---
title: "Tencent Open-Sourced a Song Model Built to Keep Voice and Band Together"
date: 2026-11-17
category: Generative Media
site: NeuroAI
canonical: https://neuroai.site/a/na-aigc-tencent-songgeneration-open-source
language: en
---

# Tencent Open-Sourced a Song Model Built to Keep Voice and Band Together

> Tencent AI Lab's SongGeneration goes open source with a design that models vocals and accompaniment on separate tracks.

A vocal line lands a fraction ahead of the drums. The singer sounds convincing, the backing track sounds convincing, and together they sound like two recordings spliced by hand. That is the specific flaw Tencent AI Lab set out to address when it open-sourced SongGeneration in June 2025.

The release arrived with a research paper, working weights and a demo. It also arrived with an unusual honesty about what the team was solving: not whether a machine can make a song, but whether it can make one that stays musically together from beginning to end.

## The model has two names

SongGeneration is the name of the open-source project. The paper behind it is titled *LeVo: High-Quality Song Generation with Multi-Preference Alignment*, posted to arXiv in June 2025. The author list reads like a small consortium: researchers from Tsinghua University's Shenzhen International Graduate School and Tencent AI Lab, with collaborators from Wuhan University, the Chinese University of Hong Kong, Shenzhen, Shanghai Jiao Tong University and Nanjing University. Tencent AI Lab released the code on GitHub and published a demo on Hugging Face.

For readers tracking the field, the two names matter. "SongGeneration" is what developers download. "LeVo" is the framework described in the paper and the term you will find in citations.

## Why separating the tracks is hard

Most song-generation systems have to make a choice. They can predict a single stream of audio tokens representing voice and instruments mixed together, which keeps the two in sync but limits how much detail each can carry. Or they can predict two separate streams, one for vocals and one for accompaniment, which raises fidelity but lets the two drift apart.

LeVo tries to have both. Its language model, LeLM, models two types of token in parallel: mixed tokens that represent the combined audio, and dual-track tokens that encode vocals and accompaniment separately. Two decoder-only transformers handle the streams, and a modular extension-training strategy is used to stop one token type interfering with the other. The paper then applies direct preference optimization (DPO) to align output with several kinds of human preference at once, using a semi-automatic process to build the training data.

The plain-language version: the system has one eye on cohesion and one eye on detail, and a post-training step nudges it toward what listeners actually prefer.

## What is in the release

Tencent says the open version supports text control, multi-track synthesis and style following. The stated use cases run from short-video soundtracks and game audio to virtual-performer shows and advertising. The compatibility list includes Chinese and English pop, hip-hop, guofeng (古风) style and electronic music, with multilingual lyric input and one-click generation among the advertised functions.

- **Weights and code:** released on GitHub under the Tencent AI Lab account, with an online demo on Hugging Face.

- **Paper:** arXiv:2506.07520, with audio samples linked by the authors.

- **Evaluation:** the paper's subjective test was built with a research team at the Communication University of China (中国传媒大学), whose School of Music and Recording Arts contributed the listening methodology.

That last point is more than a footnote. Subjective scoring in music generation is easy to run badly. Bringing in a music school to design the listening test is a reasonable way to make the numbers mean something.

## The claim, and how to read it

The paper reports that LeVo outperforms existing open-source methods on both objective and subjective metrics, while performing competitively with industry systems. Those are the authors' words, and "competitively" is doing careful work in that sentence. It is not a claim to beat every commercial product, and it should not be upgraded into one.

For a global reader, the significance is about access rather than victory. A full song model with published weights lets a developer in any country inspect the architecture, fine-tune it on local material and run it without per-second API fees. That matters more in markets where music licensing and language coverage are uneven, and where a one-size-fits-all commercial service will not fit the local sound.

The harder problem is still ahead. Generating a pleasing forty-second song is different from producing a catalogue with consistent artist identity, cleared lyrics and predictable rights. None of that is solved by an open checkpoint.

## Honest limitations

This piece is based on Tencent's official developer article, the public arXiv paper and Chinese technology-press coverage. I have not generated songs with the released model, so all performance statements are the authors'. The paper was still listed as a preprint under review at the time of the release, and peer review may change what stands. I have not independently verified the listening-test methodology or reproduced the scores, and I have not compared SongGeneration against commercial services on the same prompts. Licensing terms for generated audio deserve a separate look; this article does not resolve who owns an output track.

## What readers can do now

- Download the weights from the Tencent AI Lab GitHub repository and run a short A/B test: the same lyrics in the same style, twice. Listen for drift between voice and accompaniment near the end of the song — that is the failure mode this model targets.

- If you build audio tools, read the paper alongside the code. The dual-token design is the part worth understanding, not the headline benchmark.

- Before using any generated track commercially, check the model licence and your local rules on voice likeness and lyric rights. Open weights do not automatically mean clear rights.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-tencent-songgeneration-open-source
Free to quote with attribution and a link to the original.
