News · 2026-09-09

Tencent open-sources AuK, a speech model you direct with written instructions

Tencent open-sourced AuK on 9 September 2026, a 1.5-billion-parameter speech model that performs voice cloning, content and acoustic editing, speech enhancement and source separation — all driven by written instructions rather than separate tools. The weights are published under the MIT licence, and the repository is a 6.8 GB download. A distilled variant, AuK-Flash, runs in four steps for faster generation.

Key facts

• The release: a 1.5B-parameter unified speech model under the permissive MIT licence.
• When: weights and code published 9 September 2026.
• Who: Tencent Hunyuan, with academic collaborators including Ziyang Ma and Xie Chen.
• Primary source: the model card on Hugging Face; paper at arXiv 2609.08936.

The interesting design decision is the interface, not the architecture. Open speech tooling is normally a drawer of separate utilities: one model clones a voice, another strips background noise, another separates overlapping speakers, another edits a word in the middle of a recording. Each has its own invocation, its own quirks and its own failure modes. AuK collapses that drawer into one model you talk to. As the authors put it, AuK "exposes every task through the same natural-language instruction interface."

Underneath, the paper describes three components working together: a multimodal language model providing semantic conditioning — the part that understands what you asked for — a variational autoencoder trained jointly on speech, general audio and music to handle the acoustic side, and a hybrid rectified-flow transformer that does the generation. The flow-matching approach in that last stage is the same family of technique now common in image and audio generation, and it is what makes the four-step distilled variant possible.

The analogy is the shift from a rack of single-purpose kitchen gadgets to one good chef. The gadgets are each fine at their one job, but you must know which to reach for, and none of them can do something you did not buy a gadget for. Asking in words — make this sound like it was recorded in a smaller room, take the second speaker out, say that line again but wearier — moves the burden of knowing the tool's name off the user.

On the numbers readers actually need: the Hugging Face repository totals about 6.8 GB, comprising a 6.1 GB main checkpoint and a 637 MB audio decoder. The GPU memory needed to run it is not stated anywhere — neither the model card nor the GitHub repository publishes a minimum or recommended VRAM figure, and the documentation only shows how to select a device. Computed from the shipped files, the weights alone occupy roughly 6.8 GB when loaded at the precision they are stored in, which is a floor rather than a requirement — activations, caches and audio buffers come on top of that. Anyone planning a deployment should benchmark rather than trust an estimate.

The MIT licence is the part with the most practical consequence. Open speech models have tended to ship under bespoke community licences carrying use restrictions, non-compete clauses or commercial gates, which makes them awkward to build a product on. MIT has none of that. For a category where the obvious applications — dubbing, accessibility, audio production, podcast tooling — are overwhelmingly commercial, the licence may matter more than the quality.

That same permissiveness is the honest caveat, and it is not a small one. A capable, freely licensed voice-cloning model that anyone can download is also a capable, freely licensed voice-cloning model that anyone can download. The text-to-speech field has been steadily removing the friction that used to make voice impersonation hard, and this release removes more of it. Neither the model card nor the paper describes watermarking or provenance signals in the generated audio, which is the mitigation the field has largely converged on — see content provenance and watermarking for why that omission matters.

A second caveat is that the performance claims come from the authors' own evaluation, presented as a chart on the model card rather than reproducible numbers, and no independent comparison against established open baselines has been published yet. There is a hosted demo and a project site for anyone who wants to judge the output directly rather than take the chart on faith — which, for a generative audio model, is the only assessment that really counts.

Primary source, verified: read the paper → (arXiv 2609.08936)

Ground Truth. (2026, September 9). Tencent open-sources AuK, a speech model you direct with written instructions. Ground Truth. https://groundtruth.day/news/tencent-open-sources-a-speech-model-you-steer-with-plain-instructions.html

@misc{groundtruth:tencent-open-sources-a-speech-model-you-steer-with-plain-instructions,
  title  = {Tencent open-sources AuK, a speech model you direct with written instructions},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/tencent-open-sources-a-speech-model-you-steer-with-plain-instructions.html}
}

Topics: speech · open-weights · model-release · tencent · text-to-speech