faster-whisper
by theplasmak
Speech & TranscriptionBrief
Local speech-to-text using faster-whisper.
Description
What This Skill Does
Runs OpenAI Whisper locally via a CTranslate2 backend, achieving 4-6x faster transcription with identical accuracy. GPU acceleration pushes this to ~20x realtime. Outputs SRT, VTT, ASS, LRC, TTML, CSV, and plain text; supports speaker diarization, YouTube/URL input, podcast RSS feeds, transcript search, and chapter detection.
Runs entirely offline after the initial model download, eliminating API costs, while matching OpenAI Whisper accuracy at 4-6x the speed.
When to Use It
- Transcribing a recorded meeting with speaker labels
- Generating SRT subtitles for a video file
- Fetching and transcribing all episodes from a podcast RSS feed
- Searching a long audio file for when a specific phrase was mentioned
- Translating a foreign-language interview to English text
Example Workflow
User asks: transcribe this meeting recording and identify who is speaking
- 1
Runs ./scripts/transcribe meeting.wav --diarize
- 2
Assigns speaker labels (SPEAKER_1, SPEAKER_2) to each turn using pyannote.audio
- 3
Outputs a labeled transcript with timestamps for each speaker segment
Plain text transcript with per-turn speaker labels and timestamps
Requirements
- HuggingFace account and access token (for speaker diarization)
- Accepted model use agreements at hf.co/pyannote/speaker-diarization-3.1 (for diarization)
- Accepted model use agreements at hf.co/pyannote/segmentation-3.0 (for diarization)
Installation
clawhub install theplasmak/faster-whisperRun the CLI command, or paste the prompt into Claude Code, Cursor, Codex or whichever coding agent you use.