VocalRender — AI vocal synthesis from a musical score

VocalRender is an open-source vocal synthesis system that generates singing directly from MIDI notes, text, and BPM without explicit duration or F0 curve specifications, built on OpenBMB's VoxCPM2.

Compiled by Sergey KostenchukPublished 2026-08-12Updated 2026-09-18

2026-08-12 Multimodal HuggingFace

🎤 VocalRender — vocal synthesis directly from a musical score

Open-weights system based on OpenBMB's VoxCPM2: takes MIDI notes, syllable text, and BPM — without explicit phoneme durations or F0 curves. Two weights (~9.5 GB), Apache-2.0.

🌍 Eliminates the main "bottleneck" in SVS — explicit duration and F0 prediction. CMOS +0.42, WER 3.88% on Opencpop.

👤 Demo on HuggingFace, weights are open, runs on GPU.

Source: https://arxiv.org/abs/2607.27768 GitHub: https://github.com/pymaster17/VocalRender

Sources