Compare global voice architectures: modular cascade, integrated platforms and speech-to-speech models. Detailed recognition and synthesis comparisons remain on their dedicated pages.
The metric that counts
Measure end-of-speech → first audio, then review median, P95 and jitter. A vendor latency or isolated first audio does not describe felt silences, recoveries and variation.
The infrastructure layer
WebRTC, turn detection, barge-in, audio routing, reconnection and session memory determine continuity. An S2S model replaces neither this control layer nor its observability.
Stay alert to
Test concurrent streams, backpressure, session cost, data residency and fallback to a controlled cascade. The goal is a stable conversation, not only a fast demo.
ASR → LLM → TTS — modular, controllable, production-ready
What it brings
What to trade off
Direct audio-in → audio-out — lowest latency, natural prosody
What it brings
What to trade off
Values are design references: each experience must measure real latency from end of speech to first audio received.
References from existing voice sheets. Open a row to review advantages, limits and test conditions.
Speech-to-speech
Conversational audio loop with a particular strength in emotional intent. Test it when delivery matters more than modular composition.
Full-duplex · R&D
A research reference for speaking and listening simultaneously, especially useful for interruptions, backchannels and natural turn-taking.
Integrated platform
Unifies STT, LLM routing and streaming TTS-2. It is not a pure audio-to-audio model, but an integrated stack that reduces network hops between components.
Speech-to-speech
A WebSocket speech-to-speech stack that combines transcription, reasoning and audio response. Useful for measuring the benefit of an integrated path against a modular cascade.
Open-weights speech-to-speech
An open-weights path to test for sovereignty and reduced serialization, without attributing to it the control guarantees of a cascade.
Speech-to-speech
gpt-realtime-2.1 continues the speech-to-speech path with, according to OpenAI, better interruptions, noise and alphanumerics. Audio tokens stay at $32 / $64 per million. The mini variant is cheaper. TurnBench below still measures the documented VADs, not Live-1.
Full duplex
Full-duplex voice front end: it listens and speaks at once, and delegates tools and reasoning to a backend model. $0.05/min for that layer, plus the model behind it. The announced turn-taking gain versus gpt-realtime-2.1 is an OpenAI figure. Useful as an architecture comparison, not as a replacement for the French cascade.
A Voice-to-Voice model does not replace WebRTC, interruption handling or observability. Track end-of-speech → first audio, barge-in, jitter, concurrent streams and controlled fallback when a service slows down.
Compare real-time infrastructure responsesFor Où est Ava ?, cascade remains the reference when persona, wording and voice require control. S2S paths are comparative experiments for fluidity and emotion. Decide from end-to-end tests, not from one vendor latency claim.
Configure a stack and estimate its latency