// docs / getting-started

Getting started

Two ways to run LiveAudio: the prebuilt installer (recommended) or from source with uv. You do not need Python installed — the installer provisions its own Python 3.11.

Option A — Installer (recommended)

Download the latest release from GitHub Releases:

Windows: LiveAudio-Setup-1.2.0.exe (graphical installer).
Linux: LiveAudio-1.2.0-linux-x64.tar.gz — extract, then run ./liveaudio-launcher.

The first run downloads Python and all dependencies (~400 MB on CPU, ~2.5 GB with CUDA) and auto-detects your GPU. Later runs start instantly.

Option B — From source (developers)

Requires uv. Pick exactly one torch extra — cpu or cu121.

bash

git clone https://github.com/plynte-labs/LiveAudio.git
cd LiveAudio

# CPU only
uv sync --extra cpu

# Or with NVIDIA CUDA (exactly one torch extra)
uv sync --extra cu121

# Run the app
uv run liveaudio

# Run the tests
uv run pytest

Verify the install

Installer users: run the launcher with --self-test — it prints the detected device, the resolved paths and uv status, then exits without starting the GUI. Developers: confirm the core libraries import:

bash

uv run python -c "import torch; print('PyTorch:', torch.__version__)"
uv run python -c "import faster_whisper; print('Faster Whisper OK')"

Launcher CLI

The launcher is the primary way to force a backend, update, self-test, or recover an install. Pass these flags to the launcher executable (or, in dev, to uv run liveaudio where applicable):

Flag	What it does
`--device cpu\|cuda`	Override GPU auto-detection; the choice is persisted in `installed.json` and re-syncs the matching torch extra.
`--update [vX.Y.Z]`	Update to a tag, or to the latest release when no tag is given. Reports “already up to date” if current.
`--self-test`	Print the detected device, resolved paths and uv status, then exit.
`--reinstall`	Wipe `app/` and `.venv`, keep the device preference, and bootstrap fresh.
`--headless`	No-GUI stdout progress; used automatically when tkinter is unavailable (servers, CI).
`--src-dir PATH`	Dev mode: bootstrap from a local checkout instead of a release zip.
`--install-desktop-entry`	Linux only — add an applications-menu entry.

GPU auto-detection

On startup LiveAudio checks for an NVIDIA GPU. If one is available it defaults to cuda; otherwise it falls back to cpu automatically. CUDA needs driver ≥ 525 and VRAM ≥ 4 GiB — no manual configuration required.

Profiles

Profiles are FPS-aware presets. Editing a built-in profile turns it into Custom; changes only take effect after you click Apply changes.

Profile	Recommended for
`Fast`	Lowest latency and short phrases.
`Balanced`	Recommended for most sessions.
`Quality`	Higher accuracy, more VRAM and latency.
`Stable Streaming`	Reduces GPU load while gaming or streaming on a busy PC.

Hallucination blacklist

Whisper can invent phrases during silence. LiveAudio filters them with an editable blacklist. The default list (from config.json.example) is:

config.json.example · blacklist

amara.org, subtítulos por, suscríbete, dale like, gracias por ver,
memos, gracias, activar la campanita, ¡Hasta la próxima!, por favor

Words and phrases are comma-separated and fully editable from the Subtitles tab — add the filler your stream tends to hallucinate, or clear entries you want to keep.

Models

Pick a model in the Model & Hardware tab. Larger models are more accurate but slower and use more VRAM; each one downloads once on first use, then runs offline.

Model	First-run download	Notes
`tiny`	~150 MB	Fastest; good on low-end CPUs.
`base`	~300 MB	A step up in accuracy, still light.
`small`	~480 MB	Balanced default for CPU.
`turbo`	~1.5 GB	Highest accuracy here; best on a GPU.

Session files

Every session is saved to disk as a transcript (transcript.jsonl) and WebVTT subtitles (subtitles.vtt), plus a session.json manifest:

sessions/

sessions/
└── session_2026-05-04_143022/
    ├── subtitles.vtt
    ├── transcript.jsonl
    └── session.json