I have an introduction message for people when they pick up that goes something like this:
“How can I help you? … You can ask me anything.”
I want to give people a chance to reply - and only say “you can ask me anything” if they don’t say anything for a few seconds.
I’ve tried a few things:
..... ... .. ....
<break time="2s"/> (suggested by chatgpt)
The realtime API speeds through and reads them so fast no matter what I try sticking in there. Any ideas?
I think you would benefit from just capturing the audio of matching voices providing a few introductions, and then using these as prerecorded conversation starters played in variety to match.
This allows you several programmatic options for soliciting engagement without actually sending data to the realtime API, and keeps these turns out of the input context length also. You can keep them brief, or make them interruptible if using effective playback echo cancellation. Have personalized parts generated on startup by chat completions with audio, and saved per login.
you can play audio interruptablly?
Yes, just as you can interrupt ChatGPT in the application, with its constant stream of input audio data being sent to voice activity detection, if you have an application that is closer to hardware, you also can monitor for spoken word.
If you do not have good audio algorithms in place, it is better to stop or mute the input voice buffer when playing audio, so you don’t get a feedback loop of the AI interrupting itself either when using OpenAI’s or your own voice activity detection.
ahhhh - I’m chatting with AI lol. OK so I’m talking about the first message - what it says when you first pick up. Is there a way to get what I want - as in - a cue like elipses … that will tell the AI to pause when reading the intro message?
The AI will write until it is done. In this case, it will speak until it is done, which is much harder to parse into independent parts.
About the only suggestion I might have is to alternately send spoken recorded commands in place of the user, instructing exactly what to respond with. These could be sent at intervals in the absence of actual user speech, at expense.
Beep beep boop boop (actually human-powered with an AI sidekick if needed)