scribe

State-of-the-art local audio transcription with speaker diarization for macOS.

100% local. No cloud. No API keys. No data leaves your machine.

Features

• Transcription — Accurate speech-to-text powered by OpenAI Whisper (via WhisperKit)
• Speaker diarization — Identify who said what, powered by pyannote (via SpeakerKit)
• Apple Silicon optimized — Runs on CoreML and the Apple Neural Engine
• Multiple output formats — Plain text, JSON (with word timestamps), SRT, VTT
• 99 languages — Supports all languages that Whisper supports
• Fast — Processes audio faster than real-time on Apple Silicon

Install

brew install theam/tap/scribe

Or build from source:

git clone https://github.com/theam/scribe.git
cd scribe
swift build -c release
cp .build/release/scribe /usr/local/bin/

Usage

Basic transcription

scribe transcribe meeting.wav

With speaker diarization

scribe transcribe meeting.wav --diarize

Specify number of speakers (improves accuracy)

scribe transcribe meeting.wav --diarize --speakers 4

Output formats

scribe transcribe meeting.wav --format txt    # plain text (default)
scribe transcribe meeting.wav --format json   # structured JSON with word timestamps
scribe transcribe meeting.wav --format srt    # SRT subtitles
scribe transcribe meeting.wav --format vtt    # WebVTT subtitles

Save to file

scribe transcribe meeting.wav --format json --output transcript.json

Choose a model

scribe transcribe meeting.wav --model large-v3         # best accuracy
scribe transcribe meeting.wav --model large-v3-turbo   # best speed/accuracy (default)
scribe transcribe meeting.wav --model small             # fastest, lower accuracy

Force a language

scribe transcribe meeting.wav --language es    # Spanish
scribe transcribe meeting.wav --language fr    # French

Manage models

scribe models list                       # see available/downloaded models
scribe models download large-v3-turbo    # download a model
scribe models remove large-v3            # remove a downloaded model

Output Examples

Plain text with diarization

[00:00:12] Speaker 1: Hello everyone, welcome to the meeting.
[00:00:18] Speaker 2: Thanks for joining. Let's start with the agenda.

JSON

{
  "metadata": {
    "duration": 3612.5,
    "diarization": true
  },
  "segments": [
    {
      "start": 12.0,
      "end": 16.5,
      "text": "Hello everyone, welcome to the meeting.",
      "speaker": "Speaker 1",
      "words": [
        { "start": 12.0, "end": 12.4, "text": "Hello" }
      ]
    }
  ]
}

Requirements

• macOS 14 (Sonoma) or later
• Apple Silicon (M1 or later) recommended — Intel Macs fall back to CPU inference

Models

The default model (large-v3-turbo, ~632MB) is downloaded automatically on first use. Other models can be downloaded with scribe models download.

Model	Size	Speed	Accuracy	Languages
tiny	~75MB	Fastest	Lower	99
base	~142MB	Very fast	Fair	99
small	~466MB	Fast	Good	99
medium	~1.5GB	Moderate	Very good	99
large-v3-turbo	~632MB	Fast	Very good	99
large-v3	~3.1GB	Slower	Best	99

Acknowledgments

scribe is built on the shoulders of excellent open-source projects:

• OpenAI Whisper (Apache 2.0) — The speech recognition model that powers transcription
• WhisperKit (MIT) by Argmax — CoreML implementation of Whisper for Apple Silicon
• SpeakerKit (MIT) by Argmax — CoreML speaker diarization
• pyannote.audio (MIT) by Herve Bredin — The diarization model architecture that SpeakerKit builds on
• swift-argument-parser (Apache 2.0) by Apple — CLI argument parsing

License

Apache 2.0 — Copyright 2026 The Agile Monkeys Inc. See LICENSE.