A new voice conversion system lets you choose how much of someone's accent to keep, not just whether to erase it.
The system, called DLM-AN, works by converting speech into discrete tokens (numeric codes that stand in for small chunks of sound) and then running a "masked diffusion" process that fills in gaps in those tokens step by step, the same general idea behind image-generating diffusion models. A separate module, the Common Token Predictor, flags which of the original tokens already sound close to native pronunciation. Reuse more of those tokens and more of the original accent survives; reuse fewer and the output sounds more neutral. A second component, a flow-matching Duration Ratio Predictor, stretches or compresses timing so the result matches native speech rhythm instead of just native sounds. On multi-accent English test data, the researchers report the lowest word error rate of any system they compared against, with accent reduction and dial-in strength control that stayed smooth and predictable rather than jumping between extremes.
Most accent tools are blunt instruments: convert fully or don't bother. That's a poor fit for dubbing, where a character's regional voice might be part of the performance, or language learning, where a student may want to hear their own accent gradually fade rather than vanish. A tunable knob, if it holds up outside the lab, is the more commercially useful idea here than the diffusion architecture itself.
It's one arXiv paper with benchmark numbers and a GitHub repo, not a dubbing studio's new pipeline, so the real test is whether it holds up on messier, non-English accents and noisy audio.