W2V2-BERT-Malayalam

This model is a fine-tuned version of facebook/w2v-bert-2.0 on an these datasets: IMASC, MSC, OpenSLR Malayalam Train split, Festvox Malayalam, common_voice_16_1 It achieves the following results on the evaluation set:

• Loss: 0.1722
• Wer: 0.1299

Training procedure

Trained on NVIDIA A100 GPU

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• 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: 500
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.1416	0.46	600	0.3393	0.4616
0.1734	0.92	1200	0.2414	0.3493
0.1254	1.38	1800	0.2205	0.2963
0.1097	1.84	2400	0.2157	0.3133
0.0923	2.3	3000	0.1854	0.2473
0.0792	2.76	3600	0.1939	0.2471
0.0696	3.22	4200	0.1720	0.2282
0.0589	3.68	4800	0.1768	0.2013
0.0552	4.14	5400	0.1635	0.1864
0.0437	4.6	6000	0.1501	0.1826
0.0408	5.06	6600	0.1500	0.1645
0.0314	5.52	7200	0.1559	0.1655
0.0317	5.98	7800	0.1448	0.1553
0.022	6.44	8400	0.1592	0.1590
0.0218	6.9	9000	0.1431	0.1458
0.0154	7.36	9600	0.1514	0.1366
0.0141	7.82	10200	0.1540	0.1383
0.0113	8.28	10800	0.1558	0.1391
0.0085	8.74	11400	0.1612	0.1356
0.0072	9.2	12000	0.1697	0.1289
0.0046	9.66	12600	0.1722	0.1299

Framework versions

• Transformers 4.39.3
• Pytorch 2.1.1+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1