Jingmiao/whisper-small-chinese_base

Whisper Small Chinese Base

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

• Loss: 0.3573
• Wer: 16.6439

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: 64
• eval_batch_size: 32
• 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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0005	76.0	1000	0.3573	16.6439
0.0002	153.0	2000	0.3897	16.9749
0.0001	230.0	3000	0.4125	17.2330
0.0001	307.0	4000	0.4256	17.2451
0.0001	384.0	5000	0.4330	17.2300

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2