Shadowfetch Voice Studio

Notepad, except the text can speak in any voice you have cloned.

1. Clone a voice — record 20–30 seconds with your microphone, or pick an audio file.
2. Type what you want it to say.
3. Press Speak (or Ctrl+Enter). The speech plays as soon as it is ready; Save Audio keeps a copy.
4. Everything runs locally. No accounts, no telemetry, no cloud. The network is only used to download the models you approve, and offline mode blocks even that.

Voices are like fonts: pick one from the voice menu, type, and speak. Everything else — choosing the clean part of a recording, writing down its words (local Whisper), splitting long text, loading the engine, stitching the sentences together — happens underneath.

• Engines: Qwen3-TTS 1.7B Base (default, voice cloning from a short sample) and optional Chatterbox-Turbo.
• Transcription: faster-whisper, CPU by default.
• Shell: Tauri 2 (Rust) + React/TypeScript; Python workers for audio and inference; FFmpeg for media; SQLite for metadata.

Status: v0.2.0 standalone Linux desktop app. See docs/TEST_REPORT.md for what has been verified on Linux with an NVIDIA GPU.

Requirements

Aimed at Pop!_OS / Ubuntu 24.04-class desktops with an NVIDIA GPU.

Component	Notes
NVIDIA driver (CUDA 12.8 capable, driver ≥ 570)	Use the driver already on the system. No CUDA toolkit install is required.
ffmpeg, libportaudio2	sudo apt install ffmpeg libportaudio2
Python 3.12 for the engine runtime	Created by first-run setup or scripts/bootstrap.sh with uv (no sudo; several GB for the CUDA stack)
Model weights	Offered in-app the first time they are needed, after size, source and license are shown (Qwen ≈ 4.5 GB, whisper ≈ 0.5 GB, Chatterbox ≈ 3 GB optional)
Building from source	Rust, Node 22, and libwebkit2gtk-4.1-dev build-essential curl wget file libxdo-dev libssl-dev libayatana-appindicator3-dev librsvg2-dev patchelf

Install

From a .deb

Build a package with scripts/build.sh (or scripts/build.sh --deb-only), then:

sudo dpkg -i src-tauri/target/release/bundle/deb/*.deb

That installs one launcher (com.shadowfetch.voicestudio.desktop). The Python engine runtime is not inside the package; the app creates it under the data directory on first run.

From this tree (recommended while developing)

scripts/doctor.sh          # what is installed / missing, GPU + CUDA check
scripts/bootstrap.sh       # isolated Python environments (asks before optional Chatterbox)
scripts/install-linux.sh   # build the .deb, install it, and isolate the runtime

scripts/install-linux.sh replaces the installed desktop app and writes a real venv under $XDG_DATA_HOME/com.shadowfetch.voicestudio/runtime (default ~/.local/share/…/runtime) — not a symlink back to the checkout.

Flags: --skip-build to reuse an existing .deb; --with-chatterbox to include the optional engine env.

Development without installing

scripts/dev.sh             # desktop app in development mode
scripts/test.sh            # unit tests (add --real for GPU/model integration tests)
scripts/build.sh --deb-only  # supported Linux package
scripts/build.sh             # also tries an AppImage (best-effort)

The .deb is the supported Linux install. AppImage output is optional: linuxdeploy can fail (FUSE or packaging), even when APPIMAGE_EXTRACT_AND_RUN=1 is set.

First run

The app opens on Speak. The first time you clone a voice or press Speak, Voice Studio offers the local model it needs (name, download size, source, license) and downloads it only when you click Download. After that it works offline.

• Speak — voice menu, a large text editor (saved automatically, restored after a restart), the Speak button, the result with Save Audio, and a short Recent list.
• Voices — your cloned voices: play the sample, use, rename, add a recording, edit the sample, delete.
• Settings (gear) — auto-play, Save Audio format, offline mode, microphone. Advanced holds engine selection, the engine's own settings, seed, pauses, pronunciation, export details, model management, storage, the system check and the multi-take project editor.

Keyboard: Ctrl+Enter speaks, Esc stops.

Where data lives

Paths follow XDG. If the variables are unset, the defaults below apply.

	Default
Projects, recordings, voices, masters, SQLite, engine runtime	$XDG_DATA_HOME/com.shadowfetch.voicestudio/ → ~/.local/share/com.shadowfetch.voicestudio/
Settings	$XDG_CONFIG_HOME/com.shadowfetch.voicestudio/settings.json → ~/.config/com.shadowfetch.voicestudio/settings.json
Regenerable caches (peaks, engine prompts)	$XDG_CACHE_HOME/com.shadowfetch.voicestudio/ → ~/.cache/com.shadowfetch.voicestudio/
Model weights (Hugging Face cache layout)	$XDG_DATA_HOME/com.shadowfetch.voicestudio/models/hf/

Local storage is not encrypted. Treat the data directory like any other private files.

Documentation

• docs/USER_GUIDE.md — Speak, Clone Voice, Voices, Settings
• docs/ARCHITECTURE.md — process layout and data model
• docs/PROTOCOL.md — JSON-lines worker protocol
• docs/MODEL_LICENSES.md — engine and model licenses
• docs/FINETUNING.md — optional Qwen dataset export (no in-app training)
• docs/THIRD_PARTY_NOTICES.md
• docs/TEST_REPORT.md
• docs/screenshots/ — UI layout (browser preview mock)

License: MIT.