Tencent's Hunyuan lab open-sources AuK, a 1.5B speech model that generates, edits and separates audio from one instruction interface

Tencent's Hunyuan team open-sourced AuK on 9 September, a 1.5B-parameter foundation model for speech generation and editing that treats every task, from zero-shot text-to-speech to source separation, as the same message-shaped call with a natural-language instruction attached. The GitHub repository sits at 485 stars two days after the release. The README says the model was trained on "millions of hours of diverse audio data".

What the same call can do

The task list in the README is unusually wide for one checkpoint. Generation covers zero-shot TTS, where the model is asked to speak a target text in the voice of a reference clip, and instruction TTS, where it generates speech from a written description of the voice alone. Content editing rewrites what a recording says by describing the replacement in the prompt, and a lyric-editing mode rewrites the words of a singing recording while preserving the melody and the singer. Acoustic editing covers pitch, speed and volume; paralinguistic editing covers emotion, timbre, de-accenting, whisper conversion and adding or removing nonverbal sounds like breaths or laughs. Enhancement and separation cover denoising, dereverberation, speech separation by talking order, target-speaker extraction, and pulling the singing voice out of a music mix.

Two variants, and how they run

AuK ships as two checkpoints on Hugging Face and ModelScope. AuK Base is the higher-quality model with configurable NFE and CFG. AuK-Flash is a distilled version that runs in 4 fixed time steps with CFG set to 0, and it is meant for cases where response time is the constraint. Both share a Qwen2.5-Omni-3B multimodal LLM encoder that is loaded separately at runtime, which is why the download instructions pull a third weight directory. The Gradio interface can expose both models at once through a model selector, or a single one behind an OpenAI-compatible API.

Every task uses the same CLI call, and the only field that changes between them is a natural-language instruction. To rewrite the words of a recording, the instruction spells out the replacement; to clean up a noisy clip, it says which speakers to keep and what to remove; to change the emotion or timbre of a voice, it describes the target in prose. The uniform interface is the point of the release, more than any single quality number.

What ships around it

SGLang-Omni added day-zero serving support for both AuK and AuK-Flash, per the release notes, and Hunyuan published cookbook entries for each task with instruction templates plus Python and CLI examples. The repository lists a ComfyUI node package and a fine-tuning install target alongside the plain inference and Gradio ones. Live demos are on Hugging Face Spaces and ModelScope Spaces. The licence file in the repository is labelled Other, and the repository description lists topics including text-to-speech, zero-shot TTS, speech separation, speaker extraction and speech enhancement.

For a team that has been stitching together a separate TTS provider, a voice-cloning model, a denoiser and a source separator, the practical question is whether one open-weight checkpoint at 1.5B parameters can hold all of those jobs to the quality bar the pieced-together stack was reaching. That is a listening test on a team's own audio: any single benchmark number would settle only its own slice. What the release does settle is that the instruction-interface pattern from LLMs is now available for audio, from a lab willing to open the weights.

Source

Tencent Hunyuan's AuK repository, with the AuK Base and AuK-Flash checkpoints on Hugging Face.