wav2vec2-kanji-base-char-0916

This model is a fine-tuned version of rinna/japanese-wav2vec2-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.0080
• Cer: 0.3084
• Wer: 0.999

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: 5e-05
• train_batch_size: 4
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 77380
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer	Wer
3.5786	1.0	19348	3.4423	0.5827	1.0
1.5754	2.0	38696	2.3013	0.5027	1.0
1.2976	3.0	58044	2.1926	0.4518	1.0
1.1784	4.0	77392	1.8865	0.4352	1.0
1.1072	5.0	96740	1.8018	0.4392	1.0
1.0466	6.0	116088	1.5333	0.4289	1.0
0.9887	7.0	135436	1.4297	0.4036	1.0
0.9497	8.0	154784	1.5539	0.4024	1.0
0.9075	9.0	174132	1.6214	0.4160	1.0
0.8721	10.0	193480	1.3556	0.3964	1.0
0.8398	11.0	212828	1.1912	0.3772	1.0
0.8069	12.0	232176	1.1510	0.3529	0.999
0.7788	13.0	251524	1.0752	0.3399	0.999
0.7477	14.0	270872	1.0926	0.3373	0.999
0.7198	15.0	290220	0.9936	0.3101	0.999
0.7028	16.0	309568	1.0239	0.3123	0.999
0.6806	17.0	328916	1.0006	0.3081	0.999
0.6733	18.0	348264	1.0348	0.3156	0.999
0.6709	19.0	367612	1.0075	0.3086	0.999
0.6651	20.0	386960	1.0080	0.3084	0.999

Framework versions

• Transformers 4.42.0
• Pytorch 2.3.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1