Answering your questions about GPT-Live
Use GPT-Live for more natural, interruption-aware voice; keep its prompt focused on delegation and manage task details in the backend.
- Tell GPT-Live clearly when to delegate; put tool instructions and procedures in the backend prompt.
- Use client delegation to control what context the voice model receives and guide the next question without over-delegating.
- Save critical conversation details in your application because automatic context compaction may lose them.
The Q&A gives practical guidance, but the key advice is captured in the notes. · Best for: Developers building voice agents or migrating from GPT Realtime.
OpenAI explains how GPT-Live differs from GPT Realtime and answers developer questions about interruptions, delegation, voices, appointment workflows, and conversation context.
How GPT-Live differs from earlier voice models0:00
- Earlier cascaded systems chained speech-to-text, a language model, and text-to-speech, adding latency.
- GPT Realtime combined audio processing, response generation, and tool decisions in one model, but remained mainly turn-based.
- GPT-Live adds full-duplex listening and speaking, plus delegation of complex work to a backend model or system.
Handling background speech and real-world uses2:43
- GPT-Live uses conversation context to distinguish side conversations from speech directed at it; prompts can tell it to ignore side conversations.
- Examples mentioned: Yelp for restaurant reservations and food orders, Fin for customer support, Speak for language lessons, and Cognition for Devin.
French voices and custom voice options3:39
- The speaker recommends trying the Brise and Sillage French voices.
- Custom voices can be generated from a natural-language description or a recording with the speaker’s consent.
- Prompts can guide accent, tone, pacing, personality, formality, and pronunciation.
Improving delegation reliability4:39
- Keep the GPT-Live prompt short: explain when to delegate, while putting tool details and operating procedures in the backend prompt.
- Use explicit delegation criteria and consult current GPT-Live 1 prompt guidance, which includes the SpawnThinking token.
- Decisions API triggers can monitor for cases where delegation should occur and force it if GPT-Live has not delegated.
Avoiding over-delegation in appointment booking6:31
- Use client delegation to guide the conversation and control when the model delegates.
- Put the immediate next question in the commentary update; put remaining required fields and instructions in the thinking update.
- For an appointment, collect name, date, time, and address, then confirm them before delegating.
Managing context and compaction7:28
- Voice-model context compaction is handled automatically, but critical information may be lost; save it in the application.
- Built-in Responses or Decisions delegation automatically collects context for the backend.
- Choose client delegation when you need granular control over context sent to the backend and updates returned to GPT-Live.
Step by step
- Keep the GPT-Live prompt focused on when to delegate; put tool details and procedures in the backend prompt.
- State clear conditions for delegation. For GPT-Live 1, check the latest prompt guidance, including SpawnThinking.
- If delegation is still missed, use Decisions API triggers to detect when it should happen and force it.
- For multi-field tasks, use client delegation: give the next question in commentary and the remaining fields and conditions in the thinking update.
- Confirm collected fields before delegating, and save critical information in your application.
- Use built-in Responses or Decisions delegation for automatic context handling, or client delegation for finer control.
Tools & resources mentioned
- Brise and Sillage — French voice options mentioned in the video.
- SpawnThinking token — included in updated GPT-Live 1 prompt guidance for improving delegation.
- Decisions API — supports predefined-answer triggers and can be used to force a needed delegation.
- Responses delegation — built-in option that automatically gathers conversation context for the backend.
- Client delegation — option for controlling context passed to the backend and updates sent to GPT-Live.
- Cognition Build Hour — video mentioned as an example of how Cognition uses GPT-Live; a link was promised below the video.
Notes are AI-generated from the video's transcript and may miss nuance.