Speech-to-Speech Model
also called S2S model, Real-time voice model
A model that hears audio and answers in audio, with no text step in the middle.
A model built for live voice conversation: it takes audio in and produces audio out directly, instead of chaining speech-to-text, a text model, and text-to-speech. Skipping the text middle steps is what keeps latency low enough for a phone call, where even a second of delay makes people talk over each other. Modern voice models also reason and call tools inside the call. xAI shipped Grok Voice 3.0, a speech-to-speech model for its Voice Agent API, on October 14, 2026.
The hard parts of a voice model are conversational, not just acoustic: detecting when the caller has finished speaking, handling interruptions and background noise, and keeping the turn-taking natural. Benchmarks built for this measure pauses and interruptions (Full Duplex Bench) and simulated support calls with accents and noise (tau-Voice), rather than transcription accuracy alone.
Voice models are usually API-only services, priced per minute rather than per token, which changes how you budget them: cost scales with call duration, not with output length. The alias question matters here too: xAI moved callers on the grok-voice-latest alias to 3.0 automatically, so pinning a versioned id is what keeps a voice deployment stable.
The support line that stops talking over callers
Choosing between a speech-to-speech model and a transcription-plus-text chain for a support line.
Pick speech-to-speech when callers need to interrupt and be interrupted: the text chain adds a beat at every handoff. Judge the choice on turn-taking (Full Duplex Bench-style pauses and interruptions) and simulated support calls (tau-Voice-style accents and noise), not on transcription accuracy, and budget by call duration rather than tokens.
Judging a voice model by transcription accuracy alone. The failure modes that lose callers are pauses, talk-over, and wrong turn-taking, not misheard words.
- docsRelease Notes: Grok Voice 3.0 ↗
October 14, 2026: grok-voice-3.0 available for Speech to Speech at the same price as 2.0; grok-voice-latest now points to 3.0, so sessions on the alias upgrade automatically; 2.0 deprecated.
- reportxAI Introduces Grok Voice 3.0 Speech-to-Speech Model for Voice Agents ↗
Secondary coverage of the October 14, 2026 launch: $0.08 per minute in the Voice Agent API, Artificial Analysis Speech-to-Speech Index 86.9, Full Duplex Bench 97.0 percent, tau-Voice 71.6 percent, all as reported by xAI.