xls-r-kyrgiz-cv8 / README.md
lucio's picture
Librarian Bot: Add base_model information to model (#2)
c5d2d9e verified
metadata
language:
  - ky
license: apache-2.0
tags:
  - automatic-speech-recognition
  - generated_from_trainer
  - hf-asr-leaderboard
  - mozilla-foundation/common_voice_8_0
  - robust-speech-event
datasets:
  - mozilla-foundation/common_voice_8_0
base_model: facebook/wav2vec2-xls-r-300m
model-index:
  - name: XLS-R-300M Kyrgiz CV8
    results:
      - task:
          type: automatic-speech-recognition
          name: Automatic Speech Recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: ky
        metrics:
          - type: wer
            value: 19.01
            name: Test WER (with LM)
          - type: cer
            value: 5.38
            name: Test CER (with LM)
          - type: wer
            value: 31.28
            name: Test WER (no LM)
          - type: cer
            value: 7.66
            name: Test CER (no LM)

XLS-R-300M Kyrgiz CV8

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

  • Loss: 0.5497
  • Wer: 0.2945
  • Cer: 0.0791

Model description

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

The model vocabulary consists of the cyrillic alphabet with punctuation removed.

The kenlm language model is built using the text of the train and invalidated corpus splits.

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, dev and other of common voice official splits were used as training data. The half of the official test split was used as validation data, as and the full test set was used for final evaluation.

Training procedure

The featurization layers of the XLS-R model are frozen while tuning a final CTC/LM layer on the Kyrgiz CV8 example sentences. A ramped learning rate is used with an initial warmup phase of 500 steps, a max of 0.0001, and cooling back towards 0 for the remainder of the 8100 steps (300 epochs).

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0001
  • train_batch_size: 32
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 128
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 300.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
3.1079 18.51 500 2.6795 0.9996 0.9825
0.8506 37.04 1000 0.4323 0.3718 0.0961
0.6821 55.55 1500 0.4105 0.3311 0.0878
0.6091 74.07 2000 0.4281 0.3168 0.0851
0.5429 92.58 2500 0.4525 0.3147 0.0842
0.5063 111.11 3000 0.4619 0.3144 0.0839
0.4661 129.62 3500 0.4660 0.3039 0.0818
0.4353 148.15 4000 0.4695 0.3083 0.0820
0.4048 166.65 4500 0.4909 0.3085 0.0824
0.3852 185.18 5000 0.5074 0.3048 0.0812
0.3567 203.69 5500 0.5111 0.3012 0.0810
0.3451 222.22 6000 0.5225 0.2982 0.0804
0.325 240.73 6500 0.5270 0.2955 0.0796
0.3089 259.25 7000 0.5381 0.2929 0.0793
0.2941 277.76 7500 0.5565 0.2923 0.0794
0.2945 296.29 8000 0.5495 0.2951 0.0789

Framework versions

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