MMAE benchmark tests whether AI can edit audio without collateral damage

Why it matters

AI audio is moving from generation toward controllable editing. A credible benchmark would give labs, developers and buyers a clearer way to compare whether models can make precise changes without damaging the rest of a clip.

Tencent Hy (@TencentHunyuan) introduced MMAE, a benchmark for evaluating AI audio editing, in a post on X on Monday.

https://www.youtube.com/watch?v=6At5nTWhlXI

MMAE stands for Massive Multitask Audio Editing Benchmark. Tencent Hy says it developed the benchmark with collaborators including SJTU, SII, NTU, TJU, ZODA, PKU and FDU, and frames the test around a blunt question: whether AI can "truly edit audio, not just generate it."

That distinction matters because much of the recent audio-model race has centered on generation: turning text prompts into speech, music or sound effects. Editing asks a narrower and often harder question. A useful model needs to understand an existing clip, change only what a user asks it to change, and preserve the rest, such as speaker identity, timing, background sound or linguistic content.

Tencent Hy says current models reach an Exact Match Rate below 5% on MMAE, a sign that reliable instruction-following audio editing remains far from solved. The release says the benchmark includes 2,000 high-fidelity samples from real-world scenarios, 17,741 fine-grained rubric evaluation items, seven modality settings across sound, music, speech and mixtures, six levels of task complexity, and eight operation types spanning local and global edits.

The project materials are available through the MMAE arXiv paper, GitHub repository, Hugging Face dataset, and a demo video.

Tencent Hy claims MMAE is the first comprehensive benchmark of its kind. That priority claim still depends on how the accompanying paper defines its task scope and comparisons. But the more immediate point is the gap the benchmark is trying to expose: audio models can increasingly synthesize convincing clips, yet precise, instruction-based edits that leave unrelated content untouched remain difficult to measure and harder to execute.

For Tencent Hy, MMAE is also a positioning move. Benchmarks help define which capabilities count, and in AI audio, the lab that sets the test can shape how competitors describe progress.