How it’s built
DawIt starts with a person’s ideas about making music and learning it. AI helps write the code. Audio tools, local models and checks turn those ideas into something you can hear.
Here’s who does what, what runs on your computer, and the open building blocks behind it.
The explanation is part of the instrument.
The decisions and the doing
What should DawIt teach? Which tool would make an idea easier to hear? How should it explain that to a beginner? Those are human decisions.
A person chooses the features, their order and the way each guide explains its subject. They review what comes back, try it and ask for changes when it needs work.
The aim stays the same: help you make music and understand your choices.
Claude Code and OpenAI’s Codex help write and revise the app, its tests, its theory text and this website. A person reviews the work and decides what belongs.
AI-written code is checked with the same sound measurements and app reviews described below.
The tools used to build DawIt are separate from its optional musical tools. You can explore guides, write notes and use the built-in effects without running a language model.
See what the optional models do →Inside DawIt
C++20 and JUCE provide the desktop app, audio, MIDI and plugin hosting.
DawIt’s effects and synthesisers use signal processing to make and shape sound. The theory guides and live pictures help explain what those tools are doing.
See how sound flows through a song →The song holds the tracks, clips, key, chords, words and sections.
Instruments, effects, buses and sends turn the parts into the mix.
Meters, pictures and guide topics read the song or settings as you work.
Optional AI inside the app
Different jobs use different models. DawIt runs these tools locally. Some models are included; others download when you choose to use them. Model downloads need a connection.
Finds drums, bass, vocals and the other instruments. Runs through demucs.cpp.
DOWNLOAD ON FIRST USE · 84 MBTurns pitched audio into editable MIDI notes. ONNX Runtime runs the model.
BUILT IN · 230 KBFinds words in English vocals, through whisper.cpp. Choose Small or Medium.
DOWNLOAD ON FIRST USE · 488 MB / 1.5 GBInstall Ollama and a model separately. DawIt starts with qwen2.5:7b as its default model choice.
SEPARATE INSTALL · MODEL OF YOUR CHOICEONNX Runtime runs the voice’s models. English only. No voice comes with DawIt; each voice and vocoder has terms from its maker.
BRING YOUR OWN VOICEMusic processing stays on your computer. Ollama connects to your own machine by default; its server address is configurable.
Checks for a music app
The test suite checks things you can hear: levels, pitches, harmonics, timing and decay. The interface and explanations need their own review too.
Tests feed notes and audio through the tools, then measure what comes out. That checks whether a filter, compressor, synth or reverb behaves as expected.
The project’s review workflow opens windows and guides in DawIt and inspects app screenshots for layout, text and the right state.
Theory sources are recorded with the project. Numbers in explanations are checked against tests, standards or established references.
Credits & technical details
Open a group for the tools, libraries and models behind DawIt, including their recorded versions, makers and licences.
The language
Fast enough for audio, and the language of JUCE and of audio plugins.
8.0.9 · AGPLv3 or commercial
The whole app: the windows and controls, audio and MIDI, plugin hosting, files, JSON and HTTP, on macOS, Windows and Linux. DawIt Tune finds a voice's pitch with McLeod's method and moves it with PSOLA, plain DSP on JUCE's own FFT, so it needs no library of its own.
Steinberg · SIL Open Font License 1.1 · pinned commit
The music font that draws clefs, rests, flags and note heads in the score view and its PDF. It is the reference font of the SMuFL standard, downloaded at a pinned commit and checked by its hash. Its licence text comes with the app.
3.7.12, Steinberg · VST3 licence or GPLv3
Loading VST3 plugins in a track's chain, and building DawIt's own effects and synths as VST3 plugins (under the GPLv3). It comes with JUCE.
Apple · Apache 2.0
Building DawIt's own effects and synths as Audio Unit plugins on macOS. It comes with JUCE.
Apple · part of macOS
Loading AU plugins on a Mac.
Sevag Hanssian · MIT
Runs the Demucs model in C++, for Split into stems and Remove vocals.
MPL 2.0
The matrix maths demucs.cpp is built on.
1.30.0, Microsoft · MIT
Runs the Basic Pitch model, for Convert to MIDI, and the models of a singing voice, for Sing this.
1.9.5 · MIT
Runs the Whisper models in C++, for Detect lyrics.
Carnegie Mellon University · free for any use, with its origin named · 3.6 MB
How an English word sounds, with about 135,000 entries for about 126,000 words. It is built into the app, for Sing this and for finding rhymes and counting syllables in the lyrics window.
OpenVPI · Apache 2.0 · the model format only
The format that singing voices are exported in. DawIt links none of its code: it runs a voice's ONNX models with ONNX Runtime, for Sing this. No voice comes with DawIt. You add one yourself, and each voice (and its vocoder) has its own terms.
MIT
An open-source editor for singing voices. DawIt follows it as the reference for how a voice is laid out and in what order its models are called. No OpenUtau code is used.
MIT · a separate program
Runs a language model on your computer for the Generate panel. You install it yourself; DawIt talks to it on your machine.
Meta AI research · MIT · 84 MB
Splits a recording into drums, bass, vocals and the rest. The weights come from the demucs.cpp conversion on Hugging Face.
0.4.0, Spotify · Apache 2.0 · 230 KB
Turns audio into notes. Small enough to be built into the app.
OpenAI · MIT · 488 MB and 1.5 GB
Finds the words in English vocals. The weights come from the whisper.cpp conversion on Hugging Face.
Alibaba's Qwen team · Apache 2.0 · about 4.7 GB
The Generate panel's default model, through Ollama. Choose another installed chat model in the Generate panel if you prefer.
Anthropic
The coding tool the AI works in: it reads the project, writes the code, builds it, runs the tests and the app.
OpenAI
A coding agent used to read the project, implement features, run the checks and maintain the website.
BSD 3-Clause and Apache 2.0
CMake configures the app and tests; Ninja builds them. Versions are recorded in the build configuration, with checksums for the downloaded source archives and models that specify them.
OpenAI
Created the DawIt logo, the open book with five rising bars shown at the top of every page.
GPL 2.0
Keeps the history of every change, each with a sentence on what it does.
JUCE 8.0.9 is offered under AGPLv3 or a commercial licence. DawIt’s plugin builds use the bundled VST3 SDK under GPLv3; the Bravura music font uses the SIL Open Font License 1.1. Models, voices and vocoders retain their makers’ terms.
For the available way to build and run the app, see Get started.
And this website?
Static HTML, one shared stylesheet and a small script for navigation, search and colour themes. System fonts keep the pages self-contained. The images show DawIt itself.
There are no website accounts or analytics. ChatGPT created the book-and-levels logo; the app screenshots come from DawIt.
See the learning experience →Back to the music
Explore the guides, try a tool and follow the explanation.
