WaveGen

WaveGen

Generates English or Mandarin Chinese speech from a reference recording, its exact transcript and the text to speak. Offers the WaveGen V-Spec and WaveGen-Flash models on ZeroGPU, with 24 kHz audio output.

Made by LeeSangHoon on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: WaveGen is a text-to-speech app that produces English or Mandarin Chinese speech. You give it a reference recording, the exact transcript of that recording, and the text you want spoken. The model writes 24 kHz waveform audio directly. Two models are offered, WaveGen V-Spec and WaveGen-Flash.

Who it suits: Speech researchers and developers who want to try a text-to-speech model that works from a reference recording, and people who need short English or Mandarin clips.

Industry | Who uses it | How they use it | Example
Speech research | Researchers | Compare the two WaveGen models on the same input | Run V-Spec and Flash with one recording
Software | Developers | Study how a ZeroGPU speech app is set up | Read app.py and runtime.py
Media | Content makers | Make a short spoken line from a reference recording | Type a sentence and download the WAV file
Education | Language teachers | Hear English or Mandarin text read aloud | Generate a short passage in each language

Anyone who needs speech in other languages, or one clip longer than 30 seconds, can skip it.

Where to use it: It runs in these places.

• A web browser, on the Space's page at Hugging Face
• Windows, with Python 3.12.12 and a CUDA GPU for a local copy
• macOS, with the same needs
• Linux, with the same needs

Use case | Who | What they get back
Generate speech from text | Content makers | A WAV file of the spoken text
Compare the two models | Researchers | Output from WaveGen V-Spec and WaveGen-Flash
Run the demo locally | Developers | A local copy that needs a CUDA device

How to set it up: Using the Space online needs only a browser. A copy on your own computer needs the following.

• Python 3.12.12 and pip
• The huggingface_hub package, which gives you the hf command
• A CUDA GPU, since synthesis uses CUDA BF16 autocast
• The system packages listed in packages.txt, which the hosted Space installs with apt-get
• A Hugging Face token with read access named HF_TOKEN, only while the model repository is private
• Disk space for the repository, which holds 36 files

1. Install the huggingface_hub package.
2. Download the Space's files into a local folder.
3. Move into that folder.
4. Install Gradio at the version the Space uses, then install the packages in requirements.txt.
5. Start the app.

pip install -U huggingface_hub
hf download LeeSangHoon/WaveGen --repo-type space --local-dir WaveGen
cd WaveGen
pip install gradio==6.17.3
pip install -r requirements.txt
python app.py

How to use it: Open the Space's page, or run the app on your computer and open the page Gradio shows. Supply a reference recording, type its exact transcript, and type the text to speak. Choose a model and generate.

python app.py

Tool | What it does
WaveGen V-Spec | The recommended model, selected by default, with 50 Euler steps
WaveGen-Flash | A faster model that uses four Euler steps
WAVEGEN_MODEL_REPO | A Space variable that sets which model repository to load
WAVEGEN_MODEL_REVISION | A Space variable that pins the model revision

Example: You upload a short English recording, type exactly what it says, then type a new sentence to speak. You press generate with V-Spec and get a WAV file of the new sentence, built from your reference recording.

For newcomers and experts: Text-to-speech turns written words into sound. Here the model also listens to a reference recording to shape the result. The first small thing to try is the included example with the default V-Spec model.

The model generates waveform audio directly. Weights stay in FP32 while synthesis runs under CUDA BF16 autocast. V-Spec samples with 50 Euler steps and Flash with four. Both start from an initial Mel temperature of 0. The reference plus generated speech must fit within 30 seconds. The app does not truncate or split text for you, so split long text into shorter passages.

You give a reference recording and its transcript
You type the text to speak
The chosen model generates 24 kHz audio
The app returns a WAV file

Good to know: The app handles one generation at a time. On ZeroGPU, how long you wait and how much you can run each day depends on your own Hugging Face account.

Title: Euler steps per model
WaveGen V-Spec: 50 steps
WaveGen-Flash: 4 steps

• License: mit
• Cost to run: a Hugging Face token only while the model repository is private, otherwise none named
• Data it reaches: your reference recording, transcript and text, plus the model files that the Space downloads at startup
• Not for: languages other than English and Mandarin Chinese, and clips whose reference plus output run past 30 seconds
• Hardware: the Space is set to zero-a10g at Hugging Face

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/LeeSangHoon/WaveGen
• Files: https://huggingface.co/spaces/LeeSangHoon/WaveGen/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.