jun-han/Whisper-squeezeformer-v4

Whisper-squeezeformer-NSQU-whisper-sparse-A

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: 0.1860
• Wer: 9.1296

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: 3000
• training_steps: 36000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.89	1.0	3000	3.2878	114.5015
1.1579	2.0	6000	0.7947	42.0578
0.3888	3.0	9000	0.7379	36.9314
0.2242	4.0	12000	0.7417	35.9172
0.5221	5.0	15000	0.6811	32.7808
0.324	6.0	18000	0.6716	32.0457
0.2034	7.0	21000	0.6845	32.0073
0.2177	9.6	24000	0.1991	10.8624
0.127	10.8	27000	0.1856	10.5485
0.0909	12.0	30000	0.1838	9.5918
0.0785	13.2	33000	0.1849	9.1030
0.0595	14.4	36000	0.1860	9.1296

Framework versions

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