On-Device Voice AI Platform

Voice AI That
Lives On Device

Build private, fast, and intelligent voice experiences with a complete on-device SDK. From wake word detection to speech synthesis — all running locally, with sub-millisecond latency.

Start Building Free Read Documentation

Trusted by teams building the future of voice

AUTOMOTIVE CO SMART HOME HEALTH TECH IOT LABS DEVICE AI
<1ms
Inference Latency
39
Trained Models
10
Voice Engines
4
Languages

Complete Voice AI Platform

10 engines. One SDK. Every platform. Everything runs on-device.

Wake Word

Custom keyword detection with <1ms latency. DS-CNN architecture, 20K params, 82KB models.

Speech-to-Text

Streaming on-device STT with Zipformer and Whisper. Multiple model sizes for any hardware.

Text-to-Speech

Natural on-device TTS with voice cloning support. Kokoro and Piper backends.

Voice Activity

Neural VAD with ultra-low latency. Silero-compatible architecture for reliable detection.

Speaker ID

Verify and identify speakers with ECAPA-TDNN embeddings. Biometric-grade accuracy.

Diarization

Know who spoke when. Real-time speaker segmentation for meetings and conversations.

Noise Suppression

Neural noise removal with RNNoise and DeepFilterNet. Crystal-clear audio in any environment.

Speech Intent

Extract intent directly from speech. Skip the text step — voice commands to actions instantly.

Three Steps to Voice Intelligence

From zero to production in minutes, not months.

01

Install

cargo add saj-speak

One dependency. Rust, Python, Node.js, WASM, C, or Swift — pick your binding.

02

Configure

WakeEngine::from_model("hey_saj.onnx")

Load a model, set a threshold. Use pre-trained or train your own via Console.

03

Ship

engine.process(&audio_frame)?

Runs on-device. No cloud. No API keys. No latency. Ships with your binary.

DEVELOPER EXPERIENCE

Ship Voice Features
in Minutes

Three lines of code. That's all it takes to add wake word detection to your app.

main.rs
use saj_speak::WakeEngine;
 
let engine = WakeEngine::from_model("hey_saj.onnx")?;
engine.set_threshold(0.9);
 
// Process audio frames — that's it
if engine.process(&audio_frame)? {
    println!("Wake word detected!");
}
Rust Python Node.js WASM C FFI Swift

First-Class Multilingual

Not an afterthought. Arabic, Hindi, and Urdu are built into the foundation alongside English — serving 1.09 billion speakers with zero competitors.

EN

English

12 models · Full coverage
Hey Saj · Hey Bella · Ok Saj

عربي

Arabic

12 models · MSA + Gulf
Ya Saj · Ya Bella · Ok Saj

हिंदी

Hindi

8 models · Devanagari
Hey Saj · Hey Bella · Ya Bella

اردو

Urdu

7 models · Nastaliq
Hey Saj · Hey Bella · Ya Bella

Simple, Transparent Pricing

Start free. Scale as you grow. No hidden fees.

Free

$0 /month

For prototyping and personal projects

✓  Wake word engine ✓  VAD engine ✓  2 pre-trained models ✓  Community support ✓  Non-commercial use
Get Started
MOST POPULAR

Pro

$99 /month

For production apps and commercial use

✓  All 10 engines ✓  39 pre-trained models ✓  Custom model training ✓  Console access ✓  Priority support ✓  Commercial license
Start Free Trial

Enterprise

Custom

For OEMs, fleets, and large-scale deployments

✓  Everything in Pro ✓  Per-device OEM licensing ✓  Custom model fine-tuning ✓  MCU / embedded support ✓  Dedicated engineering support ✓  SLA & compliance guarantees
Contact Sales

Build the Future
of Voice

Join developers building private, intelligent voice experiences that run entirely on-device. No cloud. No latency. No compromise.