Skip to content
GenLovers

How to Set Up AuK: A Local, Uncensored Voice Model

Last updated: 11 min readDifficulty: Intermediate

Written by Clement

A portrait of the GenLovers mascot with a cracked black censor bar across her mouth and small redaction blocks escaping past it, beside the words Uncensored TTS, It Actually Speaks, illustrating an AI voice model with no content filter.
Cover art for the companion YouTube video, generated per docs/design/YOUTUBE_THUMBNAIL_PLAYBOOK.md.

AuK is a 1.5B open-source speech model from Tencent Hunyuan, Shanghai Jiao Tong University, and the Shanghai Innovation Institute: zero-shot voice cloning from a short reference clip, text-to-speech, and instruction-driven speech editing (replace a word, change the emotion, remove background noise) all through one model. This walks through running its official Gradio web form on a rented cloud GPU, so you don't need your own GPU to try it.

1 steps

Watch the video walkthrough

  1. 1

    Step 1 of 16

    Prefer video?

    This guide has a companion video covering the same setup end to end, from picking a GPU through the Gradio demos.

    Watch on YouTube

What AuK does

AuK exposes every one of its 14 tasks through the same plain-language Instruction box in the Gradio form: there's no per-feature menu, so what the model does depends on what you type, not which tab you're on. AuK-Flash, the distilled variant the setup below also installs, trades a small amount of quality for roughly 4.5x faster generation, useful once you're iterating on a line rather than generating it once.

Nothing about the model or this setup applies a content filter to what you type as the instruction. That's the same trade-off as the local image and companion-chat guides on this site: full control over output, and full responsibility for what you generate.

Everything you can do from the same Instruction box

All 14 come from Tencent's own task table (AuK's README), reachable from the one Gradio form: pick Zero-shot TTS when you upload a reference clip, or write the instruction for any other task instead.

Zero-shot TTS
Upload a short reference clip; speak your target text in that voice. E.g. "Say the following with the same voice: '...'"
Instruct TTS
No reference clip needed; describe the voice in words instead. E.g. "A calm, low-pitched male voice, mid-30s, saying: '...'"
Speech Content Editing
Rewrite what's said in an uploaded clip. E.g. "Replace 'yes' with 'no'"
Lyric Editing
Rewrite lyrics in an uploaded singing recording while the melody and voice stay the same
Pitch Editing
Raise or lower the pitch by semitones. E.g. "Raise the pitch by 3 semitones"
Speed Editing
Adjust the speaking rate; output length scales with the speed factor
Volume Editing
Raise or lower the volume by decibels
Emotion
Change the emotion while keeping content and voice the same. E.g. "Make this sound happy"
Timbre
Change the timbre to a description while keeping the words the same
De-accent
Remove a regional accent while preserving the speaker's voice and content
Nonverbal Editing
Add or remove breaths, laughs, or coughs at a point in the clip
Whisper Conversion
Convert between normal speech and a whisper, same speaker and content
Speech Enhancement
Denoise, dereverberate, or otherwise clean up a recording
Speech / Music / Target-Speaker Separation
Keep one speaker or the singing voice from a mix, or isolate whoever is named by what they say

2 steps

Create a RunPod account

  1. 2

    Step 2 of 16

    Sign up at RunPod

    Go to runpod.io and create an account (this is a referral link).

    Open runpod.io
    RunPod homepage with the Sign Up button in the top navigation circled.
    Click on the signup button on the top right to start Open runpod.io
  2. 3

    Step 3 of 16

    Add funds to your account

    $10 covers a solid testing session.

    RunPod dashboard top-right corner with the add-funds plus button circled.
    The + button next to your balance, top right of the dashboard.

5 steps

Configure and deploy the pod

AuK's checkpoints (AuK Base, AuK-Flash, and the Qwen2.5-Omni-3B text encoder it depends on) total about 25.5GB on HuggingFace. Tencent ships one set of full bf16 weights.

  1. 4

    Step 4 of 16

    Pick a GPU with 20GB VRAM or more, with BF16 support

    From the dashboard, click Pods in the left nav, then Deploy Pod.

    Sort by VRAM and price, and pick a current-generation card at or above 20GB.

