🎙️ SofiaVoice v2.0 — Rs4Machine

AI voice system for low-latency speech interaction

SofiaVoice is a public RS4 Lab experiment covering an asynchronous STT → LLM → TTS pipeline with session-isolated conversation handling.

🇺🇸 English (this file) · 🇧🇷 Português do Brasil

---

📑 Table of Contents

• Overview
• Key Improvements in v2.0.0
• Performance Benchmarks
• Architecture
• Tech Stack
• Running Locally
• API Endpoints
• Contributing
• License
• Contact

---

🎯 Overview

SofiaVoice is a voice-system experiment developed by Rs4Machine within the RS4 Lab. It implements a conversational pipeline that receives audio, transcribes it, generates a response, and synthesizes speech.

The public v2.0 implementation focuses on asynchronous execution, a smaller audio payload, session-isolated state, and explicit API boundaries. The system is designed for measurement and iteration rather than opaque autonomy.

Official positioning: AI Systems Engineer focused on hybrid architectures (LLM + deterministic logic) to eliminate hallucinations and ensure auditability in production.

---

⚡ Key Improvements in v2.0.0

• Asynchronous Pipeline — Full async execution (AsyncGroq + Edge-TTS) without blocking the event loop.
• Neural TTS Migration — Replaced legacy gTTS with Microsoft Edge-TTS (FranciscaNeural), reducing audio payload sizes by 52%.
• Hardened Security — Strict CORS domain mapping, 10 MB payload limits (HTTP 413), and input sanitization against XSS.
• Session Isolation — Independent conversation state handling per request session.

---

⚡ Performance Benchmarks (v1.0 vs v2.0)

Metric / Stage	Baseline v1.0 (LLaMA 3.3 70B · gTTS)	v2.0 (openai/gpt-oss-20b · Edge-TTS)	Optimization Gain
STT Latency (Whisper V3)	~0.84s	~0.84s	—
LLM Latency (LLaMA 3.3 70B → openai/gpt-oss-20b)	~0.70s	~0.70s	—
TTS Engine Synthesis	~4.41s	~1.60s – 2.50s	~40% faster
Audio Payload Size	64.5 KB	30.8 KB	52% smaller payload
Execution Architecture	Synchronous / Blocking	Pure Async / Non-blocking	Zero thread blocking

---

🏗️ Architecture

sofiavoice/
├── frontend/                        → Next.js 15 (Vercel)
│   ├── app/sofia-voice/page.jsx     → Main orchestrator
│   ├── components/SofiaVoice/       → Voice interface components
│   ├── hooks/useSofiaVoice.js       → Web Audio API + API client
│   └── styles/                      → Custom UI animations
│
└── backend/                         → Python 3.14 + FastAPI (Render)
    ├── main.py                      → FastAPI instance, CORS & security middleware
    ├── config.py                    → Environment configuration
    ├── routers/voice.py              → Async voice pipeline routes
    └── services/
        ├── stt.py                   → Whisper Large v3 via Groq
        ├── llm.py                   → openai/gpt-oss-20b via AsyncGroq
        └── tts.py                   → Edge-TTS (FranciscaNeural)

---

🛠️ Tech Stack

Layer	Technology	Specification
Frontend	Next.js 15 + React 19	App Router + Web Audio API
Styling	CSS Modules / Tokens	Rs4Machine design system
Backend	Python 3.14.2 + FastAPI	Asynchronous ASGI execution
Speech-to-Text	Groq API	Whisper Large v3
Intelligence	Groq API	openai/gpt-oss-20b via AsyncGroq
Text-to-Speech	Edge-TTS	Microsoft neural voice (pt-BR-FranciscaNeural)
Deployment	Vercel (frontend) + Render (backend)	Public deployment

---

🚀 Running Locally

Backend Setup

cd backend
python -m venv .venv

# Linux/Mac
source .venv/bin/activate
# Windows
.venv\Scripts\activate

pip install -r requirements.txt

Create a .env file inside backend/:

GROQ_API_KEY=your_groq_api_key_here
ALLOWED_ORIGINS=http://localhost:3000

Start the asynchronous API server:

uvicorn main:app --reload --port 8000

Interactive API documentation: http://localhost:8000/docs

Frontend Setup

cd frontend
npm install
npm run dev

Create a .env.local file inside frontend/:

NEXT_PUBLIC_API_URL=http://localhost:8000

Open the application at: http://localhost:3000/sofia-voice

---

📡 API Endpoints

Method	Route	Description	Status Code
GET	/health	API operational check	200 OK
POST	/api/transcribe	Audio file → text transcription	200 OK / 400 Bad Request
POST	/api/chat	Text prompt → AI response	200 OK / 400 Bad Request
POST	/api/speak	Text prompt → neural Base64 audio	200 OK / 400 Bad Request
POST	/api/voice	Full end-to-end pipeline	200 OK / 413 Payload Too Large

---

🤝 Contributing

Contributions, issues, and documentation improvements are welcome.

1. Fork the project
2. Create a feature branch (git checkout -b feature/your-change)
3. Commit your changes using Conventional Commits
4. Push the branch (git push origin feature/your-change)
5. Open a Pull Request

---

📄 License

This project is licensed under the MIT License — see LICENSE for details.

---

📬 Contact

Raphael Mendes
AI Systems Engineer & Founder · Rs4Machine

• 📧 python.dev.raphael@gmail.com
• 🔗 LinkedIn
• 🌐 Portfolio

SofiaVoice v2.0.0 · RS4 Lab · September 2026