vumichien/wav2vec2-xls-r-1b-japanese

Model description

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on my collection of Public Japanese Voice datasets for research Common Voice 7.0, JUST (Japanese speech corpus of Saruwatari-lab., University of Tokyo), JSSS (Japanese speech corpus for summarization and simplification), CSS10 (A collection of single speaker speech datasets). You can find in preprocessing dataset in here VUMICHIEN/COMMON_VOICE_LARGE_JSUT_JSSS_CSS10.

Total training data:

~60 hours

Benchmark WER result:

Benchmark CER result:

Evaluation

Please use the eval.py file to run the evaluation:

pip install mecab-python3 unidic-lite pykakasi
python eval.py --model_id vumichien/wav2vec2-xls-r-1b-japanese --dataset mozilla-foundation/common_voice_7_0 --config ja --split test --chunk_length_s 5.0 --stride_length_s 1.0 --log_outputs

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Framework versions

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