    RunPod GPU instance list with a 24GB-class card's price and VRAM circled.
    The GPU list, sorted by VRAM, RAM, vCPU, and hourly price.
  2. 5

    Step 5 of 16

    Set the pod template to a PyTorch build

    Runpod Pytorch 2.8.0 is the current default template; it ships CUDA and PyTorch pre-installed.

    If a different template is selected, hit Change template and pick it from the list.

  3. 6

    Step 6 of 16

    Expose port 7860

    Under Expose HTTP ports, set 8888,4000,7860.

    7860 is auk-gradio's default port; 8888 covers Jupyter.

  4. 7

    Step 7 of 16

    Set container disk to at least 60GB

    The checkpoints total about 25.5GB; 60GB (the setup script's own disk manifest) leaves headroom for the cloned repo, pip cache, and PyTorch/CUDA environment on top of that.

  5. 8

    Step 8 of 16

    Deploy On-Demand

    On-Demand bills per second of actual usage, no upfront commitment, right for testing before you know how much you'll use this.

Get new guides by email

One email when we publish new guides and model breakdowns. No spam, unsubscribe anytime.

4 steps

Run the setup script in Jupyter

  1. 9

    Step 9 of 16

    Open Jupyter Lab

    On the pod's Connect tab, port 8888 shows "Jupyter Lab, Ready." Click it.

  2. 10

    Step 10 of 16

    Download the setup script

    Grab setup_auk_gradio.py below and save it locally.

    Download the setup script (.py)
  3. 11

    Step 11 of 16

    Upload the setup script

    In Jupyter, upload the file you just downloaded into /workspace.

  4. 12

    Step 12 of 16

    Open a terminal and run the script

    python setup_auk_gradio.py.

    It clones the official Tencent-Hunyuan/AuK repo, patches the Gradio UI's Prompt Enhancer checkbox to default off, installs AuK with the [gradio] extra plus a runtime dependency (audioread) AuK's own install misses, and downloads AuK Base, AuK-Flash, and the Qwen2.5-Omni-3B encoder from HuggingFace. On a Blackwell-generation card (RTX 50-series, RTX PRO Blackwell), it first preinstalls a CUDA 12.8 build of PyTorch so AuK's own install doesn't downgrade to a build with no Blackwell kernels.

4 steps

Open the Gradio UI and generate

  1. 13

    Step 13 of 16

    Go back to the pod's Connect tab in RunPod

    Same screen you used to open Jupyter.

  2. 14

    Step 14 of 16

    Open port 7860

    That's auk-gradio's default port, already exposed from the pod-template step above.

    Click it once the setup script finishes.

  3. 15

    Step 15 of 16

    Fill in the form

    "Use Prompt Enhancer" starts unchecked (the setup script patches its default; see the callout below for why), so pick AuK Base or AuK-Flash from the model selector, type the text to speak, optionally upload a reference clip for voice cloning, write your instruction (e.g. "Say: '...'" or "Replace 'X' with 'Y'"), and set Duration to the number of seconds you want.

  4. 16

    Step 16 of 16

    Click Generate

    The first click loads the Qwen text encoder, the VAE, and the DiT model into VRAM before generating anything, which took about 30-45 seconds in testing; every generation after that is faster since the models stay loaded. The generated audio plays back directly in the browser; download it from the same panel.

Hardware and software checklist

Tencent's own README reports peak VRAM measured on an NVIDIA A800-SXM4-80GB with bf16 inference, after the 2026-09-16 encoder-memory fix (PR #19): about 18GB without cpu_offload, about 10.5GB with it enabled via auk-gradio's --cpu_offload flag. Actual usage still depends on input length, dtype, and hardware.

