[Enhancement] Fully continuous conversation

[mitrokun]

The project is already excellent and very stable overall, but there's one feature I'd love to have:true continuous conversation — the ability to keep talking even when the agent's last response didn't end with a question.
Currently, after LLM's response, both states can occur, and there's no way to tell whether listening is in progress without looking at the screen. One can only guess, which is especially critical for languages ​​without a clear structure for interrogative sentences.

I noticed that this option already exists in one of the forks, but I wasn't happy with its stability/reliability.
I also have my own ready-made logic for this feature that I've been successfully using on ESP32 since last summer. The implementation is actually quite straightforward — the only complication was having to add extra checks after introducing two wake words.

In short:
During the intent_progress phase, check if the agent's response ends with a question mark.
If both conditions are met (in intent_progress + ends with question), automatically start listening immediately after TTS completes.
Additionally, I'd suggest make the listening stage timeout configurable, and cancel it if VAD triggers.
The main ww should restart the session
The stop ww should interrupt any actions

[brownard]

There is certainly lots of scope to improve the continue conversation handling, both on the HA side and device side, but it is a tricky problem to get right i think.

Ideally HA would just have something more reliable than checking for a trailing questionmark to indicate whether to continue listening.

There's currently no VAD in Ava, but if this was added I could imagine an optional mode like you said where it listens for a few seconds after every response to see if any voice is detected then runs if so or cancels if not.

One thing I'm wary of is false activations, I've caught my device having a back and forth with the TV even with the current implementation, the TV playback woke it, then the device kept responding with a question about repeating the command because it didn't understand, then listening and getting background audio from the TV and repeating the loop all over again until I walked in and stopped it. This would be almost guaranteed to occur every activation with any sort of background voice. If it was optional then thats probably OK if the limitations are understood.

I will leave this open and think it over but would want to come up with a 'robust' solution if implemented.

[mitrokun]

One thing I'm wary of is false activations,

This is indeed a problem.

However, it is unlikely that this problem can be solved at the satellite side alone.

I'm experimenting with streaming models on the ASR side so that the most frequently used commands have priority over the rest of the speech stream. But due to the quality of the models, it doesn't work very reliably; a lot needs to be adjusted individually.

Another community member came up with a different, more specialized solution.

As for specialized terminals (VPN, SAT1), the community still doesn't have a solution with a phased array microphone to isolate the speaker. An Android audio system might be even more promising in this regard.

And the solution of muting/ducking the sound throughout the room is crude, but effective. Because if we start a conversation with an LLM agent, , we naturally want to eliminate unnecessary sources of voice.

Closed commercial solutions may allow a combination of these solutions, and handle complex situations in the cloud to decide whether to extend the conversation.