---
title: Ultravox.ai — Real-time, speech native voice AI
description: Ultravox is a real-time voice AI infrastructure layer that powers fast, natural, and scalable voice agents.
published: "Oct 9, 2026, 4:46 PM UTC"
---

[Resources]

[Pricing](./pricing)

[Blog](./blog)

[Sign In](https://app.ultravox.ai)

Speech-Native

Voice AI Agents

# Developer-friendly APIs. Agentic-ready primitives. Human-like conversations. Join thousands of teams building natural, conversational voice AI experiences. 

[Get Started for Free](http://app.ultravox.ai/)

[Demo Ultravox]

# Human conversations are fast, fluent, and flexible.

# Voice AI should be, too.

# We're a research lab and product company dedicated to empowering real-time voice AI experiences.

# We train the world's smartest speech model and run it on dedicated, purpose-built infrastructure.

We're the secret behind some of the world's best performing voice agents:

— 11x, Head of Growth

# The Voice AI Problem

# Human Speech ≠ Text

# Most Voice AI systems are designed to convert speech to text before an LLM can process it. This approach introduces two problems:  

# 1) Transcribing speech to text adds latency, slowing the conversation before inference can even begin.

# 2) Paralinguistic signals that influence meaning, like tone, cadence, and pitch, are lost when speech is transcribed.  This is why most Voice AI agents feel slow and robotic. Achieving real-time, natural conversations at scale requires a fundamentally new, purpose-built approach to voice AI.

# The Solution

# A Unified Stack For Audio Intelligence

# Ultravox took a first-principles approach to building a real-time voice AI infrastructure layer that can deliver fast, natural, and scalable voice agents. Ultimately, this meant we needed to create and manage our own end-to-end system.  

# We started by training a speech-native model, to ensure paralinguistic cues aren’t lost in transcription. To keep latency low, we manage our full inference stack, so there’s no waiting on external LLMs or shared inference pools. We also manage all our own infrastructure—it’s more work for our team, but it supports our goal of delivering best-in-class voice AI experiences. 

# Fast, Accurate, Smart. Pick Two Three

# Ultravox performs as well as top reasoning models when latency is factored.

Speed vs Intelligence, Big Bench Audio Score, ultravox-v0.7, gpt-realtime, gemini-live, claude-sonnet-4-5, gpt-4.1, nova-2-pro-preview

# Ready To Build? Say Hello to the Ultravox Realtime Platform

# Build with AI agents that can speak, listen, and interact in real time, just like humans do.

[Get Started](http://app.ultravox.ai/)

# Robust APIs

# Developer-friendly REST APIs for easy integration.

# Intuitive Dev Kits

# Powerful SDKs for every major platform across web + mobile.

# Empowering Tools

# Built-in tools to help you build and scale your voice agents.

# Telephony Support

# Built-in integrations with the largest telephony providers.

# Free To Start

# 5¢ per minute (including TTS) for up to 5 concurrent calls.

[Learn More](./pricing)

# Pay Go

# Perfect for just starting out and experimenting. Pay as you go, with some limits on concurrency.

# $0/month

[Get Started](https://app.ultravox.ai/)

# Pro

# Perfect for companies that are starting to scale. No hard caps on concurrency.

# $100/month

# \*when billed yearly

[Get Started](https://app.ultravox.ai/)

# Enterprise

# Designed for massive scale, we'll work with you to outline a plan that meets your needs.

# Custom

[Contact Us](./request-a-demo)

# Open Science Makes Humanity Better.

# Ultravox is built on open weight models, and we’re committed to sharing our research and findings in hopes that our work can help humanity move forward.

[View Models on Hugging Face](https://huggingface.co/fixie-ai)

# Core Model

# Ultravox v0.7

# Ultravox v0.7 is state-of-the-art on Big Bench Audio, scoring 91.8% without reasoning and an industry-leading 97% with thinking enabled.

# Dynamic Endpointing

# UltraVAD v0.1

# Our neural VAD model predicts conversation states and turn-taking by recognizing when a user is likely finished speaking, typical pause patterns, and the difference between a thoughtful pause and the end of a turn.

# Speech Generation

# Coming Soon

# We'll share more soon… ;)

Voice Agents for a future.

# A future that offers useful, productive, and accessible AGI will require models that can operate in the fast-paced and often ambiguous world of human speech.

# We believe that any genuine solution to general intelligence must encompass voice as a natural use case, and any optimal solution to voice should help illuminate the path forward to general intelligence. 

# While voice intelligence comes with its own unique challenges, they are not fundamentally different from those of general intelligence. Our team’s mission is to help close the gap between human speech and voice AI, and working on practical solutions in voice provides a concrete way to validate—or invalidate—our broader theories of general intelligence.

# -Zach Koch, Founder

[About Us](./about)

[Blog](./blog)

[Careers](https://careers.fixie.ai)

[Terms & Service](./terms)

[Privacy Policy](./privacy)

[Trust](https://trust.delve.co/ultravox)

©2026 FIXIE.AI

HELLO@ULTRAVOX.AI