New PE devices freshly unboxed and set up, all 3 dead out of the box, including fresh reflash

[coarse dock]

I've had these in a box for several months, just unboxed today, set them up in one of my existing hass units, all three connect to the mobile app, connect to the wireless network, pull updates, get to the twinkling blue ring, and then... nothing. No response to voice input (yes, I checked the switch), and show up as 'Unavailable' in the app after being discovered and added.

I guess it's unstated, but I suppsoe the 3.5mm port is intended to have an external mic and speaker attached to it?

[paper sail]

what type of home assistant install do you have?
is it up to date?
is your network standard or do you have a more complex setup (e.g. with VLANs)?
NB: the switch is the hardware mute switch and not power related.

[paper sail]

its a line out connection for attaching to external speakers

[coarse dock]

Everything is current/up-to-date on the hass side, and the PE side

I can see the units in hass, and they're set up there

[paper sail]

that is not up to date

[coarse dock]

Units 1 and 2 appear DOA out of the box, setting up unit 3 now

[paper sail]

core 2025.12.5 is latest

[coarse dock]

Weird, it says there's no newer updates available

Yep, no updates available

Trying to connect it to one of my other hass units, one moment

[paper sail]

where are you looking?

[coarse dock]

Neat, the Voice PE flasher crashes Brave and Chrome right at 99% of the initial flash, every time (on macOS)

Dead end with the other (non-containerized) hass instance

[paper sail]

first time i am hearing of that, no idea do you have another box you can flash on?

[paper sail]

is this system up to date?

[coarse dock]

Looks like the hass update server is down at the moment, throwing errors trying to update that instance to the latest core version, I'll try again later

So many papercuts with this 'simple' process

[sturdy folio]

try diff cable, check usb socket?

[paper sail]

its not. updates run through github.

[coarse dock]

"Failed to perform the action update/install. Error updating Home Assistant Core: 'HomeAssistantCore.update' blocked from execution, no host internet connection"

[paper sail]

is there a reason you have multiple HA setups running? currently everything you are saying is all over the place.
you have multiple setups of different types. which don't seem to be up to date. you didn't answer my question on your network setup.

I am happy to try and help you but we need to narrow down what you actually have and want.

[coarse dock]

I have 2 primary hass setups, one is on a pi4, containerized, manages my solar deployment and some other automation, the other is running in a VM on my vmware infra, all of them are on the same L2 subnet, a /22

Updating the containeized instance now, took forever to pull the latest container image

No complex/elaborate networking in this specific scope, no

[sturdy folio]

same hostnames?

[paper sail]

why have them separate?
which one do you want to the VPE's working with?
try full rebooting the HAOS VM then trying to update.

[coarse dock]

containerized version is now up-to-date, let me see if I can connect them there

They are separate because they have different roles which cannot overlap

They cannot, and will not ever, be combined into a single unit

[paper sail]

okay if you want them separate for a reason then that's fine. i am just trying to work out what is happening really.

[coarse dock]

Looks like with the latest hass core (containerized version), it disables transport encryption for ESPhome, the previous version did not. I proceeded without it for now.

[paper sail]

disables transport encryption for ESPhome
what exactly do you mean here?

[paper sail]

for container installs you are unable to use addons. so you will have set up STT and TTS services (usually whisper/piper if you are looking for a fully local setup) have you set up containers for these?

[coarse dock]

Yes, piper/whisper have been set up, I managed to get this far with the pi4 containrerized version

[paper sail]

and what does the voice-pipeline look like in the voice assistant settings? whichever you have set to your "preferred" assistant.

[coarse dock]

It acknowledges the wake word, but never responds back. voice pipeline looks like:

[paper sail]

ok so your using google as a TTS instead of piper. shouldn't be an issue but just noting it.

[coarse dock]

Yes, because TTS on the pi4 takes about 2 minutes just to render the 'demo' speech playback

Under 1 second for Google TTS

[paper sail]

that's odd, piper is normally pretty quick no matter what its running on.

but no matter

on the voice assistant settings hit the 3 dot menu and go to debug. this shows the trace of the pipeline call

also when you call the wakeword is it then "listening" (slow led clockwise spin) or does it just flash led and stop.

