Tencent Hunyuan open-sources universal speech model AuK with 1.5B...

Tencent Hunyuan open-sources universal speech model AuK with 1.5B parameters

On September 9, 2026, Tencent (Hunyuan) released the source code, weights, and technical report (arXiv:2609.08936) for AuK under the MIT license. AuK is a universal speech foundation model with 1.5B parameters that unifies speech generation, editing, cleaning, and separation through a single natural language interface.

Compiled by Sergey KostenchukPublished 2026-09-10Updated 2026-09-18

2026-09-10 Research HuggingFace

🤖 Tencent Hunyuan open-sources 1.5B parameter speech model AuK

On September 9, 2026, the company released the source code, MIT-licensed weights, and technical report (arXiv:2609.08936). The model understands natural language commands: TTS and voice cloning from a reference, TTS from a description, word replacement and deletion in recordings, re-singing text in a song while preserving the melody, emotion and timbre changes, denoising, speaker separation, and vocal extraction.

🌍 Open weights, code, ComfyUI nodes, Python API, and a fine-tuning pipeline with day-0 support in SGLang-Omni provide an open-source alternative to closed TTS tools.

👤 Demos are already available on HuggingFace and ModelScope: the tencent/AuK and tencent/AuK-Flash weights can be downloaded to clone voices locally, rewrite words, and clean up noise. Russian is not officially supported, but the multilingual Qwen2.5-Omni-3B encoder leaves room for hope.

Source 1: https://github.com/Tencent-Hunyuan/AuK Source 2: https://huggingface.co/tencent/AuK

Sources