seniruk/whisper-medium-si-v2

Sinscribe Medium

This model is a fine-tuned version of openai/whisper-medium on the Sinhala CSV + FLACs dataset. It achieves the following results on the evaluation set:

• Loss: 0.0657
• Wer: 22.1361
• Wer Raw: 24.8976

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: 1e-05
• train_batch_size: 4
• eval_batch_size: 24
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Wer Raw
0.1247	0.2204	2000	0.1307	36.3910	39.8362
0.1074	0.4409	4000	0.1083	32.4505	35.8267
0.0921	0.6613	6000	0.0929	29.8234	33.2830
0.0844	0.8817	8000	0.0815	27.6916	31.2998
0.0619	1.1022	10000	0.0791	27.1318	30.5885
0.0594	1.3226	12000	0.0756	25.6891	28.9286
0.0633	1.5430	14000	0.0703	24.7416	28.2388
0.0558	1.7635	16000	0.0666	22.7175	25.9323
0.0558	1.9839	18000	0.0642	23.1697	26.4497
0.0292	2.2043	20000	0.0681	22.5452	25.7383
0.0337	2.4248	22000	0.0659	22.1576	25.0269
0.0297	2.6452	24000	0.0660	22.0284	24.8976
0.0292	2.8656	26000	0.0657	22.1361	24.8976

Framework versions

• Transformers 4.54.0
• Pytorch 2.3.0+cu121
• Datasets 3.6.0
• Tokenizers 0.21.4