hubert-japanese-base-noise-0426

This model is a fine-tuned version of rinna/japanese-hubert-base on the audiofolder dataset. It achieves the following results on the evaluation set:

• Loss: 0.2302
• Cer: 0.0598
• Wer: 0.992

Model description

This model is a hiragana recognition model created by the proposed method.
The model is based on rinna's hubert base model.

Intended uses & limitations

More information needed

Training and evaluation data

Train : noisepaused_JNAS_train_0408
Test : noisepaused_JNAS_test_0408

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 12500.0
• num_epochs: 25

Training results

Training Loss	Epoch	Step	Validation Loss	Cer	Wer
11.9556	1.0	2500	9.5354	0.9998	1.0
3.8038	2.0	5000	3.6912	0.9998	1.0
1.668	3.0	7500	1.1310	0.2733	1.0
0.688	4.0	10000	0.4272	0.1880	1.0
0.4959	5.0	12500	0.3254	0.1356	0.998
0.4275	6.0	15000	0.2856	0.1026	1.0
0.3647	7.0	17500	0.2720	0.0884	0.998
0.346	8.0	20000	0.2625	0.0848	0.998
0.3273	9.0	22500	0.2646	0.0896	0.996
0.301	10.0	25000	0.2479	0.0734	0.996
0.2871	11.0	27500	0.2466	0.0778	0.998
0.268	12.0	30000	0.2403	0.0717	0.992
0.2494	13.0	32500	0.2467	0.0705	0.994
0.2336	14.0	35000	0.2411	0.0702	0.994
0.2347	15.0	37500	0.2352	0.0662	0.994
0.2261	16.0	40000	0.2400	0.0708	0.996
0.207	17.0	42500	0.2341	0.0652	0.996
0.2018	18.0	45000	0.2340	0.0635	0.994
0.196	19.0	47500	0.2323	0.0578	0.992
0.1856	20.0	50000	0.2343	0.0625	0.992
0.1788	21.0	52500	0.2303	0.0597	0.992
0.1821	22.0	55000	0.2285	0.0596	0.99
0.1824	23.0	57500	0.2305	0.0591	0.99
0.1693	24.0	60000	0.2297	0.0598	0.99
0.1807	25.0	62500	0.2302	0.0598	0.992

Framework versions

• Transformers 4.39.3
• Pytorch 2.2.2
• Datasets 2.18.0
• Tokenizers 0.15.1