AndrewMcDowell's picture
Update README.md
5262dde
|
raw
history blame
4.18 kB
metadata
language:
  - de
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
  - robust-speech-event
  - de
datasets:
  - common_voice
model-index:
  - name: XLS-R-1B - German
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: de
        metrics:
          - name: Test WER
            type: wer
            value: 16.12
          - name: Test CER
            type: cer
            value: 4.023836120707133
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: de
        metrics:
          - name: Test WER
            type: wer
            value: 36.76
          - name: Test CER
            type: cer
            value: 14.03

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - DE dataset. It achieves the following results on the evaluation set:

  • Loss: 0.1476
  • Wer: 0.1612

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: 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: 2.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
1.1842 0.07 1000 0.4461 0.4918
1.1317 0.15 2000 0.2669 0.2748
1.1029 0.22 3000 0.2638 0.2706
1.0949 0.29 4000 0.2519 0.2627
1.0923 0.37 5000 0.2475 0.2577
1.0847 0.44 6000 0.2436 0.2612
1.0667 0.51 7000 0.2472 0.2661
1.0709 0.59 8000 0.2489 0.2610
1.0472 0.66 9000 0.2354 0.2500
1.0604 0.73 10000 0.2346 0.2485
1.0375 0.81 11000 0.2286 0.2390
1.0193 0.88 12000 0.2212 0.2338
1.0077 0.95 13000 0.2152 0.2269
1.0004 1.03 14000 0.2093 0.2207
0.9649 1.1 15000 0.1993 0.2113
0.9509 1.17 16000 0.1934 0.2089
0.9533 1.25 17000 0.1874 0.2023
0.9248 1.32 18000 0.1818 0.1974
0.9216 1.39 19000 0.1776 0.1926
0.8964 1.47 20000 0.1722 0.1904
0.8941 1.54 21000 0.1690 0.1852
0.871 1.61 22000 0.1627 0.1781
0.847 1.69 23000 0.1591 0.1751
0.822 1.76 24000 0.1551 0.1701
0.8188 1.83 25000 0.1528 0.1667
0.8305 1.91 26000 0.1492 0.1631
0.8122 1.98 27000 0.1479 0.1611

Framework versions

  • Transformers 4.17.0.dev0
  • Pytorch 1.10.2+cu102
  • Datasets 1.18.2.dev0
  • Tokenizers 0.11.0

Evaluation Commands

  1. To evaluate on mozilla-foundation/common_voice_8_0 with split test
python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-1B-german-de --dataset mozilla-foundation/common_voice_8_0 --config de --split test --log_outputs