Hugging Face's logo

sammy786
/
wav2vec2-xlsr-finnish

Automatic Speech Recognition Transformers PyTorch
Finnish wav2vec2 generated_from_trainer hf-asr-leaderboard model_for_talk mozilla-foundation/common_voice_8_0 robust-speech-event Eval Results Inference Endpoints
Model card Files Community
wav2vec2-xlsr-finnish / README.md
anton-l's picture
anton-l HF staff
Upload README.md
f692779
preview code
|
raw history blame contribute delete
No virus
4.24 kB
metadata
language:
  - fi
license: apache-2.0
tags:
  - automatic-speech-recognition
  - fi
  - 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-finnish
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: fi
        metrics:
          - name: Test WER
            type: wer
            value: 13.72
          - name: Test CER
            type: cer
            value: 2.35

sammy786/wav2vec2-xlsr-finnish

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

  • Loss: 8.7555
  • Wer: 23.0231

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: 30
  • mixed_precision_training: Native AMP

Training results

Step Training Loss Validation Loss Wer
200 4.253700 0.881733 0.967007
400 0.864800 0.226977 0.420836
600 0.607000 0.157473 0.343375
800 0.380200 0.145640 0.302672
1000 0.318400 0.128028 0.293886
1200 0.261100 0.121414 0.289941
1400 0.232300 0.113451 0.279182
1600 0.216600 0.113649 0.282948
1800 0.202500 0.112375 0.276134
2000 0.190000 0.105725 0.273803
2200 0.171000 0.109715 0.270755
2400 0.156500 0.105042 0.264300
2600 0.155600 0.108337 0.260714
2800 0.149100 0.112435 0.263583
3000 0.145100 0.106193 0.261969
3200 0.131700 0.102860 0.251210
3400 0.129100 0.096058 0.246907
3600 0.121600 0.099932 0.246369
3800 0.112000 0.099041 0.244397
4000 0.114100 0.101566 0.242604
4200 0.111500 0.089498 0.239197
4400 0.099800 0.092835 0.240990
4600 0.095300 0.093518 0.238121
4800 0.094300 0.090783 0.240631
5000 0.089000 0.094046 0.238479
5200 0.088000 0.089342 0.235252
5400 0.083600 0.087770 0.234535
5600 0.083600 0.088804 0.234355
5800 0.080300 0.090168 0.231307
6000 0.078100 0.090163 0.230949
6200 0.075600 0.088876 0.232383
6400 0.078700 0.087235 0.232024
6600 0.074800 0.086825 0.231486
6800 0.076400 0.087308 0.231845
7000 0.070700 0.087695 0.230769
7200 0.075500 0.087555 0.230231

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-finnish --dataset mozilla-foundation/common_voice_8_0 --config fi --split test