README.md

metadata

language:
  - ca
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - collectivat/tv3_parla
  - projecte-aina/parlament_parla
  - generated_from_trainer
  - robust-speech-event
datasets:
  - mozilla-foundation/common_voice_8_0
  - collectivat/tv3_parla
  - projecte-aina/parlament_parla
model-index:
  - name: wav2vec2-xls-r-300m-ca-lm
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: mozilla-foundation/common_voice_8_0 ca
          type: mozilla-foundation/common_voice_8_0
          args: ca
        metrics:
          - name: Test WER
            type: wer
            value: 0.08108860330598514
          - name: Test CER
            type: cer
            value: 0.027241712812152218
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: projecte-aina/parlament_parla ca
          type: projecte-aina/parlament_parla
          args: clean
        metrics:
          - name: Test WER
            type: wer
            value: 0.06541946111307212
          - name: Test CER
            type: cer
            value: 0.02205785796827398
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: collectivat/tv3_parla ca
          type: collectivat/tv3_parla
          args: ca
        metrics:
          - name: Test WER
            type: wer
            value: 0.1506717480848443
          - name: Test CER
            type: cer
            value: 0.09562445266717665

wav2vec2-xls-r-300m-ca

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

• Loss: 0.2758
• Wer: 0.1792

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: 7.5e-05
• train_batch_size: 32
• eval_batch_size: 32
• 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: 2000
• num_epochs: 6.0
• mixed_precision_training: Native AMP

Training results (without LM)

Training Loss	Epoch	Step	Validation Loss	Wer
6.2099	0.09	500	3.4125	1.0
2.9961	0.18	1000	2.9224	1.0
2.2147	0.26	1500	0.6521	0.5568
1.3017	0.35	2000	0.3153	0.2761
1.1196	0.44	2500	0.2444	0.2367
1.0712	0.53	3000	0.2324	0.2132
1.052	0.62	3500	0.2173	0.2032
1.2813	2.13	4000	0.3326	0.2099
1.2365	2.4	4500	0.3224	0.2003
1.2193	2.66	5000	0.3198	0.1957
1.2072	2.93	5500	0.3063	0.1933
1.213	3.2	6000	0.3051	0.1980
1.2074	3.46	6500	0.3012	0.1879
1.1918	3.73	7000	0.2947	0.1829
1.1893	4.0	7500	0.2895	0.1807
1.1751	4.26	8000	0.2878	0.1776
1.1628	4.53	8500	0.2835	0.1731
1.1577	4.79	9000	0.2816	0.1761
1.1448	5.06	9500	0.2757	0.1740
1.1407	5.33	10000	0.2768	0.1798
1.1401	5.59	10500	0.2780	0.1816
1.1333	5.86	11000	0.2748	0.1750

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.1
• Tokenizers 0.11.0