[coarse dock]

No audio out, other than the chime when the wake word is heard, but it does appear to be 'responding' ... somewhere

[paper sail]

towards the bottom of the trace. does the "play audio" work? it should play the audio output locally

[coarse dock]

Yes

Plays through the browser, not the PE

[paper sail]

ok. so things appear to be running its just not able to play audio to the vpe

so to explain how things work.
api commands are send back and forth on the api connect established from HA to the VPE
for audio playback it hosts the file to play on the ha web server and then generates and sends a link to the vpe which then creates a new connection from the VPE to the HA server to stream the audio filee.

so what is happening is that it seems it cant use the link its being sent to access the audio file. this can be a number of things.

commonly it can be network settings: vlans/firewalls (but you mentioned they are on same network so this shouldnt be an issue i think)

is your container install using "host mode" networking?

[coarse dock]

It was launched with: docker run -d --name homeassistant --privileged --restart=unless-stopped -e TZ=Etc/UTC -v ~/.hass/:/config -v /run/dbus:/run/dbus:ro --network=host ghcr.io/home-assistant/home-assistant:stable

And inspect shows:

        "NetworkSettings": {
            "Bridge": "",
            "SandboxID": "3a534c4952cf38a05a283c585fd87ce46c5874c3211dc891c955283a10873363",
            "SandboxKey": "/var/run/docker/netns/default",
            "Ports": {},
            "HairpinMode": false,
            "LinkLocalIPv6Address": "",
            "LinkLocalIPv6PrefixLen": 0,
            "SecondaryIPAddresses": null,
            "SecondaryIPv6Addresses": null,
            "EndpointID": "",
            "Gateway": "",
            "GlobalIPv6Address": "",
            "GlobalIPv6PrefixLen": 0,
            "IPAddress": "",
            "IPPrefixLen": 0,
            "IPv6Gateway": "",
            "MacAddress": "",
            "Networks": {
                "host": {
                    "IPAMConfig": null,
                    "Links": null,
                    "Aliases": null,
                    "MacAddress": "",
                    "NetworkID": "53ddf2570b48b2b01283400e8d2e334bf65195cf8befb8e04c0e38b3008c7219",
                    "EndpointID": "c8ad18988af6aeae0590dc58776fc0c29886bc409be55e609ac08e0c1ef9fbaf",
                    "Gateway": "",
                    "IPAddress": "",
                    "IPPrefixLen": 0,
                    "IPv6Gateway": "",
                    "GlobalIPv6Address": "",
                    "GlobalIPv6PrefixLen": 0,
                    "DriverOpts": null,
                    "DNSNames": null
                }
            }
        }

Pretty basic, and yes, host-mode networking, everthing in hass related to this instance, works fine... except this.

[paper sail]

in settings - system - network. there is a section for "home assistant URL" is local network set to auto?

[coarse dock]

It is the IP of the hass instaance itself, the same one I use with the mobile app to connect to and manage these instances

6.148 for the pi4 instance of hass, 4.111 for the vm instance of hass

[paper sail]

if you hit the eye on the local network url is it the same as the one above it?

[coarse dock]

Yes, and when I set it to manual, then click 'Save', it automatically goes to 'Automatic', with the grayed out section above

That is, if I cut and paste the top url into the bottom, and click "Save", it changes it to Automatic, and grays it out

[paper sail]

the vpe makes the "wake sound" right?

[coarse dock]

Yes, the chime is audible when I say "Hey Jarvis"

Ooooo... wait...

When I disconnected it from my laptop and connected the vpe to a battery bank, now it works

[paper sail]

aha, could have been a power limit of the laptops usb port maybe?

[coarse dock]

No, not likely at all.. both devices output exactly the same current (measured with an inline usb meter)

[paper sail]

maybe it just needed a reboot

[coarse dock]

Seems that when connected to the laptop, it treats it as a peripheral

It's still quite slow, asking it the time takes 5-6 seconds to respond (vs. Alexa, which is sub-1-second)

[paper sail]

strange the laptop will only see a usb serial adaptor. it cant access the audio stuff.

[paper sail]

if you look in the pipeline trace you will see how long each step takes.

[paper sail]

so you can see that whisper is your bottleneck here.

if you want quicker (but a bit more limited) STT then you could run "Speech-To-Phrase" instead. (although you cant use STP with llm's if your planning to add that at some point.

or if you have other more powerful systems that you could run the STT engine on then you could do that.

