AI · Tip
mlx_whisper local transcription benchmark & usage
Fast local audio transcription benchmark and Python snippet for Apple Silicon Macs.
This note provides quick benchmark figures and minimal Python code for running OpenAI Whisper models locally on Apple Silicon using the mlx_whisper library. It demonstrates that three hours of audio can be transcribed in under five minutes on high-end chips like M2 Ultra and M3 Max using the whisper-large-v3-turbo checkpoint. It serves as a handy reference for setting up efficient, private speech-to-text without cloud APIs.
How it works
Question: how long to transcribe 3 hours of video locally on macOS with mlx_whisper?
- M2 Ultra: 4:34 mins
- M3 Max: 4:56 mins
Install: pip install mlx_whisper
Usage:
import mlx_whisper
text = mlx_whisper.transcribe(
"path/to/audio",
path_or_hf_repo="mlx-community/whisper-large-v3-turbo")["text"]
print(text)
Caveat from the author: some repetitions appear in the transcription, but this is due to the model itself, not the tool.
What it's for
- Transcribe long audio recordings locally without cloud API costs
- Estimate speech-to-text performance when sizing hardware on Apple Silicon
- Integrate private transcription workflows into local Python scripts
Key points
- Speed benchmarks across M2 Ultra and M3 Max hardware
- Minimal two-line Python snippet using the mlx_whisper package
- Direct compatibility with Hugging Face MLX community models
- Note on model-level repetition quirks with whisper-large-v3-turbo
Pricing
Open source. Free and open source
Details
- Type
- Tip
- Category
- AI
- Topic
- Text to Speech
- Open source
- Yes
- Platforms
- macOS, Code library
- For
- Developers and macOS users needing fast, local speech-to-text transcription.
- Alternative to
- whisper.cpp, OpenAI Whisper API, faster-whisper
- Tags
- whisper, mlx, apple silicon, speech to text, audio, transcription, python
Why it was saved
- @ivanfioravanti: Shows mlx_whisper can transcribe 3 hours of video locally in under 5 minutes on Apple Silicon, and gives the minimal code to do it.
Related
AI voice generator, voice cloning, and conversational agent platform
Ultra-low-latency streaming text-to-speech and instant voice cloning model.
Open-source 1.6B text-to-speech model generating multi-speaker dialogue in one pass.
Fast, low-cost text-to-speech API and studio for creators and developers
Converts spoken thoughts into structured, polished writing in customizable styles.
Official Python library for Cartesia's real-time text-to-speech API
Updated