Large Mambo Model

a voice-and-humming co-pilot that drives a DAW

A producer co-pilot you talk and hum to. It takes a single mixed utterance - spoken commands interleaved with a hummed melody, like “give me something like ♪hmm-hmm♪ but slower, and kick the drums up” - and turns it into structured musical intent that drives REAPER over MIDI. The trick is a segment-and-route architecture: a joint acoustic-and-linguistic router splits the mic stream into speech, melody, and percussion spans, sends each to a specialist decoder (faster-whisper for speech, a pYIN pitch tracker for the humming), fuses them into a versioned Utterance IR, and only then lets an LLM planner write a replayable action plan. Deliberately no LLM in the perception path, so it runs at near-zero cost on off-the-shelf parts. There’s a paper alongside the code. Source on GitHub.