Design voice turn-taking with AI
Distinguish a pause from the end of a request, handle interruptions and keep the next conversational turn consistent with what the caller actually heard.
To design voice turn-taking with AI, look beyond the transcript. A caller pausing to remember a date may sound finished to an eager endpoint detector. Someone saying “no, next week” during playback is doing something different. The voice-turn-taking skill separates these cases and helps specify the expected response.
Stopping audible speech also differs from cancelling a tool operation. A reservation might already have changed even though the caller interrupted the announcement. The workflow considers playback, conversation history and tool state separately so the next response can reflect what actually happened.
What it does
- Define how the system decides that the caller’s turn has ended.
- Distinguish short acknowledgements from genuine interruptions.
- Specify what to cancel when the caller speaks over playback.
- Keep unheard generated text out of assumptions about the conversation.
- Test corrections, overlaps and pauses with suitable audio evidence.
Install and use
- Install the skill and provide the voice system’s supported events and intended conversation rules. A permitted test call or recording is needed to verify audible timing.
- Ask, “Handle a caller correcting a date halfway through our response. Show how playback stops and how the next turn learns what was actually heard.”
- Review the audio alongside the event timeline. Check the tool outcome independently of the spoken interruption.
FAQ
Source
Made by Rob Strayer. Original skill: voice-turn-taking, published on GitHub under the MIT licence.
View on GitHub