- Since maybe 3 days the live model sometimes doesn’t hear or doesn’t react to what is said, the user has to repeat his part multiple times. The microphone works just fine, the recording is all right. Sometimes the understanding deteriorates even more and model thinks something strange alltogether is said or wrong language is used. The conversation language is German (text modality with TTS). Never seen it before in the last month of my testing with Gemini. I’m using it over Pipecat.
- After a longer conversation, the model goes rogue and creates a waterfall, inventing user answers and answering them. Haven’t seen it before, too.
Would very much like to resolve that, happy to help. These bugs make the very nice model unusable for me in its current form.
Thank you!