ksoky/whisper-small-km

Whisper Small Km - Kak Soky

This model is a fine-tuned version of openai/whisper-small on the SLR42 dataset. It achieves the following results on the evaluation set:

• Loss: 0.1471
• Wer: 35.6654

Model description

The model was fine-tuned on both encoder-decoder of transformer-based.

Intended uses & limitations

The training data is limited, thus the performance is also limited to only reading speech and a limited domain (tourism).

Training and evaluation data

The training and evaluation data was split in a 9:1 ratio from Google Text-to-speech corpus.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 4000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3639	0.76	1000	0.3452	71.9392
0.1553	1.53	2000	0.2025	49.0494
0.0565	2.29	3000	0.1664	39.9240
0.0334	3.06	4000	0.1471	35.6654

Framework versions

• Transformers 4.25.0.dev0
• Pytorch 1.12.1+cu113
• Datasets 2.7.0
• Tokenizers 0.13.2