Live Alpha • Streaming at 120ms

Voice AI that keeps up with real conversation

We built a streaming speech engine that responds in under 150 milliseconds — fast enough that people forget they're talking to a machine. Starting with native Uzbek, a language the big platforms still can't handle right.

Try the Voice Studio → Get API Access
AUDIO ENGINE RUNNING
FIRST BYTE: ~124ms • 48kHz
< 120ms
First-Byte Latency
99.4%
Word Accuracy (Uzbek)
48 kHz
Studio-Grade Output
100%
Streaming Architecture
Live Demo

Hear it yourself

These aren't pre-recorded clips. Pick a voice, type any Uzbek text, and listen to what our proprietary VoiceUZ neural engine produces in real time.

1. Pick a Voice

M

Muslima (Natural)

Uzbek • Warm tone

S

Muslima (Studio)

Uzbek • 48kHz broadcast

A

Atlas (Conversational)

Multilingual • Agent-ready

T

Turbo (Fast)

Uzbek • <110ms edge

2. Enter Text

Greeting Support Product
ACTIVE VOICE Muslima (Natural)
Time 00:00 / 00:00
Latency 124 ms
Sample Rate 24 / 48 kHz
🎙️

Tap to Start

Simulates live audio streaming through our acoustic encoder. Words appear as they're recognized.

WHAT YOU GET

  • ✔ Auto punctuation & capitalization
  • ✔ Works through background noise
  • ✔ Word-by-word WebSocket streaming
ENGINE: CONFORMER-WHISPER HYBRID ● STREAMING

Click the microphone to start the streaming demo.

Latency: 112ms Accuracy: 99.4% Language: Uzbek (uz-UZ)
Why Speed Matters

150ms is the line between natural and awkward

Linguists have known this for decades: pauses over 250ms in conversation feel wrong. Your users notice it even if they can't name it. Most voice APIs take 800ms–2 seconds to go from speech input to audio response. We got it under 150.

⚡

Stream-First Architecture

We don't wait for the full sentence. Audio starts generating the moment the first tokens arrive from your LLM.

🚀

CUDA Graph Execution

Zero CPU-GPU sync overhead. We pre-compile the entire inference graph so there's no memory allocation at runtime. 35x real-time on A100.

💎

It Still Sounds Human

Speed means nothing if it sounds robotic. Our models are trained on natural breathing patterns and prosody — not just phonemes.

How We Compare

Traditional Cloud Pipeline ~1,400ms
1400ms
Typical Voice API (Batch) ~750ms
750ms
VoiceUZ Streaming ~120ms
120ms

Measured end-of-speech to first audio byte over WebSocket. Standard fiber connection. No cherry-picking — this is our p50.

Language Coverage

Uzbek isn't an afterthought here

35 million people speak Uzbek. None of the major voice platforms handle it well — if they handle it at all. We built our acoustic models from scratch for Central Asian phonology, and we're expanding from there.

🇺🇿

Native Uzbek Phonetics

Correct handling of O', G', Q, X, H, and ng sounds that other engines butcher. Automatic Cyrillic↔Latin transliteration built in — your users write however they want.

🎭

Voice Cloning from 3 Seconds

Give us a clean 3-second clip and we'll reproduce the speaker's accent, emotion, and timbre. Works across Uzbek and English with consistent quality.

🌍

Expanding Across Turkic Languages

Same architecture scales to Kazakh, Kyrgyz, Azerbaijani, and Turkish. Shared model weights mean we don't start from zero for each new language.

For Developers

Five lines to add voice to your app

WebSocket streaming, REST fallback, SDKs for Python and Node. Connect your LLM output directly to our TTS stream — no batching required. 

# pip install voiceuz
import voiceuz

client = voiceuz.Client(api_key="vuz_live_secret_key")

# Stream TTS — audio starts before the sentence finishes generating
audio_stream = client.tts.stream(
    text="Assalomu alaykum! VoiceUZ xizmatingizda.",
    voice="muslima-natural",
    sample_rate=48000,
    latency_mode="ultra-low"
)

# Pipe directly to speaker, phone call, or WebRTC
for chunk in audio_stream:
    speaker.write(chunk)
// npm install @voiceuz/sdk
import { VoiceUZClient } from '@voiceuz/sdk';

const voice = new VoiceUZClient({ apiKey: process.env.VOICEUZ_API_KEY });

// Full-duplex: send audio in, get transcription + TTS response out
const session = await voice.connectLiveAgent({
  voice: 'muslima-natural',
  onSpeechTranscription: (text) => {
    console.log('User said:', text);
  },
  onAudioResponseChunk: (pcmBuffer) => {
    webrtcConnection.sendAudio(pcmBuffer);
  }
});
# Open a WebSocket stream
wscat -c wss://api.voiceuz.com/v1/stream \
  -H "Authorization: Bearer vuz_live_key"

# Send raw PCM audio frames
> [Binary 16000Hz 16-bit PCM]

# Get transcription events back in real time
< { "event": "transcript", "text": "Assalomu alaykum", "is_final": false, "latency_ms": 98 }
Pricing

Start free. Scale when you're ready.

No surprise bills. No minimum commitments. If you need custom volume or on-prem deployment, we'll build a plan around your infrastructure.

Developer

Prototype and test without a credit card.

$0 / month
  • ✓ 60 minutes of audio per month
  • ✓ REST + WebSocket access
  • ✓ Muslima & Atlas voices
  • ✓ Community support (Discord)
Start Building
MOST POPULAR

Pro

For production apps and voice agents that need guaranteed latency.

$49 / month
  • ✓ 5,000 minutes included
  • ✓ Guaranteed sub-150ms first-byte
  • ✓ All voices including 48kHz studio
  • ✓ 3 custom voice clones
  • ✓ Priority engineering support
Get Pro Access

Enterprise

Dedicated GPUs, on-prem options, custom SLAs.

Custom
  • ✓ Unlimited volume & concurrency
  • ✓ On-premise / air-gapped deployment
  • ✓ Custom voice training & full ownership
  • ✓ 99.99% SLA with dedicated support
Talk to Us
Our Team

Small team. Big ambition.

Four engineers building the voice infrastructure that Central Asia needs. We've shipped real-time models, not slide decks.

Founder & CEO

Mukhammadali

Founder & CEO

Product vision, business strategy, and the person who decided Uzbek deserves world-class voice AI.

GitHub →
CTO

Sardor

CTO & Speech Engineer

Owns the VoiceUZ speech architecture, CUDA optimization, and the streaming inference pipeline.

TTS Core →
ML Research Lead

Madina

ML Research Lead

Acoustic model training, Uzbek phoneme modeling, and Conformer-based STT architecture.

STT Models →
Infrastructure Engineer

Jasur

Infrastructure Engineer

GPU cluster orchestration, WebSocket infrastructure, and making sure latency stays under 150ms.

Cloud Infra →
About VoiceUZ

We built this because nobody else would

Every major voice AI platform treats languages like Uzbek as a checkbox — bolted on with bad phonetics and high latency. We decided to fix that, starting with the acoustic models and working up to a full production-grade streaming pipeline.

Our core engine and architecture are live. We're now scaling to larger GPU clusters to handle production traffic and expanding language coverage across Central Asia.

📍 Based in: Tashkent, Uzbekistan • Cloud infrastructure worldwide
⚡ Focus: Real-time conversational voice — STT, TTS, and the pipeline between them
💻 Open Source: github.com/mpython77
VoiceUZ

VoiceUZ

Real-time voice AI, built on NVIDIA GPU infrastructure and deployed on AWS.

NVIDIA GPU AWS Cloud
⭐ github.com/mpython77