xls-r-uyghur-cv8 / README.md
lucio's picture
Update README.md
f8a8e11
|
raw
history blame
5.32 kB
metadata
language:
  - ug
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
  - ug
  - robust-speech-event
datasets:
  - mozilla-foundation/common_voice_8_0
model-index:
  - name: XLS-R-300M Uyghur CV8
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: ug
        metrics:
          - name: Test WER
            type: wer
            value: 28.74
          - name: Test CER
            type: cer
            value: 5.38

XLS-R-300M Uyghur CV8

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

  • Loss: 0.2036
  • WER: 0.2977

Model description

For a description of the model architecture, see facebook/wav2vec2-xls-r-300m

The model vocabulary consists of the alphabetic characters of the Perso-Arabic script for the Uyghur language, with punctuation removed.

Intended uses & limitations

This model is expected to be of some utility for low-fidelity use cases such as:

  • Draft video captions
  • Indexing of recorded broadcasts

The model is not reliable enough to use as a substitute for live captions for accessibility purposes, and it should not be used in a manner that would infringe the privacy of any of the contributors to the Common Voice dataset nor any other speakers.

Training and evaluation data

The combination of train and dev of common voice official splits were used as training data. The official test split was used as validation data as well as for final evaluation.

Training procedure

The featurization layers of the XLS-R model are frozen while tuning a final CTC/LM layer on the Uyghur CV8 example sentences. A ramped learning rate is used with an initial warmup phase of 2000 steps, a max of 0.0001, and cooling back towards 0 for the remainder of the 18500 steps (100 epochs).

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0001
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 2000
  • num_epochs: 100.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
3.2892 2.66 500 3.2415 1.0
2.9206 5.32 1000 2.4381 1.0056
1.4909 7.97 1500 0.5428 0.6705
1.3395 10.64 2000 0.4207 0.5995
1.2718 13.3 2500 0.3743 0.5648
1.1798 15.95 3000 0.3225 0.4927
1.1392 18.61 3500 0.3097 0.4627
1.1143 21.28 4000 0.2996 0.4505
1.0923 23.93 4500 0.2841 0.4229
1.0516 26.59 5000 0.2705 0.4113
1.051 29.25 5500 0.2622 0.4078
1.021 31.91 6000 0.2611 0.4009
0.9886 34.57 6500 0.2498 0.3921
0.984 37.23 7000 0.2521 0.3845
0.9631 39.89 7500 0.2413 0.3791
0.9353 42.55 8000 0.2391 0.3612
0.922 45.21 8500 0.2363 0.3571
0.9116 47.87 9000 0.2285 0.3668
0.8951 50.53 9500 0.2256 0.3729
0.8865 53.19 10000 0.2228 0.3663
0.8792 55.85 10500 0.2221 0.3656
0.8682 58.51 11000 0.2228 0.3323
0.8492 61.17 11500 0.2167 0.3446
0.8365 63.83 12000 0.2156 0.3321
0.8298 66.49 12500 0.2142 0.3400
0.808 69.15 13000 0.2079 0.3148
0.7999 71.81 13500 0.2117 0.3225
0.7871 74.47 14000 0.2088 0.3174
0.7858 77.13 14500 0.2060 0.3008
0.7764 79.78 15000 0.2128 0.3146
0.7684 82.45 15500 0.2086 0.3101
0.7717 85.11 16000 0.2048 0.3069
0.7435 87.76 16500 0.2027 0.3055
0.7378 90.42 17000 0.2059 0.2993
0.7406 93.08 17500 0.2040 0.2966
0.7361 95.74 18000 0.2056 0.3000
0.7379 98.4 18500 0.2031 0.2976

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.18.2.dev0
  • Tokenizers 0.11.0