SiRoZaRuPa/wav2vec2-kanji-base-unigram-0914

wav2vec2-kanji-base-unigram

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: 1.6841
• Cer: 0.4426
• Wer: 1.0

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
5.0463	1.0	19348	7.1017	0.9448	0.999
2.3465	2.0	38696	4.1177	0.6757	1.0
1.6622	3.0	58044	4.1171	0.6136	1.0
1.478	4.0	77392	2.8624	0.5836	1.0
1.3639	5.0	96740	3.2231	0.5709	1.0
1.2924	6.0	116088	3.6346	0.5807	1.0
1.2191	7.0	135436	2.4840	0.5444	1.0
1.1705	8.0	154784	2.5931	0.5403	1.0
1.117	9.0	174132	2.4750	0.5263	1.0
1.0689	10.0	193480	2.1332	0.5131	1.0
1.0371	11.0	212828	2.0070	0.4903	1.0
0.9922	12.0	232176	2.0196	0.4847	1.0
0.948	13.0	251524	1.7148	0.4597	1.0
0.9143	14.0	270872	1.5651	0.4506	1.0
0.8841	15.0	290220	1.7433	0.4530	1.0
0.8593	16.0	309568	1.7191	0.4528	1.0
0.8388	17.0	328916	1.6598	0.4455	1.0
0.8304	18.0	348264	1.6671	0.4434	1.0
0.8219	19.0	367612	1.6866	0.4426	1.0
0.8184	20.0	386960	1.6841	0.4426	1.0

Framework versions

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