KeefeBuild/keefe-whisper-small-v6-ct2

Keefe Whisper Small v6 — CT2 (CTranslate2 / faster-whisper)

Accelerated inference build of KeefeBuild/keefe-whisper-small-v6 converted with ct2-transformers-converter for faster-whisper.

• Architecture: Whisper small (Transformer encoder-decoder)
• Conversion: CTranslate2 / faster-whisper
• Weights: float16 → int8 quantised for CPU efficiency
• Language: multilingual (99 languages)
• Source repo: KeefeBuild/keefe-whisper-small-v6

Files

file	description
config.json	CTranslate2 Whisper model config
model.bin	quantised weights
vocabulary.json	CTranslate2 vocabulary
tokenizer.json	fast-tokenizer vocabulary (from source)
preprocessor_config.json	audio feature extractor config (from source)

Usage

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel("KeefeBuild/keefe-whisper-small-v6-ct2",
                     device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.wav", beam_size=1)
print("".join(s.text for s in segments))

Performance

• The int8 CPU build is ~250 MB and transcribes 11 s of speech in under 10 s on an idle host, vs. the Transformers fp16 GPU fallback.

Limitations

• Synthetic-trained: model is not yet proven on real user voice.
• English-first: while multilingual (99 languages), quality is optimised for English.
• Language guard required: auto-detection can be low-confidence on short clips (e.g. a 0.19 kn probability on the JFK sample). Prefer passing language="en" when the language is known, or rely on the language_probability_threshold guard wired up in keefe/config.py.
• Production guards recommended: silence guard, language guard, and hallucination guard are wired into keefe/stt.py and keefe/config.py.
• Generation length: use max_new_tokens for length control — never max_length.
• Fallback: openai/whisper-large-v3 (the fallback path) is heavy and impractical on CPU-only deployments; enable only on GPU-capable hosts.

Production guards

Guard	Location	Purpose
Silence guard	keefe/utils.py (is_silence)	Drops near-silent audio before STT.
Language guard	keefe/stt.py (_transcribe_ct2)	Defaults to "en" when detection confidence < 0.85.
Hallucination guard	keefe/stt.py (_transcribe_ct2)	Drops segments where start >= duration.

This repo is not licensed for external use.