Whisper Small v3 Burmese - Robust Noisy Training

This model is a fine-tuned version of myatsu/whisper-small-burmese-v2 on the Google FLEURS Burmese + Kaggle Noise (minsithu/audio-noise-dataset) dataset. It achieves the following results on the evaluation set:

Loss: 0.1440
Wer: 90.6395
Cer: 52.9501

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: 8
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 16
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: 100
training_steps: 1000
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.3093	0.4057	100	0.2936	97.5581	59.4269
0.2107	0.8114	200	0.2187	95.4360	56.5003
0.1442	1.2150	300	0.1853	95.0	55.5726
0.1207	1.6207	400	0.1638	93.2558	54.6075
0.0976	2.0243	500	0.1559	92.2384	53.5169
0.0784	2.4300	600	0.1503	92.0930	52.9128
0.0795	2.8357	700	0.1444	91.6570	52.9089
0.0591	3.2394	800	0.1439	90.6977	52.6892
0.0492	3.6450	900	0.1440	90.6395	52.9501
0.0462	4.0487	1000	0.1428	91.25	52.8520