anton-l's picture
anton-l HF staff
Upload README.md
f56917a
|
raw
history blame
4.72 kB
metadata
language:
  - ckb
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
  - ckb
  - robust-speech-event
  - model_for_talk
  - hf-asr-leaderboard
datasets:
  - mozilla-foundation/common_voice_8_0
model-index:
  - name: Akashpb13/Central_kurdish_xlsr
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: ckb
        metrics:
          - name: Test WER
            type: wer
            value: 0.36754389884276845
          - name: Test CER
            type: cer
            value: 0.07827896768334217
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: ckb
        metrics:
          - name: Test WER
            type: wer
            value: 0.36754389884276845
          - name: Test CER
            type: cer
            value: 0.07827896768334217

Akashpb13/Central_kurdish_xlsr

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - hu dataset. It achieves the following results on evaluation set (which is 10 percent of train data set merged with invalidated data, reported, other and dev datasets):

  • Loss: 0.348580
  • Wer: 0.401147

Model description

"facebook/wav2vec2-xls-r-300m" was finetuned.

Intended uses & limitations

More information needed

Training and evaluation data

Training data - Common voice Central Kurdish train.tsv, dev.tsv, invalidated.tsv, reported.tsv, and other.tsv Only those points were considered where upvotes were greater than downvotes and duplicates were removed after concatenation of all the datasets given in common voice 7.0

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: 0.000095637994662983496
  • train_batch_size: 16
  • eval_batch_size: 16
  • seed: 13
  • gradient_accumulation_steps: 2
  • lr_scheduler_type: cosine_with_restarts
  • lr_scheduler_warmup_steps: 200
  • num_epochs: 100
  • mixed_precision_training: Native AMP

Training results

Step Training Loss Validation Loss Wer
500 5.097800 2.190326 1.001207
1000 0.797500 0.331392 0.576819
1500 0.405100 0.262009 0.549049
2000 0.322100 0.248178 0.479626
2500 0.264600 0.258866 0.488983
3000 0.228300 0.261523 0.469665
3500 0.201000 0.270135 0.451856
4000 0.180900 0.279302 0.448536
4500 0.163800 0.280921 0.459704
5000 0.147300 0.319249 0.471778
5500 0.137600 0.289546 0.449140
6000 0.132000 0.311350 0.458195
6500 0.117100 0.316726 0.432840
7000 0.109200 0.302210 0.439481
7500 0.104900 0.325913 0.439481
8000 0.097500 0.329446 0.431935
8500 0.088600 0.345259 0.425898
9000 0.084900 0.342891 0.428313
9500 0.080900 0.353081 0.424389
10000 0.075600 0.347063 0.424992
10500 0.072800 0.330086 0.424691
11000 0.068100 0.350658 0.421974
11500 0.064700 0.342949 0.413522
12000 0.061500 0.341704 0.415334
12500 0.059500 0.346279 0.411410
13000 0.057400 0.349901 0.407184
13500 0.056400 0.347733 0.402656
14000 0.053300 0.344899 0.405976
14500 0.052900 0.346708 0.402656
15000 0.050600 0.344118 0.400845
15500 0.050200 0.348396 0.402958
16000 0.049800 0.348312 0.401751
16500 0.051900 0.348372 0.401147
17000 0.049800 0.348580 0.401147

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.0+cu102
  • Datasets 1.18.1
  • Tokenizers 0.10.3

Evaluation Commands

  1. To evaluate on mozilla-foundation/common_voice_8_0 with split test
python eval.py --model_id Akashpb13/Central_kurdish_xlsr --dataset mozilla-foundation/common_voice_8_0 --config ckb --split test