Zero-shot TTS can already turn a few seconds of reference audio into "your voice," but one knot has long stayed untied: the reference carries not just timbre but accent. Keeping your voice while switching to another accent usually means chaining a voice-conversion model after the TTS — two pipelines, two rounds of degradation. DEFINE, recently submitted to Hugging Face Daily Papers by AMAAI Lab (arXiv:2609.32777), offers a single-model solution in a tight 5-page paper with a sharp problem statement.
Pull Apart "Who Is Speaking" from "How They Speak"
The paper's core observation: speaker identity and accent are entangled in the same reference audio in conventional zero-shot TTS. DEFINE conditions them on two separate branches — identity from one speaker exemplar, accent from another, and the two clips can come from entirely different people. Accent exemplars are mapped into a conditioning space by an exemplar encoder, supervised through learned accent prototypes during training. At inference time the system needs no accent labels, and there is no post-synthesis waveform conversion: the decoupling happens inside generation, in one pass.
One Guidance Weight, One Accent Knob
Implementation-wise, DEFINE is built on F5-TTS with parameter-efficient LoRA adaptation rather than retraining the backbone. The most practical detail: accent strength is continuously controlled by a single inference-time guidance weight, with no retraining to turn it up or down. One model covers a continuous spectrum from a faint trace to a heavy accent — for dubbing, localization, and multilingual content production, this is a must-have interaction pattern, not a paper-only demo.
The Numbers, and the Paper's Own Honest Boundaries
On accents seen in training, increasing accent guidance lifts accent-probe accuracy from 6.5% to 19.6%; a single DEFINE model matches the accent-transfer performance of a two-model TTS-plus-voice-conversion cascade on seen and out-of-domain accents, while achieving higher speaker similarity and comparable predicted speech quality. Notably, the abstract states the limit plainly: the generalization does not hold on fully held-out accents. And a 19.6% probe accuracy in absolute terms is a reminder that "controllable accent" is still some distance from "reliable accent." Reporting failure boundaries in the abstract is a restraint the field should copy more often.
So What
For speech product teams, the point is compressing "accent switching" from a two-model pipeline back inside a single model — saving latency, compute, and cascade degradation. For the broader generative-model community, this is a template of fine-grained attribute disentanglement plus inference-time continuous control, and the same recipe should transfer to emotion intensity, speaking style, or any other attribute silently photocopied into the reference audio. The next thing to watch is whether accent prototypes can be opened for custom registration — letting users add a new accent by feeding one exemplar would be the finished form of this knob. Full paper: arXiv:2609.32777 (https://arxiv.org/abs/2609.32777).