SonicMaster: AI system lets anyone fix audio quality with text instructions

Home recordings and amateur productions frequently struggle with sound problems like excessive reverb, distortion, clipping, and tonal imbalances. Fixing these issues normally requires multiple specialized software tools and skilled manual tweaking. Researchers have now developed SonicMaster, a single generative AI model that can address a wide range of audio problems using simple text instructions—or operate automatically for quick restoration.

How it works#

SonicMaster operates by accepting natural language prompts that guide the enhancement process, while also offering a fully automatic mode for general cleanup. The model was trained using a flow-matching generative framework, which learns to transform degraded audio into polished, mastered versions. To train the system, the team built a large dataset by deliberately creating common sound problems and pairing them with high-quality versions, covering five major categories: equalization, dynamics, reverb, amplitude control, and stereo imaging.

What problems it solves#

The system tackles a comprehensive set of audio artifacts through its nineteen degradation functions. These span the most common issues encountered in non-professional recording environments: excessive reverberation, distortion and clipping, tonal imbalances, and narrowed stereo width. Rather than forcing users to open separate tools for each problem, SonicMaster consolidates all these fixes into one unified workflow.

Performance and validation#

Testing using objective audio quality metrics shows measurable improvement across all degradation types. More importantly, subjective listening tests found that human listeners consistently preferred the enhanced outputs produced by SonicMaster compared to other competing approaches, indicating the model's effectiveness at producing naturally pleasing results.

The research was submitted to arXiv and represents a step toward making professional-quality audio restoration accessible without requiring specialized knowledge or multiple software licenses.