vutankiet2901/wav2vec2-large-xlsr-53-ja

Model description

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - JA dataset.

Benchmark WER result:

	COMMON VOICE 7.0	COMMON VOICE 8.0
without LM	15.74	25.10
with 4-grams LM	15.37	16.09

Benchmark CER result:

	COMMON VOICE 7.0	COMMON VOICE 8.0
without LM	9.51	9.95
with 4-grams LM	6.91	7.15

Evaluation

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

python eval.py --model_id vutankiet2901/wav2vec2-large-xlsr-53-ja --dataset mozilla-foundation/common_voice_7_0 --config ja --split test --log_outputs

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• 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: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
4.7776	4.73	1500	2.9540	0.9772	0.8489
1.9076	9.46	3000	0.7146	0.5371	0.2484
1.507	14.2	4500	0.5843	0.4689	0.2196
1.3742	18.93	6000	0.5286	0.4321	0.1988
1.2776	23.66	7500	0.5007	0.4056	0.1870
1.2003	28.39	9000	0.4676	0.3848	0.1802
1.1281	33.12	10500	0.4524	0.3694	0.1720
1.0657	37.85	12000	0.4449	0.3590	0.1681
1.0129	42.59	13500	0.4266	0.3423	0.1617
0.9691	47.32	15000	0.4214	0.3375	0.1587

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0