deAPI-ai/whisper-large-v3-ct2

Whisper Large V3 CT2 — Word/Char Timestamps + Diarization Bundle

A single, self-contained ASR bundle: faster-whisper / CTranslate2 transcription of Whisper large-v3, plus wav2vec2 forced alignment for accurate per-word and per-char timestamps, plus optional pyannote speaker diarization. Everything is resolved locally from this directory — nothing is downloaded at runtime.

Contents

Path	Component	Purpose
/	Whisper large-v3 (CTranslate2, int8_float16-ready)	transcription engine
aligners/en	wav2vec2-large-xlsr-53-english	English word/char timestamps
aligners/pl	wav2vec2-large-xlsr-53-polish	Polish word/char timestamps
aligners/multi	wav2vec2-xlsr-multilingual-56	56-language fallback aligner
diarization/	speaker-diarization-community-1	speaker labels (who spoke when)

Licenses

This is a bundle of independently-licensed models. Each component keeps its own license:

Component	Source	License
Whisper large-v3 weights	openai/whisper-large-v3	MIT
CTranslate2 conversion format	OpenNMT/CTranslate2 (SYSTRAN)	MIT
English aligner	jonatasgrosman/wav2vec2-large-xlsr-53-english	Apache-2.0
Polish aligner	jonatasgrosman/wav2vec2-large-xlsr-53-polish	Apache-2.0
Multilingual aligner	voidful/wav2vec2-xlsr-multilingual-56	Apache-2.0
Diarization	pyannote/speaker-diarization-community-1	CC-BY-4.0 (weights) · pyannote.audio code MIT

Attribution (required): the diarization weights are licensed CC-BY-4.0 — commercial use is permitted with attribution to pyannote (Hervé Bredin et al.) and the speaker-diarization-community-1 model card. Keep this notice on any redistribution.

The bundle is made available by deAPI-ai on Hugging Face, configured for integration with GamerHash AI's platform. All credit for the underlying models belongs to their respective creators listed above.

Note on file formats

• model.bin is CTranslate2's own binary format (the transcription engine) — not a PyTorch pickle.
• The aligners are stored as safetensors.
• diarization/*/pytorch_model.bin are pyannote / PyTorch-Lightning checkpoints; they carry non-tensor hyperparameters that safetensors cannot hold, so they remain .bin (identical to the official pyannote/speaker-diarization-community-1 repo).