Back to Feed
Multimodal / Efficiency & Inference

WanSong v1.0 High Fidelity Audio Generation

Original: WanSong v1.0 Technical Report

Listen to the summary

Uses a voice available on your device

Audio options
On this page 4 sections
Related concepts 7 concepts

Key Takeaways

  • Uses a continuous 1-D Variational Autoencoder to compress stereo audio with a downsampling factor of 1024.
  • Implements a dual-stem output strategy that models vocals and background music as separate, interdependent channels.
  • Outperforms Stable Audio 2 on music benchmarks with a 19 percent reduction in STFT distance and a 0.28 dB improvement in SI-SDR.
  • Optimized using three training stages, culminating in DPO and ReFL for alignment with human preferences.

Summary & Methodology Analysis

WanSong v1.0 addresses the difficulty of generating commercial-grade, long-form music by utilizing a continuous 1-D Variational Autoencoder, a tool that compresses raw audio signals into efficient latent representations. By using a downsampling factor of 1024, the system transforms stereo audio into a compact latent stream. This stream is then concatenated with text tokens provided by a decoder-only LLM, forming a unified sequence that is processed by a hybrid-transformer backbone known as an MMDit. This architecture enables the model to effectively manage the complex dependencies between lyrics and musical components.

The model improves upon traditional multi-stage pipelines by implementing a dual-stem output strategy. Instead of mixing all audio components, it models vocal and background music tokens independently while learning their cross-channel dependencies within each transformer block. The training pipeline progresses through three distinct stages involving 90s, 300s, and supervised fine-tuning sequences. These stages employ a flow-matching framework, which is a technique for modeling the transformation of data distributions via a denoising diffusion process, to ensure high-fidelity output. Final model alignment is achieved through DPO (Direct Preference Optimization), a method for tuning models based on binary choices, followed by ReFL.

Performance metrics indicate strong results, as the model achieved a 15.0 percent phoneme error rate and a quality score of 3.2 on the 90s-bench at a compression ratio of 1024. When tested at a compression factor of 2048, the system outperformed Stable Audio 2 on the music benchmark. However, the architecture faces specific limitations. Reducing the compression ratio to improve quality significantly increases token counts, which leads to higher resource consumption and slower generation speeds. Furthermore, the ReFL alignment method is restricted, as it only operates effectively during the low-noise stage of the diffusion process.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the primary contribution of WanSong v1.0?

It provides a system for generating high-fidelity, long-form, commercial-grade music with improved controllability and inference efficiency.

Q2. How does this model handle vocals and background music?

It uses a dual-stem output strategy that models vocals and background music as independent channels while simultaneously learning their dependencies.

Q3. Does this model improve upon existing baselines?

Yes, at a compression factor of 2048, it outperformed Stable Audio 2 by achieving a 19 percent reduction in STFT distance and a 0.28 dB improvement in SI-SDR.

Q4. What is the role of the Variational Autoencoder in this architecture?

The Variational Autoencoder compresses stereo audio into a latent stream to facilitate efficient processing by the transformer backbone.

Q5. What training frameworks and techniques were utilized?

The model used a flow-matching framework for the diffusion process, followed by DPO and ReFL for human preference alignment.

Q6. What happens if the compression ratio is decreased to improve audio quality?

Lowering the compression ratio increases token counts, which results in higher resource consumption and slower generation speeds.

Q7. Are there specific constraints on the ReFL alignment method?

Yes, the ReFL method only operates effectively during the low-noise stage of the training process.

Q8. How does the performance compare at different compression settings?

At a compression factor of 2048, it beats Stable Audio 2, while at a compression ratio of 1024, it achieves a 15.0 percent phoneme error rate and a 3.2 quality score on 90s-bench.

Q9. What specific datasets were used for training the model?

The provided paper does not specify the names of the datasets used for training.