wav2vec2phone-large-xlsr-jp-jdrtfw07-demo3

This model is a fine-tuned version of jonatasgrosman/wav2vec2-large-xlsr-53-japanese on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.0451
• Wer: 0.025
• Cer: 0.0195

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: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 4
• gradient_accumulation_steps: 4
• 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: 1640
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.0002	1.0	328	2.8482	0.9965	0.9980
2.2134	2.0	656	2.0964	0.8864	0.9374
1.6625	3.0	984	1.2925	0.7960	0.7917
1.0846	4.0	1312	0.4955	0.3292	0.3089
0.8593	5.0	1640	0.2443	0.1061	0.0839
0.7502	6.0	1968	0.1552	0.0620	0.0504
0.6342	7.0	2296	0.1202	0.0491	0.0422
0.6012	8.0	2624	0.1008	0.0442	0.0379
0.6214	9.0	2952	0.0846	0.0390	0.0333
0.5486	10.0	3280	0.0756	0.0361	0.0309
0.5633	11.0	3608	0.0669	0.0310	0.0261
0.4977	12.0	3936	0.0609	0.0300	0.0251
0.4863	13.0	4264	0.0556	0.0282	0.0233
0.4416	14.0	4592	0.0533	0.0279	0.0234
0.4713	15.0	4920	0.0499	0.0266	0.0214
0.4715	16.0	5248	0.0489	0.0252	0.0197
0.4724	17.0	5576	0.0472	0.0259	0.0205
0.4835	18.0	5904	0.0463	0.0259	0.0204
0.4589	19.0	6232	0.0454	0.0252	0.0199
0.4207	20.0	6560	0.0451	0.025	0.0195

Framework versions

• Transformers 4.31.0
• Pytorch 2.0.1+cu118
• Datasets 2.13.1
• Tokenizers 0.13.3