AuK

AuK

A 1.5B speech model from Tencent that generates speech from text, edits recordings, and also cleans up audio and separates sources, all through plain-language instructions.

Made by tencent on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: AuK is a 1.5B-parameter speech model from Tencent, trained on millions of hours of varied audio. One model covers speech generation, speech editing, speech enhancement and source separation, and you steer it with instructions written in plain language.

This repository holds the official weights of the base model. A second variant, AuK-Flash, is a distilled version built for fast inference in 4 steps.

Who it suits: Developers, audio researchers and content makers who want one open model for several speech jobs instead of a separate tool for each. Voice cloning is among its tasks, so use only voices you have the right to use.

Industry | Who uses it | How they use it | Example
Media | Video and podcast producers | Fix or reword spoken lines without re-recording | Replace one word in a recorded sentence
Software | App developers | Add generated voices to products | Create a voice from a written description
Music | Audio editors | Rewrite lyrics in a singing take | Change the words and keep the melody and voice
Research | Speech researchers | Study one model across many audio tasks | Test generation and separation with the same weights

People who only need a basic text reader can skip it, since AuK is built for the wider range of editing tasks.

Where to use it:

• Python and the command line, using the examples in the project's Cookbook
• Windows, macOS or Linux with Python, for downloading the files
• The Hugging Face and ModelScope demo spaces, to try it before you download

Use case | Who | What they get back
Zero-shot voice generation | Developers | Your text spoken in the voice of a reference recording
Voice from a description | Producers | Speech made from a written voice description, with no reference audio
Recording edits | Audio editors | A changed recording with new words, pitch, speed, volume or emotion

How to set it up: You need Python and enough disk space for the weights.

• Python and pip, for the Hugging Face download tools
• Disk space for the model files
• The project's Cookbook, which the card says holds instruction templates plus command line and Python examples

1. Install the huggingface_hub package.
2. Download the model files with the hf command.
3. Open the Cookbook and pick the task you want.
4. Run its example with your own text or audio.

pip install -U huggingface_hub
hf download tencent/AuK

How to use it: Every task goes through the same interface. You write an instruction in plain language and add a reference or source recording when the task needs one. The card's own run commands are cut off here, so use the Cookbook templates for the exact format.

Tool | What it does
Zero-shot TTS | Speaks your text in the voice of a reference recording
Instruct TTS | Makes speech from a voice description, with no reference audio
Speech content editing | Replaces, inserts or removes words in a recording
Lyric editing | Rewrites lyrics in a singing recording and keeps melody and voice
Pitch editing | Shifts pitch up or down in semitones
Speed editing | Changes the speaking rate, and the output length follows
Volume editing | Makes the recording louder or quieter in decibels
Emotion editing | Shifts the emotion and keeps the words and voice
Timbre editing | Changes the timbre to match a description and keeps the content

Example: You give it a recording and ask, in plain words, to replace one word in what the speaker says. It returns a new recording with that word changed.

For newcomers and experts: Think of it as an audio editor you talk to. You say what you want done to a voice recording, and it does it. A good first try is the zero-shot speech task from the Cookbook, using a short recording you own.

Underneath, one 1.5B model handles generation, editing, enhancement and separation through a single instruction interface, which saves you from stitching separate models together. The trade-off between the two variants is quality against speed, since AuK-Flash is distilled for 4-step inference.

Write an instruction in plain words
Add reference or source audio if needed
AuK generates or edits the audio
You get the new recording back

Good to know: This page covers the base model. The fast variant is a separate repository.

• Size: 1.5B parameters, according to the card
• Variants: AuK-Flash for fast 4-step inference, and Hugging Face shows 1 quantized and 3 fine-tuned versions by other people
• License: MIT
• Cost to run: no API key or paid service is named
• Data it reaches: the text and audio you give it, plus the model files you download from Hugging Face
• Not for: the card names no restrictions beyond its license, but cloning a voice needs the speaker's permission

On Hugging Face, 13 November 2026

• Task: text to speech
• Downloads in the last 30 days: 1,455
• Likes: 396
• License: mit
• Quantized versions: 1
• Fine-tunes: 3
• Tags: audio, speech, text-to-speech, zero-shot-tts, voice-cloning, speech-generation, speech-editing, speech-enhancement, source-separation, speech-separation, instruction-guided, diffusion

Links

• Model: https://huggingface.co/tencent/AuK
• Files: https://huggingface.co/tencent/AuK/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tencent/AuK

Then follow the usage in its model card. Nothing in this kit runs by itself.