sayna

Sayna is a high-performance, real-time voice layer built in Rust that unifies Speech-to-Text (STT) and Text-to-Speech (TTS) services into a seamless interface for AI agent frameworks.

For the Non-Technical Reader

Imagine building an AI assistant, but every time you switch between different voice providers (like Google, Azure, or ElevenLabs), you have to completely rebuild how your assistant listens and speaks. Sayna acts like a universal switchboard for AI voices. It allows systems to hear and talk back in real-time across 140+ languages while automatically filtering out background noise and recognizing natural pauses in speech so conversations feel human.

For the Technical Reader

• Architecture: Written in Rust for ultra-low latency, offering unified WebSocket endpoints for real-time bidirectional audio (16kHz, 16-bit PCM) alongside REST management APIs.
• Turn & Noise Management: Integrates Silero-VAD for silence detection, custom ML models for end-of-turn detection, and optional DeepFilterNet for background noise suppression.
• WebRTC Integration: Out-of-the-box support for LiveKit for room-based WebRTC streaming and participant event webhooks.
• Provider Flexibility: Pluggable provider architecture supporting Deepgram, ElevenLabs, Google Cloud (WaveNet/Neural2), and Azure (400+ neural voices), with support for per-session credential overrides.
• Audio-Disabled Mode: Allows developers to test control-plane logic and WebSocket message flows without requiring active API keys or initializing audio pipelines.

Why It Matters

As voice AI agents move into production, avoiding vendor lock-in and controlling latency are critical. Sayna provides a vendor-agnostic abstraction layer that lets developers swap STT/TTS engines dynamically, balance costs, and manage real-time media streams without rewriting core agent logic. Explore the source code on GitHub.

The "Voice AI Space Lab" Idea

The Multilingual Dispatcher Trainer: Build a high-stress simulation platform for training emergency dispatchers. Using Sayna's LiveKit routing and DeepFilterNet integration, you can dynamically simulate bad call conditions (adding or removing background noise) while switching voices across 140+ languages mid-stream to challenge trainees in realistic, unpredictable call scenarios.