whisper-small-aishell

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

• Loss: 0.1770
• Wer: 0.4068

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

train data：aishell train test data：aishell test

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 512
• total_eval_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.042	4.26	1000	0.1227	0.3990
0.0134	8.52	2000	0.1312	0.4004
0.0042	12.78	3000	0.1402	0.4027	0.051
0.0022	17.04	4000	0.1479	0.4045
0.001	21.3	5000	0.1568	0.4069
0.0007	25.56	6000	0.1568	0.3990
0.0004	29.82	7000	0.1644	0.4037
0.0003	34.08	8000	0.1697	0.4045
0.0002	38.34	9000	0.1751	0.4072
0.0002	42.6	10000	0.1770	0.4068

Framework versions

• Transformers 4.28.0.dev0
• Pytorch 1.13.0
• Datasets 2.10.1
• Tokenizers 0.13.2