Tencent Hunyuan AuK: One 1.5B Model for Speech Generation and Editing

Speech models have a habit of multiplying. One handles text to speech, another cleans up noise, a third clones a voice, a fourth separates a singer from a backing track. Tencent's Hunyuan team decided that was too many models, and on September 9, 2026, it published AuK, a 1.5B parameter foundation model that puts all of those jobs behind a single instruction interface. The code and weights are MIT licensed, which for a speech model of this range is unusually permissive.

The pitch is simple to state. You give AUk an audio clip and a sentence describing what you want changed. If you want the line "we have a meeting tomorrow morning" to become "meeting is Friday afternoon," you say so, and the model rewrites the speech without re-recording it. If you want the same voice to sound calmer, you ask for a calmer delivery. If a cough is sitting in the middle of a take, you tell it to remove the cough.

One instruction box, many jobs

The task list is long and it is deliberately flat. AUk covers zero-shot text to speech, where it reads new text in the voice of a reference clip. It covers instruction based TTS, where a written description of a voice is enough and no reference audio is needed. For editing, it can replace, insert, or delete words in an existing recording, and it can rewrite lyrics in a sung take while trying to preserve the melody and the original voice.

Then there is the acoustic layer. Pitch can be raised or lowered by semitones. Speaking rate can be adjusted, with output length scaling to match. Emotion and timbre can be changed while the words stay put, and a regional accent can be stripped out without losing the speaker's identity. The model also handles nonverbal sounds, so breaths, laughs, and coughs can be added or removed, and it can convert between normal speech and a whisper.

The cleanup work is the quiet part

Enhancement and separation round out the set. AUk can denoise and dereverb a recording, pull one speaker out of a crowded conversation by talking order, or extract a vocal from a mix. That last capability is the kind of thing that used to require a separate tool and a patient engineer. Folding it into the same interface as voice generation is the actual claim being made here.

The training numbers back the ambition. The team says the model was built on roughly 3.03 billion instruction and audio instances, drawing on about 1.95 million hours of effective supervision across generation, editing, enhancement, and separation. Architecturally, a multimodal language model supplies semantic conditioning, a jointly trained VAE supplies acoustic conditioning, and both feed a hybrid rectified-flow transformer. Pretraining, joint generation and editing, preference optimization for open-ended edits, and reward based reinforcement learning all happen before release.

Why the distilled sibling matters

There is a faster variant called AUk-Flash, distilled to run inference in four steps without classifier-free guidance. The technical report puts it at a 4.5x wall-clock speedup over the full model under matched conditions. That detail matters because speech editing is one of those tasks where the first useful result is never the final one. You change a word, listen, change the emotion, listen again. If each pass costs a minute, the workflow is unusable no matter how good the output is.

The ecosystem support arrived quickly, which is often a better signal of real interest than a benchmark table. vLLM-Omni and SGLang-Omni added day zero support, MLX inference on Apple Silicon is available, an audio.cpp integration allows GGUF offline inference, and there are official ComfyUI nodes. A 1.5B model is small enough to self-host, and the community tooling around it makes that concrete rather than theoretical.

What the licence changes

The MIT licence deserves its own paragraph. Much of the strongest speech work has shipped under terms that restrict commercial use, or that carve out whole regions. A permissive licence on both code and weights lowers the barrier for small studios, podcast editors, and localization teams that could never justify a per-seat contract for a dubbing tool. It also means the model can end up inside other products without a legal review that takes longer than the integration.

The caveat is scope. AUk's coverage is centred on Chinese and English, so teams working in other languages should test before they plan around it. Voice cloning and recording edits also carry consent questions that no licence settles. The sensible habit is the same one the team asks for: use material you have the right to use.

The dubbing workbench

Think about who this is aimed at. A localization team has a finished video in one language and needs it in five more. The old process is expensive and slow: book voice actors, record each line, match timing, pay for another round when a translation changes a single word. If a model can take the original performance and swap the words while keeping the voice and pacing, most of that pipeline collapses into an edit.

Podcasters have a lighter version of the same problem. A host misreads a sentence, fluffs a number, or wants the same episode delivered with a calmer tone for a different audience. Re-recording is the obvious fix and the wrong one. Editing the take in place, keeping the room tone and the microphone character, is what the tool promises. Whether it delivers on a full episode rather than a demo clip is the open question, but the need is not in doubt.

There is a second audience that is easy to miss: developers building voice features. Because AUk runs at 1.5B parameters and has open weights, it can be embedded in an app rather than rented per call. A small team can ship a voice feature, keep the costs fixed, and avoid sending user audio to a third party. For privacy-sensitive products, that last point may be the deciding one.

Where it can go wrong

Editing audio is a liar's errand if the model makes small mistakes, because audio leaves nowhere to hide. A generated word that fits perfectly still stands out if the breath before it is missing, or if the room tone shifts for half a second. Speech editing lives or dies on the seams, and the seams are exactly where models tend to slip. The team's own preference optimization and reinforcement learning stages were aimed at this, which is a good sign, but only sustained use will show whether the seams disappear.

There is also the consent problem that arrives with every cloning-capable model. The same tool that fixes a host's fluffed line can put words in someone's mouth, in their own voice, with no trace. AUk's licence does not address that, and no licence could. The responsibility lands on whoever operates the tool, and the teams that handle this well will be the ones that build consent into their process before they build the feature.

What to actually take from this

The interesting move here is not that another speech model exists. It is that generation and editing were treated as the same problem. For most real work, the recording already exists and only needs a handful of changes. A tool that treats "say this" and "fix that word" as two points on one curve matches how editors actually think. The breadth of tasks is what makes the unified interface more than a slogan: the same model that can speak a line can also clean it, separate it, and change how it sounds.

The practical advice is to test the editing side first, on your own messy audio, before you plan around it. Generation is the part everyone demos, and editing is the part that will decide whether the model earns a place in a real workflow. If it holds up there, the combination of small size, permissive licence, and fast distilled inference makes AUk one of the more usable open speech releases of the year.