FlowSynth: Instrument Generation Through Distributional Flow Matching and Test-Time Search
Accepted at ICASSP 2026. FlowSynth appeared in the AASP-P3: Music Generation I poster session in Barcelona, Spain.
Overview
FlowSynth addresses a central challenge in virtual instrument generation: maintaining a consistent timbre across different pitches and velocities. It combines distributional flow matching with test-time search to produce high-quality, playable instruments while preserving note-level control.
Key contributions
- Distributional flow matching models the velocity field as a learned probability distribution, capturing uncertainty instead of producing only a deterministic estimate.
- Confidence-weighted test-time sampling explores multiple generation trajectories and selects outputs that maximize timbre consistency.
- Music-specific search objectives improve consistency across an instrument’s pitch range while preserving prompt alignment and audio quality.
This work was completed during an internship at Smule.