Skippy — LocalAI Speech-to-Speech (Push-to-Talk)

A local speech-to-speech desktop app: Hold-to-talk → transcribe (STT) → LLM → TTS → play audio, with a polished DearPyGUI interface.

• Backend: LocalAI (runs on your machine)
• UI: DearPyGUI
• Input: push-to-talk hotkey + on-screen button
• Output: local TTS playback
• Persistence: ./models folder for LocalAI model caching (no repeated installs)

---

Quick start

1) Start LocalAI with persistent models

Linux/macOS

bash scripts/run_localai_docker.sh

Windows (PowerShell)

./scripts/run_localai_docker.ps1

This mounts ./models into LocalAI as /models.

2) Install models (once)

Inside the LocalAI container, install a TTS model (example: VibeVoice):

local-ai run models install vibevoice

List installed models:

curl http://localhost:8080/v1/models

3) Install and run Skippy

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/macOS: source .venv/bin/activate
pip install -r requirements.txt

python -m skippy --config config/default.toml

---

Controls

• PTT hotkey: SHIFT_R (Right Shift) by default
• Or hold the “Hold to talk” button in the UI

Change in config/default.toml:

[audio]
ptt_key = "SPACE"

---

Configuration

• config/default.toml — LocalAI URL, model names, audio devices, hotkey, persona file
• config/personas/*.md — editable personalities (system prompt)

---

Docs

• Setup
• Troubleshooting
• API Notes

---

Project structure

Skippy/
  skippy/               # app code
  config/               # config + personas
  models/               # persistent LocalAI models (ignored by git)
  scripts/              # helper scripts (docker + run)
  docs/                 # documentation

---

Notes

• Skippy defaults to LocalAI's /tts endpoint (often WAV-like output). If your TTS returns mp3/opus, install ffmpeg so the fallback decoder can handle it.
• If streaming looks weird, turn off Stream reply.