vutankiet2901/wav2vec2-xls-r-1b-ja

Model description

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - JA

Benchmark WER result:

	COMMON VOICE 7.0	COMMON VOICE 8.0
without LM	16.97	17.95
with 4-grams LM	11.77	12.23

Benchmark CER result:

	COMMON VOICE 7.0	COMMON VOICE 8.0
without LM	6.82	7.05
with 4-grams LM	5.22	5.33

Evaluation

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

pip install mecab-python3 unidic-lite pykakasi
python eval.py --model_id vutankiet2901/wav2vec2-xls-r-1b-ja --dataset mozilla-foundation/common_voice_8_0 --config ja --split test --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: 2000
num_epochs: 100.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.484	9.49	1500	1.1849	0.7543	0.4099
1.3582	18.98	3000	0.4320	0.3489	0.1591
1.1716	28.48	4500	0.3835	0.3175	0.1454
1.0951	37.97	6000	0.3732	0.3033	0.1405
1.04	47.47	7500	0.3485	0.2898	0.1360
0.9768	56.96	9000	0.3386	0.2787	0.1309
0.9129	66.45	10500	0.3363	0.2711	0.1272
0.8614	75.94	12000	0.3386	0.2676	0.1260
0.8092	85.44	13500	0.3356	0.2610	0.1240
0.7658	94.93	15000	0.3316	0.2564	0.1218

Framework versions

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

vutankiet2901
/

wav2vec2-xls-r-1b-ja