o0dimplz0o/Whisper-Large-v3-turbo-STT-Zeroth-KO-v2

Whisper Large v2 Fine-Tuned for Korean ASR

This model is being fine-tuned from openai/whisper-large-v3-turbo on a custom dataset. It currently achieves the following results on the evaluation set (still fine-tuning):

• Loss: 0.0164
• Wer: 19.9134
• Cer: 0.0660

Model Description

This model is a version of openai/whisper-large-v3-turbo, currently still being incrementally fine-tune in stages, specifically optimized for Korean automatic speech recognition (ASR) tasks. The fine-tuning process aims to deliver high accuracy and timestamped transcriptions for Korean speech.

Dataset Details

• Dataset Source: Custom dataset (https://huggingface.co/datasets/o0dimplz0o/Zeroth-STT-Korean)
• Number of Samples: 102,263
• Split: 93% train, 7% test
• Data Augmentation: 20% random, applied only to the training set

Training Details

• Hardware: L40S GPU
• Learning Rate Scheduler: Cosine
• Epochs: [pending completion]
• Optimizer: AdamW Torch Fused