Liquid Audio GUI

Hybrid local voice chat server — STT + LLM + TTS in one FastAPI app.

Built by Conrad CJ Wilson. Runs entirely on CPU: faster-whisper for speech-to-text, LM Studio for LLM inference, and edge-tts for text-to-speech. No cloud APIs, no GPU required.

What It Does

Component	Tool	Model
Speech-to-Text	faster-whisper	small (CPU, int8)
Language Model	LM Studio local API	lfm2.5-audio-1.5b
Text-to-Speech	edge-tts	en-US-JennyNeural

WebSocket-based real-time audio chat with a browser GUI.

Architecture

Browser GUI (index.html)
    │  WebSocket
    ▼
FastAPI Server (server/server.py)
    ├── faster-whisper  →  STT (speech → text)
    ├── LM Studio API   →  LLM (text → response)
    └── edge-tts        →  TTS (response → audio)

Quick Start

# Clone
git clone https://github.com/cjps4linux-creator/liquid-audio-gui.git
cd liquid-audio-gui

# Install
pip install -r requirements.txt

# Configure
cp .env.example .env
# Set LM_STUDIO_URL, WHISPER_MODEL, TTS_VOICE as needed

# Ensure LM Studio is running with lfm2.5-audio-1.5b at http://127.0.0.1:1234

# Run
python server/server.py

# Open browser
http://localhost:8765

Configuration

Variable	Default	Description
LM_STUDIO_URL	http://127.0.0.1:1234/v1/chat/completions	LM Studio API endpoint
LM_STUDIO_MODEL	lfm2.5-audio-1.5b	Model identifier in LM Studio
WHISPER_MODEL	small	faster-whisper model size
WHISPER_DEVICE	cpu	Device: cpu or cuda
WHISPER_COMPUTE	int8	Quantization: int8, float16
TTS_VOICE	en-US-JennyNeural	Edge TTS voice ID
HTTP_HOST	127.0.0.1	Server bind address
HTTP_PORT	8765	Server port

Prerequisites

• Python 3.10+
• LM Studio running with LFM2.5-Audio-1.5B model loaded
• pip install -r requirements.txt

Requirements

fastapi>=0.100.0
uvicorn[standard]>=0.23.0
faster-whisper>=1.0.0
edge-tts>=6.1.0
sounddevice>=0.4.0
soundfile>=0.12.0
numpy>=1.24.0
requests>=2.31.0
pydub>=0.25.0

Security

• No API keys in code — LM Studio runs locally, no cloud auth needed
• Server binds to 127.0.0.1 by default (localhost only)
• No data leaves the machine
• .env is gitignored

Performance Notes

• CPU-only operation tested on Intel i7-1185G7
• faster-whisper small model: ~1-2s transcription latency
• LM Studio with int8 quantization keeps RAM under 4GB
• edge-tts streams audio chunks for low perceived latency

Roadmap

• Confirm LM Studio audio streaming endpoints
• Complete GUI wiring for multi-turn conversation
• Add conversation history / session memory
• Support for additional Whisper model sizes

License

MIT — see LICENSE

Author

Conrad CJ Wilson — @cjps4linux-creator