VocalRender

VocalRender: Score-Native Singing Voice Synthesis for Real-World Composition

• Paper
• Code and documentation
• Online inference demo
• Open-source training dataset

VocalRender is a score-native singing voice synthesis (SVS) model designed for real-world composition. It synthesizes 48 kHz singing directly from lyrics, MIDI pitches, symbolic note values, and tempo through an original combination of an interleaved lyric--note representation, continuous acoustic latents, and autoregressive diffusion modeling. It does not require phoneme-level durations, an explicit duration predictor, or a time-aligned acoustic reference.

For musicians: try it without writing code

You do not need to understand machine learning or install the model to try VocalRender. Open the online VocalRender demo and work with familiar score formats in your browser.

What VocalRender does

VocalRender sings a melody that you provide. You choose the lyrics, notes, rhythm, tempo, and a short voice reference; the model renders them as a singing voice. It is a score renderer rather than an automatic composer: it does not write the melody or arrange an accompaniment for you.

What to prepare

• A melody: paste ABC notation, upload an ABC file, or upload MusicXML from notation software such as MuseScore. You can also enter notes manually.
• Chinese lyrics: embedded ABC w: or MusicXML lyrics can be imported. If the score has no note-aligned lyrics, type or paste them in the lyrics box.
• A voice reference: select one of the included voices, or upload 2–8 seconds of clean, unaccompanied singing that you have permission to use. The clip supplies the vocal color; it does not need to be the same song.

Score-to-song workflow

1. Open Import ABC notation or MusicXML, provide your score, and press Parse score.
2. Check the suggested vocal part, lyric line, and measure range. The demo preselects a short range that fits the model; you can change it.
3. Press Load selected range into editor. Review the lyrics, pitches, durations, and BPM, then press Generate Singing.

In the editor, MIDI pitch 60 means middle C (C4), and pitch 0 means a rest. + Melisma note lets one lyric syllable continue across multiple notes. SP represents a rest or breath.

The released checkpoint currently works best with Mandarin Chinese and a single vocal melody. Chord names in ABC, such as "C" or "G7", are treated as accompaniment labels and ignored, but actual simultaneous notes are not supported. For piano or ensemble scores, select a monophonic vocal/melody part. The demo reports unsupported passages instead of silently guessing which note should be sung.

How it works

VocalRender combines three components:

1. A score-native interleaved representation serializes BPM followed by each lyric syllable and its associated (pitch, note-value) pairs. This retains lyric-to-note alignment and supports melisma.
2. An Audio VAE represents singing as compact continuous acoustic latents.
3. An autoregressive diffusion model generates the latent sequence patch by patch and decides when to stop. The VAE decoder renders the result as a waveform.

Checkpoints

This repository contains two variants with the same architecture, parameter count, and speech-pretrained base-model initialization. Their only difference is the training recipe (training corpus and schedule):

Variant	Repository path	Training data and strategy	Paper results
VocalRender	VocalRender/	CrawlSinger-OS (>2,300 h); 40k-step synthetic pretraining followed by 20k-step real-data finetuning	Stronger subjective score following (MS-MOS 2.96)
VocalRender-Pro	VocalRender-Pro/	CrawlSinger (>5,600 h of in-house real singing); 160k training steps	Better intelligibility, speaker similarity, naturalness, and OOD robustness; MS-MOS 2.71

VocalRender-Pro reduces WER from 4.44 to 3.88 and improves speaker similarity from 0.922 to 0.929 on Opencpop. On CrawlSinger-Eval, WER changes from 4.52 to 4.45 and similarity from 0.919 to 0.926. The paper relates these improvements to the larger amount of real singing and broader singer coverage. VocalRender's higher score-following score may reflect more reliable score annotations in its real-data finetuning subset.

Each directory contains the model weights, AudioVAE, model configuration, and extended SVS tokenizer required by the inference code. Both checkpoints have the same model size; each is approximately 9.5 GB to download.

Installation

git clone --recurse-submodules https://github.com/pymaster17/VocalRender.git
cd VocalRender
uv sync
source .venv/bin/activate

Download

Download either checkpoint into the repository's pretrained_models/ directory:

# VocalRender
hf download pymaster/VocalRender \
    --include "VocalRender/*" \
    --local-dir pretrained_models

# Or VocalRender-Pro
hf download pymaster/VocalRender \
    --include "VocalRender-Pro/*" \
    --local-dir pretrained_models

Quick inference

The GitHub repository includes three ready-to-use prompt/score pairs selected from OpenCpop. This command runs the first demo:

python scripts/infer_vocalrender_svs_single.py \
    --ckpt_dir pretrained_models/VocalRender \
    --json_file examples/opencpop_demo.json \
    --item_name 2003000087 \
    --prompt_audio examples/prompt_audio/2003000081.wav \
    --output demo_2003000087.wav

Additional bundled pairs are 2017000646 with prompt 2017000644.wav, and 2044001652 with prompt 2044001666.wav. The prompt clips are 4.19-6.17 seconds long and do not overlap their paired target segment. Clone the GitHub repository to obtain the scripts and examples. The excerpts remain subject to the OpenCpop terms.

Use pretrained_models/VocalRender-Pro as --ckpt_dir to run the Pro model. Full generation requires a CUDA-capable compute node. Prompt audio is required: the released checkpoints were trained with prompt audio on every sample (prompt_audio_prob=1.0), so prompt-free inference is unsupported and may substantially degrade quality. Use a clean 2-8 second singing clip, which also provides the target timbre. See the repository README for batch-inference options.

Input format

{
  "item_name": "demo",
  "word": ["我", "的", "孤", "独"],
  "pitch": [65, 64, 64, 65, 67, 65],
  "note": ["<NOTE_8>", "<NOTE_32>", "<NOTE_16>", "<NOTE_16>", "<NOTE_16>", "<NOTE_8>"],
  "pitch2word": [0, 1, 2, 2, 2, 3],
  "bpm": 64
}

pitch2word[i] identifies the lyric index associated with note i, allowing one lyric syllable to span multiple notes. word_dur and pitch_dur are optional fields used only for visualization and evaluation; they are not model inputs.

Limitations

• The released checkpoints primarily target Mandarin Chinese singing.
• Checkpoint loading and generation require substantial disk space, system memory, and GPU memory.
• Output quality depends on the musical validity and lyric-to-note alignment of the input score.

Model foundation

The released implementation initializes its ARDM backbone from the speech-pretrained VoxCPM2 weights. The score-native representation, SVS formulation, data pipeline, and singing training recipe are introduced by VocalRender.

Citation

@article{chen2026vocalrender,
  title   = {VocalRender: Score-Native Singing Voice Synthesis for Real-World Composition},
  author  = {Chen, Yukun and Wang, Tianrui and Mu, Zhaoxi and Yang, Xinyu and Chng, EngSiong},
  journal = {arXiv preprint arXiv:2607.27768},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.27768}
}

License

Apache License 2.0. See the repository's LICENSE file for details.