Bugs

Sometimes EVI will detect an interruption when there really isn't one (eg just background noise or the user backchanneling), and sometime it fails to recover from these false positive interruptions. We are working on improvements to reduce these issues.

Often EVI 2's words do not match the transcript exactly - it will say something that is not in the transcript, or mis-speak something that is in the transcript. We're currently working on preventing this.

HI Team can you please work on latency minimisation, so conversation should be very smooth, as right now we have to wait till it responds, it does not work when there is noise in the background.

During a conversation we want to update the System Prompt to guide the conversation more specifically. We do that using the session_settings Message. After a new System Prompt is active we need a way to interrupt the model gracefully so it starts responding in the context of the new System Prompt. We can trigger a new response using a User Input message with little or no content, but this will not clean up the current generation and playback gracefully. Suggestion: One way could be to have a flag in the session_settings message that signals EVI2 to gracefully clear generation and respond like a reconnect happened. Workaround: Reconnect manually with new System Prompt and Chat Resume feature

EVI 2 sometimes does not speak the transcript correctly. Sometimes, ah or ums in the transcript turn into screeching noises. In longer responses especially, EVI tends to drop the last phrase or sentence, which is present in the transcript but not spoken.

If EVI is falsely interrupted (eg by some background noise rather than the user voice), leads to a blank user message, then EVI doesn't continue until prompted again by the user. This was fixed on Thu Sep 19 - EVI will now continue naturally after these interruptions.

When I talk to Kora, I say "who" and she thinks I say "two". She should know when I say "two" or "who" not mistake "who" as saying "two".