[coarse dock]

It looks like I'll need to know all of the possible questions anyone would ask of vpe, so I can set up each individual automation to respond to them in hass

[paper sail]

for most basic interaction commands stuff works pretty nicely if maybe a little strict on how its said. if you add a "brain" to it as an llm then you get a bit more flexibility

if you want super specific stuff then yeah automations are a way to go.

[coarse dock]

I need functioanl equivalency to Alexa, with weather, wikipedia, pubmed, etc. lookups

[paper sail]

so you will likely need a llm with access to to be able to call online services.

[coarse dock]

I have 45 smart plugs here, lots of various devices, automations, 'scenes', grouped by room, which alexa happily manages (eg: "Alexa, office on" will turn on my cameras, govee lights, rack lighting, overhead lights, power up my streaming setup, etc.)

I'd rather not have to reinvent all of that with brittle, hand-curated automations in hass

[paper sail]

for something like local weather you can have a weather integration which shares its data to a local llm. but if you want the llm itself to search for "live" "online" data then its a bit more complex.

adding something like an open AI or Gemini connection. helps unlock these kind of things. but then you are depending on an external service which your exposing your entities too.

i don't use these myself, I have a local LLM but it's not set up to search for things online for live data.

[coarse dock]

Took me a bit to get my other hass fixed, apparently Supervisor doesn't look at host networking, had to convert it from netplan to NetworkManager, and inject nameservers with 'nmcli con mod', because it also doesn't use systemd-resolved or read /etc/resolv.conf.

[paper sail]

sometimes haos needs a reboot to kick the supervisor into connecting to stuff. i am not sure how it gets itself into a mess but it happens.

[coarse dock]

Yep, had to reconfigure a number of things, but now it's updated, refused to upgrade from the UI, even though all other plugins and automations could upgrade, core would not, until I reconfigured the host to use NetworkManager, and manually injected the nameservers (despite them being correctly defined in netplan)

Ok, let's see if I take one of these other vpe units and connect it to the other hass, if I can improve performance a bit there.

Presumably I'll need whisper and piper + wyoming there as well?

[paper sail]

yes, although theres nothing stopping you adding the whisper/piper instances on your other system to the wyoming. but then you would be in the same bottleneck.

[coarse dock]

ah interesting, on the vm-based hass, it prompts for and installs these, but on the pi4 hass, I had to manually install and configure them as separate docker containers

[paper sail]

the VM is running HAOS which supports addons so it can install them as addons

addons are just containers

[coarse dock]

If this works, I'll just build a dedicated box for the LLMs, and point these at it

[paper sail]

but with some extra wizzardry

[paper sail]

yeah if you have a powerful system you can run llm's/whisper on there and connect to them.

my llm runs on a seperate system to my HA instance.

[coarse dock]

When I add a new, unique vpe to the vm-based hass, after it's set up piper + whisper, why do I have 2 assistants showing up, but only the bottom one is 'valid', and shows the correct debug?

In the pi4 hass, there's only one listed there, "Home Assistant", and that one (on the pi4) shows the active debug logs.

[paper sail]

during the onboarding when you make selections it creates one. the onboarding works slightly differently on haos than it does on container.

if you check the vpe device page then it will propbabyl show its using the "fully local" assistant.

for quicker STT/TTS from a trustworthy source you could also sign up for home assistant cloud which includes STT/TTS cloud service. and having it actively supports the project. you would still need "brain" llm if you want to do more complex stuff though.
NB: a HA cloud sub can only be active on a single HA instance.

[coarse dock]

Progress, and quite a bit faster than the pi4 version

I'll tear down the pieces on the pi4, and use the vm version for now

Thank you for your help getting me this far! 🙏

[paper sail]

cool, no problem. a lot of getting this working is based around knowledge of how things work and how all the pieces fit together. hopefully you are a bit further down that road now.

[coarse dock]

Been a hass/tts/stt user for many, many years (tts for ebooks for decades now, back in the festival days), but the oddities with docker, sandboxing the the workflow in hass, tangled me up a bit, good to know that even though both instances use docker, they're quite different in infrastructure

[coarse dock]

Success!