devkubnet/whisper-th-distill-ct2

Thai Whisper (distil large-v3) — CTranslate2

A CTranslate2 (float16) conversion of biodatlab/distill-whisper-th-large-v3 (MIT) for faster-whisper, made for Devkub Edit Kit. All credit for the model goes to biodatlab; their original model card is in ORIGINAL_MODEL_CARD.md.

Includes the model's own tokenizer.json (large-v3 vocabulary of 51,866 tokens; without it faster-whisper falls back to whisper-tiny's tokenizer, which shifts the language ids).

Decoding note

The model was fine-tuned without timestamp tokens: transcribe speech-trimmed pieces of ≤ 20 s cut at pauses rather than relying on Whisper's 30-second walk. On four real Thai clips (CPU, int8) this gave a character error rate of about 3.3% versus 8.5% for large-v3-turbo.