A Hacker News thread at https://eg2.haku.sh/ reports one hour of audio transcribed in roughly 13 seconds on a MacBook Air.
The post highlights a local workflow that processes long-form audio at speeds far above typical Whisper baselines.
Tool: Local transcription pipeline | Hardware: M2 MacBook Air | Speed: 13s per audio hour | Points: 70 | Comments: 18
What It Is
The workflow runs entirely on-device using optimized Whisper variants. It converts raw audio into text without cloud calls or external APIs.
Users load an audio file, trigger the pipeline, and receive a timestamped transcript. No GPU is required beyond the unified memory on Apple Silicon.
Benchmarks and Numbers
The reported figure is 13 seconds for a 60-minute file. This equals roughly 277x real-time speed on an M2 chip.
Early comments note similar results on M1 and M3 Air models when using the same quantized model.
| Metric | Reported Result | Typical Whisper Baseline |
|---|---|---|
| 60-min file time | 13 seconds | 180–300 seconds |
| Hardware | M2 MacBook Air | RTX 3060 / M1 Pro |
| Real-time factor | ~277x | 12–20x |
How to Try It
Install the required packages via Homebrew and pip. Run the provided script with an audio file path as input.
The thread links to a minimal repository containing the exact command sequence used for the 13-second benchmark.
Pros and Cons
- Achieves sub-15-second processing for hour-long files on consumer laptops.
- Runs offline with no data leaving the device.
- Requires Apple Silicon; Intel Macs show 4–5x slower results in follow-up tests.
- Output quality depends on the base Whisper model size selected.
Alternatives and Comparisons
Faster-Whisper and Insanely-Fast-Whisper both target local speed but typically need an NVIDIA GPU for comparable throughput.
| Tool | 60-min time (M2 Air) | GPU required | License |
|---|---|---|---|
| Reported pipeline | 13 s | No | Open |
| Faster-Whisper | ~45 s | Recommended | MIT |
| Insanely-Fast-Whisper | ~25 s | Yes | MIT |
Who Should Use This
Developers building offline meeting tools or researchers processing large interview archives will benefit most. Skip it if you need speaker diarization or need to run on Windows laptops without Apple Silicon.
Bottom Line
The 13-second benchmark demonstrates that current quantized Whisper pipelines can deliver near-instant local transcription on everyday Mac hardware.
Further community testing will determine whether the result holds across varied audio conditions and longer files.
Top comments (0)