anton-l's picture
anton-l HF staff
Upload README.md
55558ca
metadata
language:
  - bas
license: apache-2.0
tags:
  - automatic-speech-recognition
  - bas
  - 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: XLS-R-300M - Basaa
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: bas
        metrics:
          - name: Test WER
            type: wer
            value: 38.057
          - name: Test CER
            type: cer
            value: 11.233

wav2vec2-large-xls-r-300m-basaa-cv8

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - BAS dataset. It achieves the following results on the evaluation set:

  • Loss: 0.4648
  • Wer: 0.5472

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: 7e-05
  • train_batch_size: 32
  • eval_batch_size: 16
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 100.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
2.9421 12.82 500 2.8894 1.0
1.1872 25.64 1000 0.6688 0.7460
0.8894 38.46 1500 0.4868 0.6516
0.769 51.28 2000 0.4960 0.6507
0.6936 64.1 2500 0.4781 0.5384
0.624 76.92 3000 0.4643 0.5430
0.5966 89.74 3500 0.4530 0.5591

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.18.3
  • Tokenizers 0.11.0