README.md

metadata

language: ru
datasets:
  - mozilla-foundation/common_voice_8_0
metrics:
  - wer
  - cer
tags:
  - audio
  - automatic-speech-recognition
  - generated_from_trainer
  - hf-asr-leaderboard
  - mozilla-foundation/common_voice_8_0
  - robust-speech-event
  - speech
model-index:
  - name: XLS-R 1B Wav2Vec2 Russian 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: ru
        metrics:
          - name: Test WER
            type: wer
            value: 10.83
          - name: Test CER
            type: cer
            value: 2.41
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: ru
        metrics:
          - name: Test WER
            type: wer
            value: 37.71
          - name: Test CER
            type: cer
            value: 12.98
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Test Data
          type: speech-recognition-community-v2/eval_data
          args: ru
        metrics:
          - name: Test WER
            type: wer
            value: 31.89

wav2vec2-xlsr-1b-ru

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 0.1352
• Wer: 0.0971

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: 5e-05
• train_batch_size: 32
• eval_batch_size: 8
• 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: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.5462	0.35	500	0.4027	0.3575
0.498	0.69	1000	0.2588	0.2513
0.4279	1.04	1500	0.2265	0.2204
0.4099	1.38	2000	0.2189	0.1979
0.4688	1.73	2500	0.2100	0.1920
0.2241	2.07	3000	0.1980	0.1767
0.2056	2.42	3500	0.2020	0.1683
0.3423	2.76	4000	0.1862	0.1606
0.2478	3.11	4500	0.1787	0.1563
0.3079	3.45	5000	0.1759	0.1555
0.2477	3.8	5500	0.1713	0.1423
0.1718	4.14	6000	0.1695	0.1391
0.1675	4.49	6500	0.1677	0.1372
0.1631	4.83	7000	0.1652	0.1333
0.1429	5.18	7500	0.1605	0.1308
0.1505	5.52	8000	0.1612	0.1245
0.1385	5.87	8500	0.1487	0.1225
0.1285	6.22	9000	0.1526	0.1201
0.1153	6.56	9500	0.1464	0.1172
0.1159	6.91	10000	0.1505	0.1143
0.1061	7.25	10500	0.1444	0.1106
0.1016	7.6	11000	0.1427	0.1075
0.1125	7.94	11500	0.1386	0.1045
0.0937	8.29	12000	0.1403	0.1022
0.1059	8.63	12500	0.1406	0.1022
0.0857	8.98	13000	0.1372	0.0992
0.0901	9.32	13500	0.1380	0.0977
0.0913	9.67	14000	0.1352	0.0971

Framework versions

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