wav2vec2-xlsr-1b-et / README.md
anton-l's picture
anton-l HF staff
Upload README.md
a72c8ad
|
raw
history blame
2.47 kB
metadata
language: et
datasets:
  - mozilla-foundation/common_voice_8_0
metrics:
  - wer
  - cer
tags:
  - generated_from_trainer
  - mozilla-foundation/common_voice_8_0
  - audio
  - automatic-speech-recognition
  - speech
  - robust-speech-event
  - hf-asr-leaderboard
model-index:
  - name: XLS-R 1B Wav2Vec2 Estonian by Rasmus Toivanen
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: et
        metrics:
          - name: Test WER
            type: wer
            value: 20.12
          - name: Test CER
            type: cer
            value: 3.82
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: et
        metrics:
          - name: Test WER
            type: wer
            value: 40.77
          - name: Test CER
            type: cer
            value: 12.32
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Test Data
          type: speech-recognition-community-v2/eval_data
          args: et
        metrics:
          - name: Test WER
            type: wer
            value: 41.97

wav2vec2-xlsr-et-lm-1B

This model was finetuned with mozilla_foundation/common_voice_8_0 et with train+other+validation splits. It achieves the following results on the test set: (Loss reported with last eval step at step 2000/2040 during training)

  • Loss: 0.2150
  • Wer: 0.2012

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.00005
  • train_batch_size: 32
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 1
  • total_train_batch_size: 16
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 10
  • mixed_precision_training: Native AMP

Training results

Framework versions

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