• Home
  • About
    • Breno Brito photo

      Breno Brito

      Breno Brito is an AI engineer, data scientist, and writer covering Bitcoin, finance, language models, talks, and applied AI projects.

    • Learn More
    • Twitter
    • LinkedIn
    • Github
  • Writing
  • Projects
  • Talks & Media
  • Tags

fast-transcript

03 May 2026

Reading time ~1 minute

fast-transcript is a local CLI for transcribing large lectures quickly on a laptop.

On the development machine, it handled 30 minutes in 2!*

* Benchmark run on a MacBook Pro M1. The exact long-run measurement was 29m47s of lecture audio transcribed in about 2m14s (13.38x real-time).

The project packages an ONNX Parakeet workflow behind a very short command:

fscript lecture.mp3

Or install it with Homebrew:

brew install brenorb/fast-transcript/fast-transcript

On Apple Silicon, that Homebrew install now pours a real Homebrew bottle, so it does not need a local Rust toolchain.

That command is enough to:

  • fetch the default model if needed
  • normalize the audio with ffmpeg when the input is not already in the expected format
  • transcribe with long-audio defaults tuned for unattended runs
  • write a JSON transcript next to the source audio

Why I built it

I wanted a tool for 30-minute to 2-hour classes and lectures that could run locally, stay fast, and avoid the usual friction of manual conversion steps and fragile command lines.

In local benchmarks on the same development machine, fast-transcript outperformed the other workflows I tested for this specific use case, including mlx-whisper, insanely-fast-whisper, and the tuned parakeet-mlx path.

What shaped the implementation

This project was strongly informed by:

  • Handy
  • GLaDOS
  • transcribe-rs

The core direction came from the ONNX Parakeet path used in Handy and the packaging ideas surfaced in both Handy and GLaDOS.

Links

GitHub Repo



aiaudiotranscription Tweet