Silly transcript mistakes that accumulates the err...
# general-englishm
Manuel Saelices
12/18/2024, 6:07 PMWhy VAPI stores the transcribed AI spoken messages into the call instead of storing the actual prompt? It seems silly to me. Except for the first system message and the greeting, e.g., if we instruct the assistant to say "Are you an existing client of Ciao business?" and the LLM produces "Are you an existing client of Ciao business?" so the exact text is sent to the TTS, why at the end we have stored a message like "Are you an existing client of Shaw Business?" ? I understand that the pronunciation may be wrong, what I don't understand is VAPI is storing the message in call.messages after being transcribed back, even if we have the exact prompt before passing it to the TTS. I can understand we do it when the caller speaks, as we only have their audio, but why we do that also for the AI messages?!?
Manuel Saelices
12/18/2024, 6:32 PMTo me it's way better to store the message just after receiving the next phrase from the LLM and before passing it to the TTS pipeline. We need to understand that the minimum error in words like the business name are carried in the next interaction with the AI, so the pronunciation and the transcription will be worse and worse as the call goes on
Manuel Saelices
12/18/2024, 6:33 PMI guess VAPI engineers are using the exact same logic for both human and AI messages, which does not make any sense to me
3 Views