Repository navigation
Getting Started
Start LocalSoundsAPI (see Installation) and open http://127.0.0.1:5006 in your browser.

Each tool lives in its own collapsible panel. Click a panel's title bar to open it.
| Panel | What it's for | Guide |
|---|---|---|
| Upload & Transcribe | Add reference voices, transcribe audio with Whisper | [[Upload & Transcribe |
| XTTS-v2 TTS | Voice cloning and 58 built-in speakers, 17 languages | [[XTTS Voice Cloning |
| Fish Speech Voice | Highest-quality voice cloning, 17 languages | [[Fish Speech |
| Kokoro Voice | Very fast TTS, 54 voices in 9 languages, runs on CPU too | [[Kokoro TTS |
| Stable Audio Music/SFX | Sound effects and short music from a text prompt | [[Stable Audio |
| ACE-Step Music | Songs with lyrics, up to 60 seconds | [[ACE-Step Music |
| Video Production Studio | Narration + subtitles + backgrounds → video | [[Video Production |
| Chatbot | Local LLM to help write scripts, prompts and lyrics | Chatbot |
| Master Cheat Sheet | Recommended settings for every engine | below |
Kokoro is the quickest way to hear the app working: its model is small (~350 MB) and it runs on CPU or GPU.
- Open Kokoro Voice.
- Pick a device at the bottom left (
cpu, or a GPU number) and click Load. The badge turns green: LOADED. The first load downloads the model. - Choose a voice, type some text in Text to Speak, and click Generate.
- Watch the progress bar, then press ▶ on the player or click Download.
That's the pattern for every engine: pick device → Load → set up → Generate.
Speech panels share the same three-column layout:
- Left: voice and model. Voice picker, language, device, Load/Unload and the status badge. The ⚙ gear opens the inference settings (temperature, accuracy tolerance, Verify with Whisper, …).
- Center: your text, plus the API Payload button.
- Right: output. Speed, Clean (de-reverb), De-ess, output format, save path, Generate / Stop and the results.
When you click Generate, the panel locks its controls (everything except Stop) and shows a progress bar with the stage, chunk or step count, elapsed time and an ETA:

- Preparing: loading the model, splitting text into chunks
- Generating: real percentage (Chunk 3/9 for speech, Diffusion step 42/100 for Stable Audio)
- Cleaning up: joining chunks, post-processing, saving
Stop cancels only your own job. Jobs started from another tab or by a script show up as started elsewhere. Different engines can generate at the same time; see Concurrency & Progress.
Leave Save Path empty to just play and download the result. Type a project name (e.g. MyAudiobook) and output is saved to projects_output/MyAudiobook/ together with a job.json that allows job recovery: if a long job fails part-way, type ##recover## as the text with the same save path and press Generate. Only the missing chunks are redone.
Every engine panel has an API Payload button that shows the exact JSON your settings would send to the REST API. It's the fastest way to script anything you did in the UI:

See API Reference.
The toolbar at the top saves and restores every slider, text field and choice in the UI:
- Type a name in preset name and click Save Settings.
- Pick it from the dropdown later to restore everything.
Presets live in settings/ as JSON files, so you can copy them to another machine.
Not sure what temperature or tolerance to use? The Master Cheat Sheet panel at the bottom lists recommended settings for every engine:

| Folder | Contents |
|---|---|
voices/ |
Reference voice clips for cloning |
projects_output/ |
Saved projects, job.json files, final audio and videos |
output_tts/ |
Temporary output (cleared on startup by default) |
models/ |
Downloaded AI models |
settings/ |
Presets |
brain/ |
Chatbot history and system prompts |
LocalSoundsAPI · Latest release · PDF manual · Issues · MIT licensed
Get started
Speech
Music & sound
Production
Developers
Help