Early voice assistants waited for you to finish, turned the recording into text, and read a reply aloud. New systems listen while speaking. A spoken correction can arrive before the first task has stopped.
New voice systems can listen and speak at the same time.
Two processing patterns
Turn-based voice
Speech→Text→Answer→Speech
Each handoff adds latency and can lose timing, tone or interruption context.
Full-duplex voice
Listen↔Speak
↳ Reasoning or tools continue separately
A native audio model processes incoming and outgoing sound together. WebRTC keeps the audio stream open while slower work runs behind an asynchronous boundary.
Claude now has hands-free product Voice with interruption support. Anthropic publishes no comparable realtime audio API or details of the app’s internal audio pipeline.
In Speak’s early evaluation, GPT-Live-1 cut interruptions during thinking pauses by almost 80% compared with its previous turn-based system. OpenAI and Speak reported the result.
A native audio model handles both directions over a continuous WebRTC stream. Reasoning and tools run behind an asynchronous boundary, so new speech can arrive while work continues.
MicrophoneContinuous encrypted audio leaves the device.Session beginsMuting does not close GPT-Live.
→
WebRTC + turn detectionTransport, noise handling and voice activity stay responsive.No separate bandwidth feeRealtime API; provider infrastructure still runs.
→
Audio model ↔ speechThe model consumes sound and generates sound.$0.05 / session minuteGPT-Live-1 voice layer.
→
Conversation stateThe session clock continues during silence and backend work.Ten minutes = $0.50Transcription alone: $0.045 with gpt-transcribe.↘ Reasoning, search and tools: billed separately
December 202460% cheaperOpenAI cut full-model audio rates to $40 input and $80 output per million tokens. A smaller realtime model launched at $10 and $20.Efficiency improvements + smaller modelAugust 2025Another 20%The full realtime model moved to $32 input and $64 output. Context limits and truncation reduced repeated conversation history.Less repeated contextJuly 202625% lower p95 latencyImproved caching reduced reported tail latency. The mini path remained $10 input and $20 output per million audio tokens.Caching + model choice
Native audio removed serial transcription and speech-generation handoffs. WebRTC improved transport. Smaller models, cached input and shorter retained context reduced work inside the session.
OpenAI offers two billing models. GPT-Live-1 charges for the full session, including silence and backend work. The Realtime API charges audio tokens and resends conversation history on each turn. Smaller models, cached input and shorter context reduce the token bill.
The application must track the conversation and the task separately.
Illustrative request: change a reservation
Conversation
“Book a table at 7.”“Actually, 8.”Assistant hears the change
Booking task
7 PM request sent7 PM result may arriveCheck before booking 8 PM
Press the button while the 7 PM booking request is in progress.
The app must check whether the first booking was committed before sending another. The sequence is illustrative, not a reported incident.
A spoken correction changes the conversation. The application still has to learn whether the first booking completed before it sends a second request.
OpenAI states that speech interruption does not automatically cancel backend work. Google exposes interaction status separately from spoken utterances.