SPEAK-ASR/whisper-si-exp-10

This model is a fine-tuned version of openai/whisper-small on the SPEAK-ASR/openslr-sinhala-asr-norm-noise-rem-preprocessed dataset. It achieves the following results on the evaluation set:

• Loss: 0.1739
• Wer: 12.6991

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: 0.0006926233815806411
• train_batch_size: 128
• eval_batch_size: 256
• seed: 42
• optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 295
• num_epochs: 20.0

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2032	1.0	473	0.1951	22.2112
0.1473	2.0	946	0.1551	18.1276
0.1243	3.0	1419	0.1471	16.4627
0.1083	4.0	1892	0.1319	14.6748
0.0976	5.0	2365	0.1371	15.8588
0.0883	6.0	2838	0.1271	14.3492
0.0823	7.0	3311	0.1247	13.4447
0.0738	8.0	3784	0.1311	13.9303
0.0675	9.0	4257	0.1251	13.3315
0.0602	10.0	4730	0.1254	13.0560
0.0534	11.0	5203	0.1285	12.9820
0.0466	12.0	5676	0.1330	12.8602
0.0395	13.0	6149	0.1382	13.2603
0.0350	14.0	6622	0.1410	12.7947
0.0282	15.0	7095	0.1477	12.6035
0.0228	16.0	7568	0.1557	12.7520
0.0176	17.0	8041	0.1637	12.7514
0.0137	18.0	8514	0.1739	12.6991

Framework versions

• PEFT 0.18.1
• Transformers 5.0.0
• Pytorch 2.9.0.dev20250821+rocm7.0.0.git125803b7
• Datasets 4.5.0
• Tokenizers 0.22.2