An opinionated CLI to transcribe Audio files (or YouTube videos) w/ Whisper on-device! Powered by MLX, Whisper & Apple M series
TL;DR - After our actual testing, the Whisper supported by MPS achieves speeds comparable to a 4090!
80 mins audio file only need 80s on APPLE M1 MAX 32G! ONLY 80 SECONDS
We've added a CLI to enable fast transcriptions. Here's how you can use it:
Install whisper-mps with pip:
# please install ffmpeg first: brew install ffmpeg
pip install whisper-mpsRun inference from any path on your computer:
# filetype should be wav/mp3/mp4 etc.
whisper-mps --file-name <filename>Run inference with a specific model size:
# for example, using the base model size. Other available models: "tiny", "base", "small", "medium", "large".
# Larger models require more loading time.
# filetype should be wav/mp3/mp4 etc.
whisper-mps --file-name <filename> --model-name baseSpecify the language to skip auto-detection (useful for short clips or lower-param models):
# Use an ISO 639-1 code (e.g. "en", "fr", "de", "zh") or the full language name (e.g. "french")
whisper-mps --file-name <filename> --language frRun inference from a YouTube URL on your computer:
Note
Deprecated: The YouTube URL inference feature is deprecated and may be removed in future releases.
Please uses yt-dlp to download YouTube videos.
# whisper-mps --youtube-url https://www.youtube.com/watch\?v\=jaM02mb6JFMNote
The CLI is highly opinionated and only works on Apple MPS.
The whisper-mps repo provides all-round support for running Whisper in various settings. More command-line support will be provided later.
--file-name FILE_NAME
Path or URL to the audio file to be transcribed.
--model-name MODEL_NAME
Size of the OPENAI Whisper model name, like tiny (default), base, small, etc.
--language LANGUAGE
Language spoken in the audio (e.g. "en", "fr", "de", "zh", or full names like
"french"). Skips automatic language detection. Helpful when auto-detection fails
on short clips or with lower-param models.
--youtube-url URL_ADDRESS
The YouTube video URL. (Deprecated)
--output-file-name OUTPUT_FILE_NAME
The output file name for the transcribed text JSON.
--log-level LOG_LEVEL
Optional. ["DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"]