Yes the result is pretty good

[runic lark]

Would you mind sharing how you got it to work? I'm trying to do the same but I cant seem to get it to work properly. I'd be incredibly grateful for the help!

[digital fulcrum]

I keep a buffer of tokens. If the next token starts with a space or is a special character, I sent the previous buffer and create a new on with the new token.

[runic lark]

i'm basically doing the same, here's my code:

the problem is that the splitters get sent to the websocket as a single token, so i guess thats' contributing to the problem

and here is how i handle the gpt stream of tokens and send it to the websocket, but the quality i get is awful

this is my getAudioData() function for reference. here i send the api request to the websocket, i decode the audio and play it as it get streamed

[digital fulcrum]

How horrible we talking? Does it play ok? Or is the voice quality bad?

I don’t do much else but send each word or punctuation with a space at the end in each websocket send.

I think the voice quality is ok for my use.