---
title: WaveGen
emoji: 🔊
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.17.3
python_version: 3.12.12
app_file: app.py
fullWidth: true
license: mit
models:
- LeeSangHoon/WaveGen
short_description: WaveGen V-Spec and Flash text-to-speech on ZeroGPU
---

# WaveGen

**WaveGen: Truly End-to-End Waveform Generation via Internal Spectral Trajectory Learning**  
Sang-Hoon Lee and Ha-Yeong Choi · Ajou University and KT · NeurIPS 2026

[Source Code](https://github.com/sh-lee-prml/WaveGen) · [Models](https://huggingface.co/LeeSangHoon/WaveGen) · [Demo Page](https://wavegen26.github.io/demo/) · [Benchmark Demo](https://sh-lee-prml.github.io/wavegen-tts-demo/)

Generate English or Mandarin Chinese speech from a reference recording, its exact
transcript, and the text to speak. The model emits 24 kHz waveform audio directly.

| Model | Default sampling |
|---|---|
| **WaveGen V-Spec — Recommended, selected by default** | 1400k EMA, 50 Euler steps, Prompt Anchor, Scheduled K3, CFG 1.25 |
| **WaveGen-Flash** | 48k EMA, four Euler steps, time grid `[0, 0.1, 0.25, 0.5, 1.0]`, CFG 0 |

Both models default to initial Mel temperature 0 and skip the unused final Mel
output head. Inference dropout follows each model's configuration. Model weights
remain FP32; synthesis uses CUDA BF16 autocast. The reference plus predicted
generated speech must fit within **30 seconds**. Split longer text into shorter
passages; the app does not silently truncate it or automatically split it.

## Deploy this folder to Spaces

1. Create a **Gradio Space** under the intended account. Keep it **Private** while
   validating it, and select **ZeroGPU** hardware if the account is eligible.
   This README's SDK settings do not allocate hardware or change visibility.
2. Upload the contents of this folder to the Space repository root, including
   `src/`, the license files, `requirements.txt`, `constraints.txt`, and
   `packages.txt`. The root must contain `app.py` and this README.
3. If the model repository is private, add `HF_TOKEN` under the Space's **Settings
   → Secrets**, using a token with read access to `LeeSangHoon/WaveGen`. Keep the
   token out of source files, public variables, and commit messages. No token is
   needed once the selected model revision is publicly readable.
4. Let the Space build, then test the included example with **both models** and
   verify a downloaded WAV. Make the Space public only when ready for release.

The Space downloads each model's Safetensors checkpoint, YAML, and vocabulary at
startup. The model repository and revision can be set with Space **Variables**:

| Variable | Default |
|---|---|
| `WAVEGEN_MODEL_REPO` | `LeeSangHoon/WaveGen` |
| `WAVEGEN_MODEL_REVISION` | `da53c835a39272c05e01526fe3c10b2b9d4ce60b` |

The default revision pins both model bundles to the same release. Use another
full commit SHA to update them deliberately. Model files and authentication
credentials are not included in this source bundle.

## ZeroGPU runtime

The entry point imports `spaces` before PyTorch, loads the models during startup,
and places their modules on CUDA before serving requests. The synthesis callback
is wrapped with `@spaces.GPU`; model selection does not download weights or reload
a model inside that callback. The app processes one generation at a time and
returns a WAV file to preserve the model's output amplitude.

ZeroGPU allocates a GPU for each request and releases it afterward. Queue time
and daily usage quotas depend on the visitor's Hugging Face account. Running the
app on a local CUDA device verifies model integration, but does not itself verify
Hugging Face's GPU allocation and worker lifecycle. A successful Space build and
on-Space request are still required before claiming ZeroGPU deployment is tested.

See the official [ZeroGPU documentation](https://huggingface.co/docs/hub/spaces-zerogpu)
and [Spaces configuration reference](https://huggingface.co/docs/hub/spaces-config-reference).

## Versions and local verification

The existing development environment is not modified. The Space creates a
separate Linux environment. Its inference and UI dependencies are pinned to the
versions inspected in that environment, including:

- Gradio **6.17.3**
- Spaces SDK **0.51.3** in the current build configuration, managed by ZeroGPU (local SDK **0.50.4** unchanged)
- PyTorch and TorchAudio **2.9.1+cu128**
- Hugging Face Hub **0.36.2**
- NumPy **2.2.6**

The remote Python runtime is **3.12.12**, a documented ZeroGPU runtime. The local
Python **3.12.14** installation remains unchanged. Linux CUDA dependencies are
provided by the pinned PyTorch wheels; **hf-xet 1.8.0** is supplied for the Hub's
Linux download transport. One existing transitive dependency cannot be copied
verbatim into a fresh build: the local **Pillow 13.0.0** exceeds Gradio 6.17.3's
declared `Pillow <13` requirement. The isolated Space therefore pins **Pillow
12.2.0**, while the local installation remains untouched. The platform's automatic
Gradio MCP extra also requires **Pydantic 2.12.5** and **pydantic-core 2.41.5**;
the local versions remain **2.14.2** and **2.50.2**. Exact deployment pins
are inline in `requirements.txt`; `constraints.txt` is the reference manifest,
including the observed platform-managed Spaces SDK. Spaces mounts only the requirements file during installation, so it
cannot include a relative constraints file at that stage. The Torch requirement
uses `2.9.1` for ZeroGPU's configuration parser, with the CUDA 12.8 wheel index.

For local CUDA verification with already downloaded model folders, run this from
the bundle folder using the existing Python interpreter:

```bash
python app.py --local --model-dir /path/to/WaveGen/ckpts
```

The `--local` path uses the same models and sampling settings without requesting
a Hugging Face GPU allocation. No package installation is necessary when using
the already validated development environment.

The `src/` files are unchanged copies of the inference APIs and the generator
modules they import. Shared generator definitions include their original loss
definitions, but training entry points, dataset loaders, evaluation tools,
optimizer state, and checkpoints are excluded. `SOURCE_SHA256SUMS` records the
copied files for comparison with the original source.

## Responsible use and license

Use reference speech only with the speaker's consent and permission to use the
recording. Do not use generated audio for deceptive impersonation, fraud, or
harassment. Identify shared outputs as AI-generated and avoid uploading private
or sensitive recordings to a public demo.

Source code is distributed under the [MIT License](LICENSE), with the original
[third-party notices](docs/THIRD_PARTY_NOTICES.md) preserved. Model weights are
licensed separately under **CC BY-NC 4.0**, as stated in the
[model card](https://huggingface.co/LeeSangHoon/WaveGen).