Tencent's Hunyuan team open-sourced AuK on Sept. 9, 2026: a single AI model that can create speech and edit existing recordings, following instructions written in everyday language. Tencent released it under the MIT license, a license that lets anyone use, change and share it free of charge. The model files are on Hugging Face and ModelScope, and the code to train and run it is on GitHub. Most strong speech-editing tools still sit behind closed services, or split creating, editing and cleaning up audio across separate programs.
AuK does more than read text aloud. Tencent describes one way of working for every job: the user writes an instruction and, if needed, adds an audio clip. The same model can then make new speech, change the words in a recording, clean up audio, separate voices or sounds that are mixed together, or change how words are delivered and how a recording sounds, without a different tool for each task.
Tencent's technical report, submitted Sept. 8, says the core of AuK has about 1.5 billion parameters, the numbers a model tunes as it learns. It was trained on the equivalent of about 1.95 million hours of instruction-and-audio examples. Later rounds of training used human feedback to improve open-ended editing, and trial and error with rewards to improve speech generation. A faster version, AuK-Flash, was trained to imitate the full model; Tencent says it runs about 4.5 times faster on the same hardware.
Tencent reports that AuK came out ahead of the rival systems it listed, including Qwen3-TTS and Seed-TTS, on a standard test of reading text aloud accurately in a given voice. The paper calls its results leading for creating speech and following editing instructions, and competitive for restoring audio. All of these results come from Tencent's own tests, and no outside comparison has been published yet.
The paper notes its own limits. It says AuK still cannot reliably handle free-form editing requests on its own, and works better when each request is first sorted by type of task and passed through a helper Tencent calls the Prompt Enhancer. According to a news report, neither the paper nor the model's Hugging Face page describes a watermark or any other signal that marks AuK's audio as AI-made.
AuK is also the reference system for one track of the ICASSP 2027 Audio Editing Challenge, which IEEE lists among its grand challenges. The challenge site says Tencent sponsors the prize pool and that several organizers are also AuK's authors, so AuK's place in the contest is not an outside verdict on its quality.