whisper-ctranslate2 - local faster-whisper transcription CLI with word-level timestamps

check: whisper-ctranslate2 https://github.com/Softcatala/whisper-ctranslate2

whisper-ctranslate2 — a Whisper-compatible command-line client built on faster-whisper (CTranslate2). Runs locally on CPU or CUDA with no API key and writes word-level timestamps to JSON, which the captions capability's default provider uses for word-timed burned captions. Installed with uv tool install, pinned to an exact version; Whisper models download from Hugging Face on first use (small ≈ 480 MB, cached afterwards).

Key Commands

Command	Description
`whisper-ctranslate2 vo.wav --model small --word_timestamps True --output_format json --output_dir out`	Transcribe with per-word start/end times (`segments[].words[]`)
`whisper-ctranslate2 vo.wav --language en --initial_prompt "Octomind, HeyGen"`	Fix the language and bias spelling of names and jargon
`whisper-ctranslate2 talk.wav --vad_filter True --output_format srt`	Skip silence with Silero VAD and write a sentence-level SRT
`whisper-ctranslate2 vo.wav --compute_type int8 --device cpu`	Low-memory CPU run

Common Usage

bash

# 16 kHz mono WAV in, word-timed JSON out (./video-out/<slug>/captions/vo.json)
ffmpeg -i final-mix.mp4 -vn -ar 16000 -ac 1 captions/vo.wav
whisper-ctranslate2 captions/vo.wav --model small --language en \
  --word_timestamps True --output_format json --output_dir captions

# Larger model for noisy creator footage or non-English speech
whisper-ctranslate2 captions/vo.wav --model large-v3-turbo --word_timestamps True --output_format json --output_dir captions

Model guide: small / small.en for clean voiceover (fast on a laptop CPU), medium or large-v3-turbo for noisy footage, accents, or non-English. Whisper drops some filler words ("um", "uh"); an --initial_prompt written with fillers makes it keep more of them when you need to cut them.

Links

View on GitHub