Tencent Hunyuan open-sourced its voice generation and editing model AuK on September 9. The code sits under the Tencent-Hunyuan organization on GitHub, the weights are up on both Hugging Face and ModelScope, and the technical report is listed as arXiv 2609.08936.
The model isn't positioned as just another TTS system. It folds work that used to be spread across a dozen-plus specialized models into a single entry point: give it a natural-language instruction plus a reference clip describing what you want done, and it outputs the edited audio.
A Dozen Jobs Through One Door
Tencent's documentation splits the tasks into six categories. On the generation side there's zero-shot text-to-speech and instruction-driven TTS. Editing is split into three layers — the content layer changes the words or lyrics being spoken, the acoustic layer adjusts pitch, speed and volume, and the paralinguistic layer changes emotion, timbre, accent and non-verbal sounds, including turning normal speech into a breathy voice. The processing side covers denoising, dereverberation, music separation and target-speaker extraction.
Each of those used to require its own dedicated model. A dubbing team would run one TTS stack, swap in a separate voice-conversion tool to change emotion, and bring in yet another separation model to strip background noise — with each tool using its own audio format, sample rate and dependency versions, so most of the engineering effort went into stitching them together. AuK's approach is to treat the instruction itself as the unified interface, letting the model decide which path to take.
The encoder uses Qwen2.5-Omni-3B as its multimodal backbone. Alongside the Python API, the repository ships a Gradio interface and ComfyUI nodes, both common entry points in today's open-source audio-video community. The demo clips use Chinese and English examples.
Where AuK-Flash Gets Its Speed
AuK-Flash is a fast version distilled from the base model, with NFE fixed at 4 and CFG set to 0. Tencent describes the distillation as a two-stage process: first, trajectory-level consistency initialization gets the few-step student model up and running, then training switches to a task-routed, decoupled DMD objective, with the separation task supervised separately through clean-prediction regression.
For users, what that means in practice is this: a typical diffusion-based voice model needs dozens of sampling steps per generation, plus an extra unconditional pass to keep quality up. AuK-Flash cuts both of those costs, at the price of landing "close to" rather than identical to the teacher model's quality. Tencent's own figure is a 4.5x wall-clock speedup, with no published number for the corresponding quality loss on standard metrics.
How It Stacks Up Against Other Open Voice Models
Open-source voice models have been coming out fast this year. Mistral released a 4B voice model benchmarked against ElevenLabs; Alibaba's Qwen3.5-Omni folds text, audio and video into one system with voice-cloning support; and Hunyuan itself had already open-sourced the Hy4 base model and an inference-acceleration framework.
AuK's real point of difference isn't the benchmark scores — it's the word "editing." Those other models are mainly built around generation: give them text, get speech back. Half of AuK's functionality instead takes an existing audio clip and modifies it, while keeping the original speaker's timbre and rhythm intact. That's a different engineering problem from generation, with stricter consistency requirements, and it sits closer to how actual post-production work happens.
The MIT license is worth calling out on its own. It permits commercial use, modification and redistribution as long as attribution is kept, which is more permissive than the custom licenses that come with a number of domestic open-source models. For small teams trying to build voice features into a product, the license terms can matter more than parameter count when it comes to picking a model.
Tencent hasn't yet published WER, speaker-similarity or MOS numbers comparing AuK against other models. The repository lists performance comparisons across four task categories — generation, editing, enhancement and separation — but the specific scores still need to be reproduced independently. Claims that it's state-of-the-art at launch are, for now, coming from community reports, with no third-party evaluation to back them up yet.
Sources: Tencent-Hunyuan/AuK repository and technical report, CocoLoop, Hugging Face paper page; parameter count, license and speedup figures checked against the official repository, evaluation scores not independently verified.