Nabu Casa Voice Preview - Has a lot of difficulty understanding voice commands

[river saffron]

I have the Home Assistant Voice Preview Edition and I'm trying to get it to recognize speech better, at least as good as the Google mini speaker (which isn't a high bar 😅). The Google devices can still recognize my wake word and command even if a TV is playing in the background, but the Voice Preview struggles with wake words and voice commands.

Any tips on placement or any adjustments to the settings that might help? I have it in a large open room so I have to speak pretty loud for it to hear me. It's near the TV soundbar since that's the only flat surface within USB range of my HA Green device.

[stuck kraken]

You won't get acceptable results having the device near the soundbar. PE isn't that good - and so are all other ESP32/XMOS based devices currently. Background noise, that's not coming from the device itself, cannot be cut completely.

Also when you're saying "at least as google speaker", remember that those devices are result of decade-long research by multi-billion corporation, unlike PE/Respeaker/SAT1...

[stone jetty]

that's the only flat surface within USB range of my HA Green device.

you know it doesnt have to be plugged into the green right?
you can power it from any USB power supply

[river saffron]

Based on the marketing material for the PE, they claim "dual mic and XMOS cuts out most background noise"

[river saffron]

Yes but I assumed it was using the USB for data as well, to minimize latency from bluetooth?

[stone jetty]

it doesnt use bluetooth (except for initial onboarding) everything is done over wifi

[river saffron]

Doesn't the HA Green recognize it as a USB microphone when directly plugged in?

[stone jetty]

no

thats not how it works

[stone jetty]

it can filter out stuff that it is itself playing using the xmos chip

usb is only used for either manually flashing firmware or pulling logs.
bluetooth is only used as an option for initial provisioning of the device
all connectivity for main functions is over wifi

maybe now armed with this info, perhaps can find a better location

but from a realistic point of view, its going to be a while before its "as good as" google/amazon devices etc...

part of wake word detection is needing real world data. home assistant gets a few thousand volunteers. amazon/google dont ask they just slerp millions of audio feeds from peoples houses to train on.

new firmware has an option to make it more sensitive to wake word detection though. i havent tried it yet but i guess the trade off is possibly more false positives

[strange coral]

You can go my route and have it set up where the volume lowers on the soundbar and a snapshot is taken of the previous level when the pe switches to listening and restores volume upon idle.

I had the same problems.

[lunar goblet]

Just a note i hsve no problems with pe nor atom echo near a music source. But i use easy commands Levenshtein distance and LLM to filter out spelling errors.

[night ore]

@strange coral I'd like to implement this, can you give an example of how you snapshot the pre-listenting volume?

[stuck kraken]

@night ore check this blueprint: https://github.com/formatBCE/Respeaker-Lite-ESPHome-integration/blob/main/blueprints/automation/formatbce/duck_players_on_satellite_working.yaml

May import it from here: https://github.com/formatBCE/Respeaker-Lite-ESPHome-integration/blob/main/readme/blueprints.md

[night ore]

thanks!