Audio-to-MIDI in Your DAW, Honestly Offline

What it does NeuralNote is a VST3/AU/Standalone plugin that transcribes audio into MIDI without making you leave your DAW. Drop in a recording—or capture one live—and it renders polyphonic MIDI, complete with pitch-bend data, in an internal piano roll. You can tweak parameters, quantize to scale, and drag the finished MIDI onto a track.

The interesting bit The authors wrapped Spotify’s basic-pitch model for DAW duty, but the original CNN had to be split into four sequential chunks to fit through RTNeural—a process that required upstreaming 2D convolution support. Feature extraction still relies on ONNXRuntime. It is a pragmatic exercise in parking an offline model inside a real-time host while candidly admitting it cannot run in real time.

Key highlights

• Polyphonic transcription with pitch-bend detection for any tonal instrument, voice included
• Runs as a VST3/AU plugin or standalone app on Windows, macOS, and Linux
• Offline by necessity: the Constant-Q transform needs >1 second of audio, the CNN adds ~120 ms, and the note-decoding algorithm is non-causal
• Lets you adjust transcription settings and quantize to scale while auditioning the synthesized result
• The transcription engine (Lib/Model) is reusable in other projects under Apache-2.0

Caveats

• Real-time transcription is explicitly impossible; the architecture is fundamentally non-causal and requires long audio buffers
• Windows binaries are unsigned, so expect OS security friction
• The model-weight generation pipeline is manual and not included in the repository

Verdict Grab it if you want to salvage MIDI from recordings without switching applications. Look elsewhere if you need real-time transcription or a fully reproducible model-export pipeline.