jun-han/Whisper-squeezeformer-v3

Whisper-squeezeformer-NSQU-whisper

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

• Loss: 0.1322
• Wer: 5.6642

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: 20
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2500
• training_steps: 50000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.8718	1.0	2500	3.8609	111.8590
2.5628	2.0	5000	0.2978	15.6193
0.1698	3.0	7500	0.2218	11.0906
0.0867	4.0	10000	0.2011	10.1891
0.1697	5.0	12500	0.1641	8.9851
0.0993	6.0	15000	0.1553	7.8039
0.0651	7.0	17500	0.1555	7.2448
0.0468	8.0	20000	0.1569	7.1497
0.2168	9.0	22500	0.1509	7.0507
0.1467	10.0	25000	0.1494	6.9671
0.1113	11.0	27500	0.1493	6.7597
0.0914	12.0	30000	0.1511	6.8035
0.1946	13.0	32500	0.1391	6.4212
0.1425	14.0	35000	0.1369	5.8753
0.1145	15.0	37500	0.1368	5.7536
0.1776	16.0	40000	0.1302	5.5995
0.1416	17.0	42500	0.1298	5.6204
0.1239	18.0	45000	0.1297	5.6204
0.3373	19.0	47500	0.1353	5.7403
0.2785	20.0	50000	0.1322	5.6642

Framework versions

• Transformers 4.45.2
• Pytorch 2.4.0
• Datasets 3.1.0
• Tokenizers 0.20.0