anton-l's picture
anton-l HF staff
Upload README.md
0000611
metadata
language:
  - cs
license: apache-2.0
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_8_0
  - generated_from_trainer
  - robust-speech-event
  - xlsr-fine-tuning-week
  - hf-asr-leaderboard
datasets:
  - mozilla-foundation/common_voice_8_0
model-index:
  - name: Czech comodoro Wav2Vec2 XLSR 300M CV8
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 8
          type: mozilla-foundation/common_voice_8_0
          args: cs
        metrics:
          - name: Test WER
            type: wer
            value: 10.3
          - name: Test CER
            type: cer
            value: 2.6
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Dev Data
          type: speech-recognition-community-v2/dev_data
          args: cs
        metrics:
          - name: Test WER
            type: wer
            value: 54.29
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Robust Speech Event - Test Data
          type: speech-recognition-community-v2/eval_data
          args: cs
        metrics:
          - name: Test WER
            type: wer
            value: 44.55

wav2vec2-xls-r-300m-cs-cv8

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice 8.0 dataset. It achieves the following results on the evaluation set while training:

  • Loss: 0.2327
  • Wer: 0.1608
  • Cer: 0.0376

The eval.py script results using a LM are: WER: 0.10281503199350225 CER: 0.02622802241689026

Model description

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Czech using the Common Voice dataset. When using this model, make sure that your speech input is sampled at 16kHz.

The model can be used directly (without a language model) as follows:

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

test_dataset = load_dataset("mozilla-foundation/common_voice_8_0", "cs", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("comodoro/wav2vec2-xls-r-300m-cs-cv8")
model = Wav2Vec2ForCTC.from_pretrained("comodoro/wav2vec2-xls-r-300m-cs-cv8")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset[:2]["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset[:2]["sentence"])

Evaluation

The model can be evaluated using the attached eval.py script:

python eval.py --model_id comodoro/wav2vec2-xls-r-300m-cs-cv8 --dataset mozilla-foundation/common-voice_8_0 --split test --config cs

Training and evaluation data

The Common Voice 8.0 train and validation datasets were used for training

Training procedure

Training hyperparameters

The following hyperparameters were used during first stage of training:

  • learning_rate: 7e-05
  • train_batch_size: 32
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 20
  • total_train_batch_size: 640
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 150
  • mixed_precision_training: Native AMP

The following hyperparameters were used during second stage of training:

  • learning_rate: 0.001
  • train_batch_size: 32
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 20
  • total_train_batch_size: 640
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 50
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
7.2926 8.06 250 3.8497 1.0 1.0
3.417 16.13 500 3.2852 1.0 0.9857
2.0264 24.19 750 0.7099 0.7342 0.1768
0.4018 32.25 1000 0.6188 0.6415 0.1551
0.2444 40.32 1250 0.6632 0.6362 0.1600
0.1882 48.38 1500 0.6070 0.5783 0.1388
0.153 56.44 1750 0.6425 0.5720 0.1377
0.1214 64.51 2000 0.6363 0.5546 0.1337
0.1011 72.57 2250 0.6310 0.5222 0.1224
0.0879 80.63 2500 0.6353 0.5258 0.1253
0.0782 88.7 2750 0.6078 0.4904 0.1127
0.0709 96.76 3000 0.6465 0.4960 0.1154
0.0661 104.82 3250 0.6622 0.4945 0.1166
0.0616 112.89 3500 0.6440 0.4786 0.1104
0.0579 120.95 3750 0.6815 0.4887 0.1144
0.0549 129.03 4000 0.6603 0.4780 0.1105
0.0527 137.09 4250 0.6652 0.4749 0.1090
0.0506 145.16 4500 0.6958 0.4846 0.1133

Further fine-tuning with slightly different architecture and higher learning rate:

Training Loss Epoch Step Validation Loss Wer Cer
0.576 8.06 250 0.2411 0.2340 0.0502
0.2564 16.13 500 0.2305 0.2097 0.0492
0.2018 24.19 750 0.2371 0.2059 0.0494
0.1549 32.25 1000 0.2298 0.1844 0.0435
0.1224 40.32 1250 0.2288 0.1725 0.0407
0.1004 48.38 1500 0.2327 0.1608 0.0376

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.17.1.dev0
  • Tokenizers 0.11.0