BARS: Beat-Adaptive Rap Synthesis via Flow Matching, accepted to ISMIR 2026

Excited to share that our work at BandLab Technologies, “BARS: Beat-Adaptive Rap Synthesis via Flow Matching,” has been accepted to ISMIR 2026!

Given lyrics, a beat track, and a short reference voice, BARS generates rap vocals that perform the lyrics over the beat using the timbre of the reference voice.

Unlike prior autoregressive approaches, BARS uses a non-autoregressive Flow Matching architecture, achieving a real-time factor of 0.027 on an NVIDIA RTX 4090 GPU. That’s about 1.6 seconds to generate one minute of rap, end-to-end. It also requires no explicit duration prediction or phoneme-level alignment during training, and allows independent control over rap tempo for flexible-length generation while maintaining rhythmic coherence with the beat.

Huge thanks to my wonderful co-authors Héctor Martel - 何可拓 and En Yan Koh for their help along the way.

The full paper will be available later. For now, we’ve put together a demo page where you can listen to examples generated by BARS.

See you at ISMIR 2026!