Ultravox (formerly Fixie.ai)

Ultravox (formerly Fixie.ai) is a developer-focused voice AI platform that takes a fundamentally different architectural approach to building conversational agents. Instead of stitching together separate speech-to-text (STT), large language model (LLM), and text-to-speech (TTS) services in a sequential pipeline, Ultravox uses a single speech-native model that ingests raw audio and produces conversational output directly. This collapses what would normally be three latency-inducing hops into one, and it preserves paralinguistic signals — tone, pacing, hesitation, emotion — that traditional STT systems strip away when they convert audio into plain text.

The platform is aimed at engineers building production voice agents for use cases like inbound and outbound calling, customer support, scheduling, voice-enabled SaaS features, IVR replacement, and embedded in-app voice assistants. Developers interact with Ultravox primarily through an API and SDKs (JavaScript and others), and the platform is designed to slot into existing telephony stacks via providers such as Twilio, as well as into web and mobile applications via WebRTC. Sub-second response latency is one of the key selling points, putting Ultravox on the same footing as other modern real-time voice frameworks while distinguishing itself by the speech-native model architecture rather than a cascaded pipeline.

Ultravox supports tool calls and function execution, allowing the voice agent to query databases, hit internal APIs, transfer calls, send SMS, schedule appointments, or take any other action a normal LLM agent could take — but inside a live phone or voice conversation. Because the underlying model also has access to acoustic features, it can react more naturally to interruptions, barge-in, and conversational turn-taking than pipeline-based approaches that depend on voice activity detection heuristics. The platform exposes a generous free tier for prototyping and offers usage-based pricing for production traffic, with self-hosted and enterprise deployments available for teams with stricter data, latency, or compliance requirements. The original Fixie.ai brand pivoted from broader agent infrastructure into voice specifically, and the company open-sourced an Ultravox model on Hugging Face, which has helped it gain traction among developers who want either a managed API or the option to run their own inference.

In short, Ultravox is best understood as infrastructure: it does not ship a finished voice product, no-code builder, or canned IVR. It gives developers the lowest-latency, most speech-aware building block they can wire into their own application logic, telephony provider, and tool stack to produce a voice agent that feels closer to a person than a phone tree.