whisper-large-v2-anime

This model is a fine-tuned version of clu-ling/whisper-large-v2-japanese-5k-steps on joujiboi/japanese-anime-speech (https://huggingface.co/datasets/joujiboi/japanese-anime-speech) dataset.

Model description

Whisper large v2 finetuned with japanese common voice 11 dataset (clu-ling) which was then fine tuned with transcripts from joujiboys dataset. First of three. *next model 10k steps with anime dataset. Possibly add another dataset.

Intended uses & limitations

Intended for alternative forms of media where language and pitch often deviate from mainstream. (anime tv jav etc)

Training and evaluation data

joujiboi/japanese-anime-speech mozilla-foundation/common_voice_11_0

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 3000

Training results

Qualitative eval. Improved Japanese to English translation over whisper-large-v2 and whisper-large-v2-japanese-5k-steps.

Framework versions

• PEFT 0.8.2
• Transformers 4.30.2
• Pytorch 2.2.1+cu118
• Datasets 2.17.0
• Tokenizers 0.13.3