GPU
20GB VRAM or more, with BF16 support; 12GB or more if running with --cpu_offload
Container disk
60GB or more (checkpoints total about 25.5GB; the rest covers the cloned repo, pip cache, and Python environment)
License
AuK's own weights are MIT; the bundled Qwen2.5-Omni-3B text encoder ships under Alibaba's Qwen Research license, not MIT. Check that license before commercial use of the encoder

Troubleshooting

Out-of-memory error

Re-run with python setup_auk_gradio.py --serve --cpu_offload to trade inference speed for roughly 7-8GB less resident VRAM, or move to a larger GPU

"Prompt Enhancer failed" / missing LLM_API_KEY error

The checkbox got checked back on. Uncheck "Use Prompt Enhancer" and set Duration manually, or set up LLM_API_KEY / LLM_BASE_URL / LLM_MODEL_NAME first; this setup script doesn't configure that optional LLM

"CUDA error: no kernel image is available for execution on the device"

Your GPU is a newer Blackwell-generation card (RTX 50-series, RTX PRO Blackwell). The setup script detects this and preinstalls a Blackwell-compatible PyTorch build automatically; if you still see this, you're running an older copy of the script. Re-download setup_auk_gradio.py from this guide and re-run python setup_auk_gradio.py --install

"ModuleNotFoundError: No module named 'audioread'" on Generate

AuK's own install doesn't declare this runtime dependency of qwen-omni-utils. Run pip install audioread in the same terminal, then Generate again; the setup script installs it automatically on a fresh run

Generated speech cuts off mid-word

Increase the duration setting in the form

Output doesn't sound like the reference clip

Use a longer, cleaner reference clip with minimal background noise; a few seconds of clear speech clones more reliably than a noisy one

Frequently asked questions

What is AuK, and who made it?
AuK is a 1.5B open-source speech model from Tencent Hunyuan, Shanghai Jiao Tong University, and the Shanghai Innovation Institute, released under an MIT license, covering zero-shot voice cloning, text-to-speech, speech editing, and audio cleanup in one model.
Do I need a GPU to run AuK locally?
Yes, an NVIDIA GPU with BF16 support. Tencent's own measured peak VRAM is about 18GB without cpu_offload and about 10.5GB with it enabled, so 20GB or more is a safe target without offload, 12GB or more with it.
What's the difference between AuK Base and AuK-Flash?
AuK-Flash is a distilled variant that runs 4-step inference without classifier-free guidance, roughly 4.5x faster wall-clock than Base, at a small quality cost. Both are installed by the setup script; pick per generation depending on if you're iterating or finalizing.
Can AuK clone a voice from a short clip?
Yes, zero-shot: upload a short reference clip through the Gradio form's audio input, and AuK speaks your text in that voice with no fine-tuning or training step required.
Is running AuK on RunPod free?
No, RunPod bills per second for GPU time, and AuK's VRAM footprint (about 18GB without cpu_offload, about 10.5GB with it) puts it above the cheapest cards on offer. It's not free the way running fully on CPU (like this site's local companion-chat guide) is.
How is this different from using ElevenLabs?
ElevenLabs is a hosted service: no setup, but usage-based pricing, an account, and its own content policy. AuK runs entirely on hardware you control, with no per-generation fee once the pod is running and no content filter on the instruction text, at the cost of the RunPod setup in this guide and the GPU rental while it runs.
Does this work on newer Blackwell-generation GPUs (RTX 50-series, RTX PRO Blackwell)?
Yes. AuK's own PyPI install pulls a PyTorch build with no Blackwell kernels, which detects the GPU but crashes on the first real generation with a CUDA kernel-image error. The setup script checks the GPU's compute capability and preinstalls a Blackwell-compatible build first when needed, so this is handled automatically.
Should I turn Prompt Enhancer on?
Not unless you've set up an OpenAI-compatible LLM endpoint (Tencent Cloud TokenHub, DeepSeek, OpenAI, or similar) via the LLM_API_KEY, LLM_BASE_URL, and LLM_MODEL_NAME environment variables. It's a convenience layer that turns a free-form request into a ready-to-run instruction; every core capability (zero-shot TTS, voice cloning, content editing) works with it off and Duration set manually, which is how this setup ships by default.
Does AuK support French, or other languages besides English and Chinese?
No, not reliably. We tested French and the output was worse: a weaker accent and less natural delivery than the same instruction generated in English. Tencent's own Cookbook scopes AuK to "English and Chinese instruction templates" only, with no other language documented or benchmarked anywhere. Use English or Chinese for anything you need to sound right.

Keep reading

Get new guides by email

One email when we publish new guides and model breakdowns. No spam, unsubscribe anytime.

Add GenLovers as a preferred source in Google