whisper-medium-si-ta-en-split1

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

• Loss: 0.2412
• Wer Si: 34.6399
• Cer Si: 7.4300
• Wer Ta: 27.6527
• Cer Ta: 7.6979
• Wer En: 5.0041
• Cer En: 3.1183
• Wer: 22.4322
• Cer: 6.0821

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: 0.0002
• train_batch_size: 4
• eval_batch_size: 8
• seed: 3407
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_8BIT 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: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Cer	Cer En	Cer Si	Cer Ta	Validation Loss	Wer	Wer En	Wer Si	Wer Ta
0.246	0.2279	1000	8.4308	3.2317	12.1226	9.9381	0.2259	31.5085	5.1075	53.7830	35.6348
0.2012	0.4558	2000	7.3432	3.0479	9.9484	9.0333	0.2043	27.2708	4.8801	44.7584	32.1738
0.1664	0.6837	3000	6.6231	2.9148	8.3529	8.6015	0.1906	24.9899	4.7146	38.6509	31.6043
1.0048	0.9116	4000	0.2638	47.1285	10.4020	31.2451	8.7173	4.9835	3.0479	27.7857	7.3890
0.7362	1.1397	5000	0.2519	39.1978	8.2277	29.7380	8.4446	4.9628	3.0831	24.6329	6.5851
0.7605	1.3676	6000	0.2638	35.6427	7.8680	28.4851	7.9633	4.9214	3.0987	23.0164	6.3100
0.6403	1.5955	7000	0.2412	34.6399	7.4300	27.6527	7.6979	5.0041	3.1183	22.4322	6.0821

Framework versions

• PEFT 0.20.0
• Transformers 4.56.2
• Pytorch 2.11.0+cu128
• Datasets 3.6.0
• Tokenizers 0.22.2