Qwen-Music: Melody-CoT for Full-Song Generation with Vocals

Introduction

Qwen has released a technical report for Qwen-Music, a music generation model aimed at producing complete songs rather than short audio snippets. The system is designed to generate high-fidelity music with full vocal singing, while also supporting more controlled tasks such as lyric-conditioned songwriting and style-shifted cover generation.

Key Points

• Two main tasks: Qwen-Music handles Text to Music Generation, where new songs are created from text descriptions, lyrics, and musical attributes. It also supports Cover Song Generation, where an existing song can be reinterpreted with different styles or vocal characteristics.
• A three-part architecture: Qwen-Music-Tokenizer compresses audio into a 25 Hz single-codebook stream of Music Semantic Tokens. These tokens are intended to preserve semantic and melodic information in a form suitable for prediction by an LLM. Qwen-Music-LLM then performs autoregressive music semantic modeling, while Qwen-Music-Render converts the discrete semantic representation into high-fidelity stereo waveforms.
• Melody-CoT: The report’s central novelty is a melody-token-based chain-of-thought mechanism. Instead of generating an entire song directly, the model first plans melodic structure and then expands it into the full musical output. The authors say this improves creativity, musicality, structural coherence, and reference-audio-based melody cloning.
• Large-scale multilingual training: Qwen-Music-LLM is trained on more than 5 million hours of multilingual music data covering hundreds of languages. The training pipeline includes a quality-aware pre-training curriculum, followed by progressive post-training with supervised initialization, offline DPO, and online GSPO to improve musicality and instruction following.

Why It Matters

Full-song generation is harder than simply producing plausible sound. A useful music model must keep melody, rhythm, vocals, lyrics, timbre, arrangement, and structure aligned over a longer duration. Qwen-Music reflects a broader design pattern in audio AI: separate the symbolic or semantic planning layer from the final acoustic rendering layer. This lets the LLM work on a compressed musical representation while a dedicated renderer restores stereo detail and audio fidelity.

The report states that, across 600 Chinese and English prompts, Qwen-Music achieves state-of-the-art results in 13 of 16 objective metrics for musicality and audio quality. Professional evaluators are also reported to prefer Qwen-Music over leading proprietary systems. For cover song generation, the report says the model preserves reference melodies more accurately than leading proprietary alternatives.

One important caveat is availability. The provided material does not state that model weights are open, and community comments are already asking about open sourcing. For now, Qwen-Music is best read as a technical blueprint and performance claim. If Qwen later releases weights, demos, or reproducible evaluation details, it could become a significant reference point for controllable AI music generation.

Source: Hugging Face Daily Papers