transcribe-whisper-local

name	transcribe-whisper-local
description	Transcribe audio locally using Whisper (offline, no cloud). Use when the user asks for a local transcription, offline transcription, privacy-preserving transcription, or explicitly requests Whisper.

Transcribe (local Whisper)

Offline transcription with Whisper. Requires setup-whisper to have been run once.

Preferred implementation

faster-whisper (CTranslate2 backend) — 4x faster than openai-whisper at similar accuracy.

uv run --with faster-whisper python <<'PY'
from faster_whisper import WhisperModel
model = WhisperModel("medium", device="cpu", compute_type="int8")
segments, info = model.transcribe("INPUT.wav", beam_size=5, vad_filter=True)
for s in segments:
    print(f"[{s.start:.2f}s -> {s.end:.2f}s] {s.text}")
PY

Model selection

• base / small → fast, decent quality, good default for English
• medium → good balance (default)
• large-v3 → best quality, slow on CPU

Config whisper_model in config.json (default medium).

GPU

Daniel's workstation has an AMD GPU. faster-whisper supports ROCm via CTranslate2 builds but it's fragile; default CPU unless user has set whisper_device: "cuda" / "rocm".

Output convention

<source_stem>.whisper.md — timestamped plain text. <source_stem>.whisper.json — segments with start/end/text for downstream use.

If not set up

If which faster-whisper fails, tell the user to run /setup-whisper first.