
1000 0.661700 0.142229 0.258938
1200 0.594400 0.137321 0.256275
1400 0.583900 0.132922 0.248418
1600 0.565100 0.117214 0.238640
1800 0.369600 0.116954 0.238291
2000 0.292800 0.109973 0.227509
2200 0.255400 0.104955 0.228120
2400 0.266800 0.097268 0.220525
2600 0.232700 0.096055 0.213584
2800 0.213700 0.097770 0.218866
3000 0.209900 0.091633 0.210485
3200 0.196800 0.090342 0.208739
3400 0.200500 0.082326 0.204767
3600 0.176800 0.085491 0.204068
3800 0.170000 0.081289 0.201231
4000 0.166200 0.080762 0.200227
4200 0.161700 0.076671 0.198001
4400 0.147000 0.077383 0.196997
4600 0.141900 0.076057 0.195862
4800 0.144800 0.074612 0.195120
5000 0.138900 0.073138 0.193985
5200 0.143900 0.072802 0.192894
5400 0.131100 0.072764 0.193723
5600 0.137000 0.072697 0.193679
5800 0.133300 0.072651 0.193286

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.0+cu102
  • Datasets 1.17.1.dev0
  • Tokenizers 0.10.3

Evaluation Commands

  1. To evaluate on mozilla-foundation/common_voice_8_0 with split test
python eval.py --model_id sammy786/wav2vec2-xlsr-czech --dataset mozilla-foundation/common_voice_8_0 --config cs --split test