DuDE

DuDE

Generates a two-person conversation as speech from an XML script, with turn-taking, backchannels and overlapping talk, one voice per speaker. It suits researchers and builders of dialogue audio, for non-commercial use.

Made by penguinfish1688 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: DuDE is a text-to-speech model for full-duplex conversation. It takes an XML script of two speakers and generates both voices together, including turn-taking, short backchannel replies and overlapping speech. It writes a 24 kHz stereo WAV file, with speaker A on the left channel and speaker B on the right.

Who it suits: It suits speech researchers and developers who need natural two-person dialogue audio, for example to build or test conversational systems. The license allows non-commercial use only. Use only voices you have the right to use.

Industry | Who uses it | How they use it | Example
Research | Speech and dialogue researchers | Generate overlapping two-speaker audio from scripts | Making test conversations with interruptions
Software | Voice agent developers | Create training or test audio for conversational systems | Producing sample calls with backchannel replies
Media | Audio experimenters | Prototype dialogue scenes with two distinct voices | Drafting a short interview scene

People who need commercial use, or a single narrator, can skip it.

Where to use it:

• Python 3.12 on a computer with a CUDA-capable NVIDIA GPU
• Dialogue research and data generation pipelines
• Audio tools that play stereo WAV files

Use case | Who | What they get back
Dialogue audio | Researchers | A stereo WAV with one speaker on each channel
Turn-taking tests | Voice agent builders | Audio with interruptions, backchannels and overlaps
Scripted scenes | Experimenters | Two-voice scenes built from an XML script

How to set it up:

• Python 3.12
• A CUDA-capable NVIDIA GPU
• PyTorch and torchaudio 2.8.0, using the CUDA builds that match your system
• Disk space for about 7.7 GB of F32 weights

1. Install the Hugging Face hub package.
2. Download the repository, which bundles inference code, an XML frontend, an audio codec, the model and four voice templates.
3. Install the requirements with matching CUDA builds of PyTorch.

pip install -U huggingface_hub

hf download penguinfish1688/full-duplex-speech-data-engine

How to use it: Write the dialogue as XML, with one tag per speaker turn and markers that say when each part starts. Choose a voice template for each speaker, then run the included inference code. The command for that is not in the text provided here.

Tool | What it does
XML frontend | Resolves turn dependencies into interleaved text for the model
Voice templates | Four reference voices that condition each speaker
Event tags | Mark sounds such as laughs, coughs, sighs and gasps inside the script
Stereo output | A 24 kHz WAV, with speaker A on the left and B on the right

Example: You script a fitness chat in which speaker B asks a question and speaker A replies while B answers in short phrases. You pick two voices and run the inference code. You get one stereo file where the speakers sometimes overlap, as people do in real talk.

For newcomers and experts: Most speech tools read one voice at a time. This one speaks both sides of a chat in one pass, so they can talk over each other. A first small step is to listen to the card's two samples, then read their input scripts.

DuDE first self-distills Qwen3-TTS on separate utterances so that its shared backbone can handle two interleaved audio channels. It is then fine-tuned on about 1,922 hours of synchronized conversation recordings. Both channels decide for themselves when speech ends, while they share one timeline.

XML script for speakers A and B
Frontend interleaves the text
Two channels generate speech together
Stereo WAV comes out

Good to know:

• License: CC BY-NC 4.0, so non-commercial use only.
• Cost to run: none named. You need your own NVIDIA GPU.
• Data it reaches: it downloads files from the Hugging Face Hub, then generates audio on your machine.
• Not for: commercial use. Reference-speaker voices should only be ones you have the right to use.
• The card's sample scripts contain strong language.

On Hugging Face, 13 November 2026

• Task: text to speech
• Parameters: 1.9B
• Weights: about 7.7 GB in F32; about 958 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 101
• Likes: 6
• License: cc-by-nc-4.0
• Languages: en
• Based on: Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
• Tags: safetensors, qwen3_tts, full-duplex, conversation-tts, dialogue, stereo, qwen3-tts, text-to-speech

Links

• Model: https://huggingface.co/penguinfish1688/full-duplex-speech-data-engine
• Files: https://huggingface.co/penguinfish1688/full-duplex-speech-data-engine/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download penguinfish1688/full-duplex-speech-data-engine

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