Skip to content

Releases

WaveGen Studio 0.1 (beta)

WaveGen Studio 0.1 (beta) Pre-release
Pre-release

Choose a tag to compare

View all tags
@sh-lee-prml sh-lee-prml released this 12 Nov 06:30
v0.1.0-beta.1
0be8ec6

WaveGen Studio 0.1 (beta)

A free, local voice studio — no subscription, no paid API.

Free for non-commercial use. WaveGen model weights use CC BY-NC 4.0. Original Studio code is separately MIT-licensed; models, runtime components, and recordings retain their own terms.

This is the first public prerelease, 0.1.0-beta.1: a Windows C++ application for generating and editing English speech locally, with no Python required at runtime.

For the original models, paper, training, and research implementation, please refer to WaveGen.

Install and start

  1. Download WaveGen-Studio-0.1.0-beta.1-app-windows-x64.zip from the assets below. Extract it into a writable folder, such as C:\Apps\WaveGen Studio. The small app ZIP is the starting point; you do not need GitHub's source-code ZIP to run the app.
  2. Install the Microsoft Visual C++ x64 Redistributable from Microsoft if it is not already present. A compatible NVIDIA driver is also required. Microsoft's official runtime guidance.
  3. Double-click Install-WaveGen-Studio.cmd in the extracted folder. It downloads the runtime and all three model packs—Flash, V-Spec, and RE-USE—and verifies their SHA-256 hashes against the release manifest. Keep the full folder together.
  4. When installation finishes, double-click Launch-WaveGen-Studio.cmd. Select Mira, Rowan, or Kai, enter a short English sentence, and generate. The latest audio is beside the composer on wide screens and directly below Generate on small screens.

The default full installation downloads about 3.9 GB. Allow at least 15 GB of free space for initial downloads, extraction, and installation working files; recordings and future update backups need additional space. The application does not automatically launch after installation.

The app and runtime are hosted on GitHub Releases. Native Flash, V-Spec, and RE-USE model packs are hosted in the original LeeSangHoon/WaveGen Hugging Face repository. The installer uses an immutable repository revision, verifies its config.json, and reads that configuration to locate the real model packs. Model and configuration hashes are bound by the release manifest. Verified models and valid cached assets are reused, so an app-only update does not download unchanged models again.

For a smaller Flash-only installation, open PowerShell in the extracted folder and run:

.\Install-WaveGen-Studio.cmd -Models flash

This installs the common runtime and Flash; V-Spec and enhancement require their respective additional model packs. You can add them later. To pin this exact beta rather than selecting the latest published release, add -ReleaseTag v0.1.0-beta.1. The installer does not install a GPU driver. Microsoft runtime installation can be explicitly requested with -InstallVcRuntime.

Included in this beta

  • Flash (4 steps) and V-Spec (50 steps): resident models, long-text splitting, seed controls, speed, and target duration.
  • Your own voice profiles: reference audio plus matching Reference Text, optional images, trash/restore, and three licensed CMU ARCTIC starter references. No private user recordings or history are included.
  • Reference preparation: waveform selection, original preservation, preview, undo/redo, edge-silence trimming, and optional peak adjustment.
  • NVIDIA RE-USE enhancement: optional reference and generated-audio enhancement, with before/after comparison and preserved originals.
  • Speech Edit: keep, remove, or regenerate a selected audio region and save a new result.
  • A local workspace: searchable recording history, text/settings reuse, saved drafts, cancellation, and connection recovery.
  • Four interface languages: English by default, 한국어, 简体中文, and 日本語.

Current scope

Hardware qualification is currently limited to Windows x64 with an NVIDIA RTX 5080 (16 GB). Other GPUs and a lower VRAM minimum have not been established; this beta does not provide supported CPU-only, AMD/Intel GPU, macOS, or Linux application builds.

Speech generation and Reference Text are English-only in this application. Interface language selection does not change that. Native BF16 inference is not bit-identical to upstream PyTorch, and output quality, pacing, voice similarity, and edit boundaries can vary. Enhancement may change timbre or breath; compare with the original. A first request can take longer while a model is prepared.

Models and audio processing run locally after installation. Closing the browser tab leaves the application running; use Quit app to release its memory.

Update and keep your work

Use Quit app, back up the complete data/ folder, then run Update-WaveGen-Studio.cmd. Updates are explicit, preserve personal data and unmanaged files, and back up replaced managed files. Keep this whole application folder when moving it. The Microsoft runtime and NVIDIA driver remain system prerequisites.

The updater requires a managed installation; it does not overwrite an older unmanaged development folder. See installation and update/backup instructions. Checksums verify file integrity relative to the manifest; they are not digital signatures.

Help shape the next version

Report a bug with your version, GPU/driver, selected model, and a reproducible example. Please keep private audio and data folders out of reports.

Voice contributions are welcome: submit a clean sample, its exact Reference Text, and source/license/speaker-permission details through the voice submission form. Samples may be added after review; submission is not automatic inclusion. Voice contribution guide.

Thank you to the WaveGen authors, NVIDIA RE-USE, Mamba, ONNX Runtime, CMU ARCTIC, and the supporting open-source projects. License map and acknowledgments · Source and build instructions

Loading