# Hearty's Booth on your own hardware

The full `studio.py` remains the application: lyric-conditioned songs, 4,160
public tags, User-Named Style, reference remixing, masked replacement, intro/outro
extension, waveform editing, vocal editing, stems, library and WAV/MP3/OGG export.
Its song-duration default is now three minutes, with controls through five minutes.
The optional browser MusicGen clip engine is not an equivalent full-song engine.

## Personal native installation (macOS / Linux)

Extract `Complete-Studio-Code.zip` into a folder on the rendering user's computer.
Install Git and a C++ compiler first. macOS Command Line Tools provide the compiler;
Homebrew can supply Python and FFmpeg: `brew install git python@3.12 ffmpeg`.
On Debian/Ubuntu: `sudo apt-get install git build-essential python3-venv ffmpeg`.

```sh
python3 -m venv .venv
.venv/bin/python -m pip install gradio==6.2.0 'numpy>=1.26,<2' 'scipy>=1.12,<1.17' soundfile imageio-ffmpeg==0.6.0 httpx==0.28.1 huggingface_hub cmake==3.31.6
.venv/bin/python setup-user-device.py --build --download
.venv/bin/python personal_studio.py
```

Open http://127.0.0.1:7875 on that computer. This launcher forces local rendering,
removes inherited cloud routing, and uses the native ACE GGUF engine. No owner
machine or GCP renderer is contacted. Model weights are downloaded once using
the user's internet connection; cache reuse verifies their complete SHA256 hashes.
Quantized core weights total about 2.57 GB. RAM usage is higher than download size.
The runtime compiled and all core weights passed hash verification on the 8 GB
Intel Mac. A 60-second lyric-conditioned trial did not complete within eight
minutes while that host was under sustained memory pressure; it was stopped and
its logs retained. This is not a successful full-song or listening-quality test.
CPU, Apple Silicon Metal, or CUDA is selected at build; Vulkan is an explicit
option for a host with its SDK. One heavy model phase runs at a time; tiled VAE
decode reduces transient memory. Quantization is not a guarantee of equal fidelity.

The GGUF adapter supplies singing, full-song synthesis, source remixing, masked
editing and timbre-reference conditioning. It does not itself supply Demucs,
Seed-VC, RVC or Vevo2. Their existing complete code paths are retained. Install
the full dependencies and voice setup from `UNIVERSAL-BUILD.md` on a supported
Apple Silicon / Linux rendering host to use strict vocal removal, four-stem
separation and the independent singer/lyric editors. These operations fail clearly
when their dependencies are unavailable; no model or cloud fallback is substituted.
The full PyTorch application can also launch with:

```sh
.venv/bin/python personal_studio.py --runtime pytorch
```

Outputs stay in that user's `~/Desktop/Treblo_Complete_Studio/`. Keep the local
application open during generation. Current named style prompts and reference
voices remain model directions; exact artist identity and Treblo fidelity are
not certified.

## Phones: native runtime required for full synthesis

The upstream ACE C++ project includes `buildtermux.sh` for Android. Its native
engine can use text, lyrics, uploaded source audio and timbre references; it is a
candidate for a phone-native Hearty application. Android compilation, sustained
full-song inference, thermal behavior, memory limits and all neural editing tools
still need physical-device qualification. This package is not a signed Android
app, and the complete Gradio/PyTorch installation is not verified in Termux.

No tested iPhone-native full studio is included. iPhone requires a signed native
app embedding the inference engine, qualified model conversions and a memory-safe
separation/voice stack. Opening the website or installing its PWA does not install
those native libraries. A phone controlling its user's computer is a different
compute arrangement and is not described as rendering on the phone itself.

Sources checked October 10, 2026:

- https://github.com/ServeurpersoCom/acestep.cpp
- https://github.com/ServeurpersoCom/acestep.cpp/blob/master/docs/ARCHITECTURE.md
- https://github.com/ServeurpersoCom/acestep.cpp/blob/master/buildtermux.sh
- https://github.com/ace-step/ACE-Step-1.5
- https://github.com/Emb1ter/ai-music-js (desktop browser only; no mobile or reference/repaint support)
- https://stability.ai/guides/stable-audio-3-prompt-guide (not intended for intelligible vocals)