SPEAK-ASR/whisper-si-exp-5

SPEAK-ASR/whisper-si-fullds-tm4

This model is a fine-tuned version of openai/whisper-small on the SPEAK-ASR/openslr-sinhala-asr-preprocessed-1 | SPEAK-ASR/openslr-sinhala-asr-preprocessed-2 | SPEAK-ASR/openslr-sinhala-asr-preprocessed-3 | SPEAK-ASR/youtube-sinhala-asr-preprocessed dataset. It achieves the following results on the evaluation set:

• Loss: 0.1704
• Wer: 21.6000

Model description

• WandB ID: ruby-deluge-67
• Used full OpenSLR dataset + Our Youtube dataset
• Added Youtube data test split to eval set
• Used LoRA adapters
• Previous name was SPEAK-ASR/whisper-si-fullds-tm4
• Used WER to choose the best model

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: 3e-05
• train_batch_size: 32
• eval_batch_size: 256
• seed: 42
• optimizer: Use OptimizerNames.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: 200
• num_epochs: 5.0

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1940	0.6380	3000	0.2560	30.9152
0.1719	1.2761	6000	0.2136	26.7417
0.1846	1.9141	9000	0.1999	25.0293
0.2162	2.5521	12000	0.1869	23.3360
0.1663	3.1901	15000	0.1790	22.4653
0.2172	3.8282	18000	0.1738	21.9250
0.2025	4.4662	21000	0.1704	21.6000

Framework versions

• PEFT 0.18.1
• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2