What is yt-transcribe?

yt-transcribe is a small command-line script that downloads a YouTube video and converts its speech into a text transcript, saving the result as a Markdown file. The entire process runs locally on your machine, no audio is sent to any cloud transcription service. It ties together two tools: yt-dlp, which downloads just the audio track from YouTube, and FluidAudio, Apple's Core ML-based speech recognition tool built for Apple Silicon chips (the M-series processors in modern Macs). Core ML is Apple's machine learning framework that runs models directly on the device hardware. The workflow is: grab the video title via yt-dlp, download only the audio in m4a format and cache it in /tmp/yt-transcribe so re-running the same URL skips the download, then pass the audio to FluidAudio's command-line tool for transcription. The transcript is wrapped in a small Markdown header and saved as a file named after the video title. The default model (v2) handles English. Passing --model-version v3 switches to a multilingual model. Useful extra flags include --output-json for a full transcript with per-word timings, --word-timestamps for word-level timing data, and --custom-vocab to bias the recognizer toward domain-specific vocabulary. The first run downloads Core ML model weights from Hugging Face (about 1 GB) and caches them locally, subsequent runs are fully offline. Requirements are macOS 14 or later on Apple Silicon and a Swift toolchain. Installation is via Homebrew (recommended) or manually by cloning and building FluidAudio from source.