AndrewMcDowell/wav2vec2-xls-r-1b-japanese-hiragana-katakana

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

• Loss: 0.5500
• Wer: 1.0132
• Cer: 0.1609

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: 8
• 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: 1500
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
1.7019	12.65	1000	1.0510	0.9832	0.2589
1.6385	25.31	2000	0.6670	0.9915	0.1851
1.4344	37.97	3000	0.6183	1.0213	0.1797

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-japanese-hiragana-katakana --dataset mozilla-foundation/common_voice_8_0 --config ja --split test --log_outputs

2. To evaluate on mozilla-foundation/common_voice_8_0 with split test

python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-1b-japanese-hiragana-katakana --dataset speech-recognition-community-v2/dev_data --config de --split validation --chunk_length_s 5.0 --stride_length_s 1.0