Ultravox

Ultravox is a real-time voice AI platform for building natural conversational voice agents. Its speech-native model processes audio directly instead of transcribing speech to text first, which cuts latency and preserves tone and rhythm. Developers get REST APIs, SDKs, telephony integrations, and built-in agent tools. The hosted service starts free with 30 minutes of calls, then charges per minute. The underlying model is also MIT open-source. In late 2026, Ultravox was acquired by Inworld. Benchmarks show strong results on audio understanding tasks.

Key Features

1. Speech-native model

Audio goes straight into the model with no transcription step. Conversations feel faster and more natural.

2. UltraVAD turn detection

Neural voice-activity detection predicts when a speaker finishes. Interruptions and pauses are handled gracefully.

3. Low-latency infrastructure

Purpose-built servers run the full inference stack. You never wait on shared external model pools.

4. Developer APIs and SDKs

REST APIs and SDKs cover web, mobile, and telephony. Built-in tooling connects agents to your systems.

5. Benchmarked quality

The v0.7 model scores 91.8 percent on Big Bench Audio, rising to 97 percent with thinking enabled. Results are published openly.

Frequently Asked Questions (FAQ)

How much does Ultravox cost?

The first 30 minutes of calls are free. After that, usage is $0.05 per minute, with Pro at $100 a month and custom Enterprise plans.

What does speech-native mean?

Most voice AI converts speech to text, reasons, then converts back. Ultravox reasons on audio directly, keeping speed and vocal nuance.

Is the model open source?

Yes. The Ultravox model repo is MIT-licensed on GitHub. The hosted platform with telephony is the paid product.

Who owns Ultravox now?

Inworld acquired Ultravox in October 2026 to strengthen its speech-to-speech research. The product continues under Inworld.

Conclusion

Ultravox is a strong choice when voice quality and latency matter. The speech-native design is a genuine technical edge over transcription pipelines. Free starter minutes make testing cheap. With Inworld's backing, the platform looks stable for builders planning real voice products. Test the free minutes against your hardest conversation. Compare its latency with your current stack.