Wav2vec2 Xlsr Czech

🚀 sammy786/wav2vec2-xlsr-czech

This model is a fine - tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - cs dataset. It can be used for automatic speech recognition tasks, achieving certain performance on evaluation sets.

📄 License

This project is under the Apache - 2.0 license.

📚 Documentation

Model Information

Model description

"facebook/wav2vec2-xls-r-1b" was finetuned.

Intended uses & limitations

More information needed

Training and evaluation data

Training data - Common voice Finnish train.tsv, dev.tsv, invalidated.tsv and other.tsv

Training procedure

For creating the train dataset, all possible datasets were appended and a 90 - 10 split was used.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.000045637994662983496
• train_batch_size: 8
• eval_batch_size: 16
• seed: 13
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e - 08
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_steps: 500
• num_epochs: 7
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.0+cu102
• Datasets 1.17.1.dev0
• Tokenizers 0.10.3

Evaluation Results

The model achieves the following results on different evaluation sets:

1. MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - cs evaluation set (10 percent of train data set merged with other and dev datasets):
   • Loss: 7.26
   • Wer: 19.32
2. Other evaluation results:
   • Common Voice 8:
     • Test WER: 11.22
     • Test CER: 2.52
   • Robust Speech Event - Dev Data:
     • Test WER: 97.02
   • Robust Speech Event - Test Data:
     • Test WER: 69.7

Evaluation Commands

Basic Usage

# To evaluate on `mozilla-foundation/common_voice_8_0` with split `test`
python eval.py --model_id sammy786/wav2vec2-xlsr-czech --dataset mozilla-foundation/common_voice_8_0 --config cs --split test