drepic/whisper-small-jp-ct2

This repository contains the CTranslate2 export of the fine-tuned model.

• Base Transformers model: drepic/whisper-small-jp
• Use with faster-whisper:

from faster_whisper import WhisperModel
model = WhisperModel("drepic/whisper-small-jp-ct2", device="cuda", compute_type="float16")

OTHER FINETUNES

• Want better accuracy? Try drepic/whisper-medium-jp-ct2

whisper-small-jp

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

• Loss: 0.6168
• Wer: 0.2600
• Cer: 0.2600

Model description

Better suited for transcribing japanese youtube content.

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: 5e-06
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• total_train_batch_size: 16
• total_eval_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.6589	1.0	7154	0.6615	0.2735	0.2735
0.6273	2.0	14308	0.6457	0.2699	0.2699
0.6251	3.0	21462	0.6359	0.2660	0.2660
0.6427	4.0	28616	0.6283	0.2642	0.2642
0.6389	5.0	35770	0.6243	0.2631	0.2631
0.6078	6.0	42924	0.6242	0.2615	0.2615
0.5788	7.0	50078	0.6195	0.2603	0.2603
0.5801	8.0	57232	0.6180	0.2596	0.2596
0.5866	9.0	64386	0.6145	0.2598	0.2598
0.6052	10.0	71540	0.6168	0.2600	0.2600

Framework versions

• Transformers 4.56.1
• Pytorch 2.8.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.0