Legal

Licences and attribution

Rooster is built on open source work by others. This page credits the code and models it ships and carries a copy of every licence they are used under. Rooster itself processes your recordings on your Mac.

1. Parakeet TDT 0.6B v3

   Transcribes your recordings into the script you edit.

   By NVIDIA

   Changes: Converted to Core ML by FluidInference. Rooster uses the converted models unmodified.

   CC BY 4.0 · Source

2. FluidAudio

   Runs the Parakeet models on the Apple Neural Engine.

   By FluidInference

   Apache 2.0 · Source

3. whisper.cpp

   Transcribes recordings when you choose Local Whisper in Settings, for languages Parakeet does not cover.

   Copyright (c) 2023-2026 The ggml authors

   MIT · Source

4. DeepFilterNet

   Removes background noise when you turn on Studio Voice.

   Copyright (c) 2021 Hendrik Schröter

   MIT · Source

5. fastcluster

   Hierarchical clustering inside FluidAudio. It is compiled into Rooster, which does not call it.

   Copyright © 2011 Daniel Müllner; changes from version 1.1.24 on © Google Inc.

   Changes: Wrapped in C for Swift by FluidInference as FluidAudio's FastClusterWrapper.

   BSD 2-Clause · Source

6. VBx

   Speaker clustering inside FluidAudio. It is compiled into Rooster, which does not call it.

   Copyright 2021-2024 BUT Speech@FIT

   Changes: Reimplemented in Swift by FluidInference as part of FluidAudio.

   Apache 2.0 · Source

7. misaki

   Turns Japanese text into pronunciations inside FluidAudio. It is compiled into Rooster, which does not call it.

   By hexgrad

   Changes: Parts ported to Swift by FluidInference as part of FluidAudio.

   Apache 2.0 · Source

8. cutlet

   Turns Japanese text into romaji inside FluidAudio. It is compiled into Rooster, which does not call it.

   Copyright (c) 2020 Paul O'Leary McCann

   Changes: Ported to Swift by FluidInference, by way of misaki, as part of FluidAudio.

   MIT · Source

9. Convert-Numbers-to-Japanese

   Writes numbers as Japanese words inside FluidAudio. It is compiled into Rooster, which does not call it.

   Copyright (c) 2018 David Wilson

   Changes: Ported to Swift by FluidInference, by way of misaki, as part of FluidAudio.

   MIT · Source

10. Sparkle

    Installs updates in the copy downloaded from rooster.cam. The Mac App Store version carries no part of it.

    Copyright (c) 2006-2013 Andy Matuschak and the Sparkle contributors

    Changes: Unused Downloader XPC service removed before embedding. Not otherwise modified.

    MIT · Source

11. LAME

    Writes MP3 exports in the copy downloaded from rooster.cam. The Mac App Store version carries no part of it.

    Copyright (c) 1999-2011 The LAME Project

    LGPL 2.0 · Source

LAME’s source

Rooster ships LAME as a separate library, libmp3lame.0.dylib inside the app, and opens it by name rather than building it in. To use your own build of LAME, replace that file and sign the app again yourself: macOS opens a library inside a signed app only when the two signatures match. The terms of use permit that change. LAME’s source is served here exactly as Rooster builds it: the LAME 4.0 archive, unmodified, and the script that builds it.

Transcription models you supply

Rooster can transcribe with a Whisper model file that you download and choose yourself. Those model files are not distributed with Rooster, so they are covered by whatever licence you obtained them under.

Corrections

If a credit here is wrong or incomplete, please tell us and we will correct it.