devkubnet/whisper-th-distill-mlx

Thai Whisper (distil large-v3) — MLX

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

Files

File
config.json, weights.safetensors	mlx-whisper model (float16)
alignment_heads.json	word-timing heads from the source generation_config.json — mlx-whisper's loader ignores them, so apply after loading: model.set_alignment_heads(np.array(heads))
devkub.json	decoding hint used by Devkub Edit Kit

Decoding note

The model was fine-tuned without timestamp tokens, so Whisper's usual 30-second walk drops the end of each window. Transcribe speech-trimmed pieces of ≤ 20 s (cut at pauses, no long trailing silence) one at a time. On four real Thai clips this gave a character error rate of about 3% versus 4.5% for whisper-large-v3-turbo, with matching word timestamps.