noflm/whisper-base-ja-elite

Whisper Base Japanese Elite

This model is a fine-tuned version of openai/whisper-base on the Elite35P-Server/EliteVoiceProject twitter dataset. It achieves the following results on the evaluation set:

• Loss: 0.4385
• Wer: 17.0732

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: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 200
• training_steps: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0002	111.0	1000	0.2155	9.7561
0.0001	222.0	2000	0.2448	12.1951
0.0	333.0	3000	0.2674	13.4146
0.0	444.0	4000	0.2943	15.8537
0.0	555.0	5000	0.3182	17.0732
0.0	666.0	6000	0.3501	18.9024
0.0	777.0	7000	0.3732	16.4634
0.0	888.0	8000	0.4025	17.0732
0.0	999.0	9000	0.4178	20.1220
0.0	1111.0	10000	0.4385	17.0732

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.8.1.dev0
• Tokenizers 0.13.2