What it does

NeuralNote is an audio effect plugin (VST3, AU, or standalone app) that turns audio into MIDI inside a DAW. Version 2 runs Kyutai and Mirelo's MuScriptor transcription model locally, replacing the earlier Basic Pitch model, and adds multi-instrument transcription for full mixes plus a built-in multi-instrument synth for previewing results (README, MuScriptor).

You record live or drop an audio file onto the plugin, choose the instruments (or Automatic), and press Transcribe. The piano roll fills in as decoding proceeds, and you can drag the MIDI to a MIDI track or save it as a file. Transcription runs in the background and can be paused and resumed, even after reopening the plugin or DAW session (usage). It is aimed at producers and musicians who want to recover notes from recordings without uploading audio anywhere.

Getting started

1. Download the installer for your platform from the website download page or the releases page. macOS has signed and notarized installers for Apple Silicon and Intel; the Windows x64 installer is not code signed and may trigger a warning. Linux prebuilt binaries are not available yet (install notes).
2. Install, then open the plugin or standalone app and download a model on first launch. The transcription model is not bundled.
3. In the plugin, record the track or drop an audio file (.wav, .aiff, .flac, .mp3, .ogg; .m4a on macOS).
4. Select instruments or Automatic, click Transcribe, and optionally audition the result through the built-in synth.
5. Drag the MIDI into a MIDI track or save it to disk.

Model sizes and hardware

MuScriptor is available in small, medium, and large, downloaded and managed from the Model button in the top bar (models and performance). Downloads are roughly 210 MB, 620 MB, and 2.7 GB respectively. Models live under ~/Library/NeuralNote/models on macOS, %APPDATA%\NeuralNote\models on Windows, and ~/.config/NeuralNote/models on Linux; files placed there by hand are only picked up if their name and size match the official downloads.

The GPU is used through Metal on macOS and Vulkan on Windows/Linux; Auto, a specific GPU, or CPU can be chosen under Settings > Compute device. On an Apple M1 Pro the README reports roughly 3.5x real time for small on Metal and 2x on CPU, 1.5x for medium on Metal, and 0.5x for large on Metal, with CPU not recommended for large. Only a few GPUs have been tested, and users are asked to report hardware issues (performance notes).

Building from source

Requirements are git, CMake, a C++23 compiler, Python 3, and internet access on the first configure, which fetches ggml and the synth soundfonts (build instructions). Clone with submodules, then configure and build:

git clone --recurse-submodules https://github.com/DamRsn/NeuralNote
cd NeuralNote
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

Artefacts are written to build/NeuralNote_artefacts/Release/. macOS GPU builds need Xcode's Metal toolchain; Windows GPU builds need the Vulkan SDK, and ASIO support requires signing Steinberg's ASIO license agreement or configuring with -DNEURALNOTE_ASIO=OFF (build instructions).

Reusing the engine

The transcription engine lives in a separate repository, muscriptor.cpp, a C++/ggml port of MuScriptor linked here as a submodule. It accepts 16 kHz mono float32 audio and returns note events; the public API is documented in docs/API.md.

Licensing

NeuralNote's code is Apache-2.0, but the MuScriptor model weights are not open source: they are distributed under CC BY-NC 4.0 and may only be used non-commercially (license, CC BY-NC 4.0). Plan commercial use accordingly.

HotRepo's take

NeuralNote moved to MuScriptor in v2, a transformer with 103M–1.4B parameters versus fewer than 17K for Basic Pitch, and extended transcription from single instruments to full mixes. The app is still fully local, which matters for musicians who prefer not to upload session audio. HotRepo recorded it at rank 11 on the C++ daily list on 2026-11-03 with 17 stars gained that day, and a 41-point Hacker News thread in May 2026 (2 comments) is too sparse to draw conclusions from. With only one snapshot and no long-running star series in the collected data, there is not enough history here to describe a trend.