Local speech-to-text using faster-whisper. 4-6x faster than OpenAI Whisper with identical accuracy; GPU acceleration enables ~20x realtime transcription. SRT/VTT/TTML/CSV subtitles, speaker diarization, URL/YouTube input, batch processing with ETA, transcript search, chapter detection, per-file language map.
Faster Whisper is a local speech-to-text skill built on faster-whisper, a CTranslate2 reimplementation of OpenAI's Whisper that the docs describe as running 4-6x faster with identical accuracy and, with GPU acceleration, around 20x realtime. It solves the problem of transcribing audio and video locally and offline (after model download), avoiding API costs while producing high-quality transcripts and subtitles for meetings, interviews, podcasts, lectures, and videos.
The skill wraps a bundled `transcribe` script exposing a broad flag surface: subtitle formats (SRT, VTT, ASS, LRC, TTML, plus CSV and HTML transcripts), speaker diarization with optional named speakers, URL and YouTube input that auto-downloads via yt-dlp, podcast RSS feed transcription, batch processing with ETA and skip-existing support, per-file language mapping, translation to English, 99+ language auto-detection, transcript search, chapter detection from silence gaps, per-speaker audio export, filler-word cleanup, channel selection, and audio preprocessing such as normalize and denoise. Its guidance stresses keeping invocations minimal by defaulting to the plain command and adding flags only when the user asks for that capability. Optional dependencies are ffmpeg and yt-dlp, and diarization can read a Hugging Face token from the local cache.
It targets developers, journalists, podcasters, and researchers who want private, cost-free, scriptable transcription and subtitling. Use cases span captioning videos, identifying speakers, translating foreign-language audio, searching recordings for a phrase, and bulk-transcribing mixed-language file sets.
Python 3 is required; ffmpeg and yt-dlp are optional (yt-dlp enables URL/YouTube download), and diarization can use a Hugging Face token cached at ~/.cache/huggingface/token. Supported platforms are Linux, macOS, and WSL2.
No cloud API is used; transcription runs locally and works offline after the model is downloaded. Internet is only needed for downloading models or for URL/YouTube/RSS input.
Plain text plus SRT, VTT, ASS, LRC, and TTML subtitles, CSV spreadsheet output, and HTML confidence-colored transcripts. Multiple formats can be produced at once.
Diarization labels who said what and can map real names to speaker slots; it supports 99+ languages with auto-detection, per-file language maps for batches, mixed-language audio, and translation to English.
The core rule is to use the default command as the fastest path and add flags such as --diarize, --format, or --translate only when the user explicitly asks for that capability.
Quick Setup:
.claude/skills/Repository
theplasmak/faster-whisper