anton-l's picture
anton-l HF staff
Upload README.md
9503ad1
metadata
language:
  - ba
license: apache-2.0
tags:
  - automatic-speech-recognition
  - ba
  - 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-bashkir
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: ba
        metrics:
          - name: Test WER
            type: wer
            value: 11.32
          - name: Test CER
            type: cer
            value: 2.34

sammy786/wav2vec2-xlsr-bashkir

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

  • Loss:
  • Wer:

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 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: 16
  • eval_batch_size: 16
  • seed: 13
  • gradient_accumulation_steps: 2
  • 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: 30
  • mixed_precision_training: Native AMP

Training results

Step Training Loss Validation Loss Wer
200 5.387100 1.982867 1.000000
400 1.269800 0.369958 0.545755
600 0.903600 0.287705 0.465594
800 0.787300 0.235142 0.417091
1000 0.816300 0.206325 0.390534
1200 0.700500 0.197106 0.383987
1400 0.707100 0.179855 0.381368
1600 0.657800 0.181605 0.370593
1800 0.647800 0.168626 0.358767
2000 0.650700 0.164833 0.351483
2200 0.490900 0.168133 0.363309
2400 0.431000 0.161201 0.344350
2600 0.372100 0.160254 0.338280
2800 0.367500 0.150885 0.329687
3000 0.351300 0.154112 0.331392
3200 0.314800 0.147147 0.326700
3400 0.316800 0.142681 0.325090
3600 0.313000 0.138736 0.319553
3800 0.291800 0.138166 0.315570
4000 0.311300 0.135977 0.322894
4200 0.304900 0.128820 0.308627
4400 0.301600 0.129475 0.307440
4600 0.281800 0.131863 0.305967

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-bashkir --dataset mozilla-foundation/common_voice_8_0 --config ba --split test