MuseTimbre: Zero-Shot Timbre Transfer by Controlling a Frozen Music Generator

Yuan-Chiao Cheng (Music Informatics Group, Georgia Institute of Technology) and Zhiyao Duan (University of Rochester)

Instrument timbre transfer from an audio reference clip, driven by two plug-in control modules on a pretrained music generator: a pitch module that reads a binary piano roll through decoupled cross-attention, and a timbre module that reads a fine-tuned CLAP embedding through adaptive layer norm.

Zero-shot timbre transfer

Each row takes the notes of the source clip and the timbre of the reference clip. Source and reference come from different datasets and different instruments; none of the recordings was seen in training. Baselines that read the timbre from audio: CTD, SS-VQ-VAE, TokenSynth-A. Baselines that name the target instrument by text: TokenSynth-T, MuseControlLite, DDIM-inv, DDPM-inv. Headphones recommended.

source → targetSourceReferenceMuseTimbre (ours)CTDSS-VQ-VAETokenSynth-ATokenSynth-TMuseControlLiteDDIM-invDDPM-inv
1
bassoon

piano
2
electric guitar

piano
3
electric guitar

horn
4
viola

electric guitar
5
viola

saxophone
6
clarinet

viola
7
clarinet

electric guitar
8
electric guitar

clarinet
9
double bass

clarinet
10
clarinet

flute
11
flute

violin
12
piano

violin

MIDI score input and text control

The pitch module accepts a symbolic score in place of a transcription, and the generator's original text conditioner still works on top of the transferred timbre. The last column adds the text prompt shown under the player.

exampleSource audioReferenceOurs, pitch from audioOurs, pitch from MIDI scoreOurs, MIDI + text prompt
Guitar to electric piano
guitar (GOAT)

electric piano
“in a large reverberant concert hall”
Guitar to organ
guitar (GOAT)

organ
“with heavy rain and distant thunder in the background”
Violin to distorted guitar
violin (URMP)

distorted guitar
“in a huge cathedral with a long reverb tail”
Flute to electric piano
flute (URMP)

electric piano
“lo-fi vintage radio recording, warm and distant”

Paper

Submitted to ICASSP 2027. Link and BibTeX will appear here.