ReVoiceAI-W2V-BERT-Thai-IPA

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2532
• Wer: 0.0847
• Cer: 0.0276

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: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 30

Training results

Training Loss	Epoch	Step	Cer	Validation Loss	Wer
1.6842	1.0	1154	0.1430	0.7152	0.5133
0.4338	2.0	2308	0.0879	0.3823	0.2990
0.3128	3.0	3462	0.0705	0.2929	0.2408
0.2398	4.0	4616	0.0571	0.2670	0.1937
0.2061	5.0	5770	0.0576	0.2584	0.1878
0.1834	6.0	6924	0.0514	0.2383	0.1687
0.147	7.0	8078	0.0475	0.2141	0.1547
0.1412	8.0	9232	0.0465	0.2291	0.1530
0.1187	9.0	10386	0.0438	0.1988	0.1400
0.1055	10.0	11540	0.0466	0.2579	0.1581
0.1001	11.0	12694	0.0400	0.2064	0.1304
0.0923	12.0	13848	0.0374	0.2115	0.1188
0.081	13.0	15002	0.0379	0.2080	0.1196
0.0734	14.0	16156	0.0347	0.1915	0.1110
0.06	15.0	17310	0.0371	0.2192	0.1169
0.0576	16.0	18464	0.0363	0.1919	0.1156
0.0495	17.0	19618	0.0345	0.1928	0.1073
0.0493	18.0	20772	0.0338	0.1889	0.1089
0.0396	19.0	21926	0.0321	0.1961	0.0994
0.0371	20.0	23080	0.0319	0.1946	0.0992
0.0312	21.0	24234	0.0313	0.2144	0.0972
0.0263	22.0	25388	0.0314	0.2076	0.0972
0.0224	23.0	26542	0.0309	0.2165	0.0958
0.0202	24.0	27696	0.0301	0.2221	0.0924
0.0165	25.0	28850	0.0314	0.2358	0.0955
0.016	26.0	30004	0.0296	0.2357	0.0913
0.0111	27.0	31158	0.0296	0.2265	0.0906
0.0079	28.0	32312	0.0289	0.2357	0.0892
0.0081	29.0	33466	0.2474	0.0861	0.0279
0.0053	30.0	34620	0.2532	0.0847	0.0276

Framework versions

• Transformers 4.53.0
• Pytorch 2.7.1+cu126
• Datasets 3.6.0
• Tokenizers 0.21.2