Ultravox

Ultravox is a real-time voice AI platform that processes speech natively through a single multimodal model, eliminating the traditional ASR-to-LLM-to-TTS pipeline to deliver conversational agents with sub-300ms time-to-first-token latency. Pricing starts at $0.05 per minute on the managed cloud with a free tier that includes 30 minutes of usage and up to 5 concurrent calls, making it accessible for prototyping before scaling to production.

Unlike conventional voice AI architectures that chain together separate speech recognition, language model, and text-to-speech components, Ultravox ingests audio tokens directly into its multimodal model and produces semantic output without an intermediate transcription step. This speech-native approach preserves paralinguistic cues such as tone, pace, hesitation, and emotion that are typically lost during text conversion. The result is more natural-sounding conversations where the agent can respond to how something is said, not just what is said.

The platform is built around an open-weight model architecture, with model weights published on Hugging Face for teams that need to self-host for HIPAA compliance, GDPR data-residency requirements, or air-gapped deployments. This gives organizations the flexibility to run inference on their own GPU infrastructure, fine-tune models for domain-specific vocabulary and speech patterns, or use the managed cloud API for convenience.

Ultravox supports three primary transport protocols: WebRTC for browser-based real-time audio, WebSocket for server-to-server communication, and SIP for telephony integration with providers like Twilio. This means a single voice agent can serve web visitors, mobile app users, and inbound or outbound phone callers without requiring separate implementations. The platform provides SDKs for Python, JavaScript, and Go to accelerate integration across different technology stacks.

A native tool-calling system allows voice agents to invoke external APIs, query databases, retrieve CRM records, process transactions, and hand off to human agents using structured function calls defined at session start. Combined with RAG integration for dynamic knowledge retrieval, agents can access and relay real-time information during conversations rather than relying solely on training data.

The Pay and Go tier charges $0.05 per minute with no monthly fee and includes the first 30 minutes free. The Pro tier adds a $100 monthly base fee for priority support and no hard concurrency limits while maintaining the same per-minute rate. Enterprise plans offer custom pricing for large-scale deployments, on-premise installation, custom SLAs, and dedicated account management.

Ultravox is best suited for engineering teams building production voice agents who need infrastructure-level control over their voice stack. It serves use cases including enterprise customer service automation, outbound sales qualification, healthcare intake and triage, IVR modernization, in-car voice assistants, and interactive applications where natural turn-taking is essential. Teams that prioritize speed to launch over infrastructure control may find higher-level platforms like Vapi or Retell a better starting point.