Why wont the Voice PE recognize any of my commands... in Dutch

[Mariusthvdb (Marius)]

Started out with some hesitation, but finally decided to buy the Voice PE. Support the project.
Figured to throw the family something to play with over the holidays.

But what a disappointment this is. It wont recognize many if not most of the most basic commands we throw at it. Using Home Assistant Cloud, no other settings added, nor do I want other commercial Assistants.

Figured it might be the language, and I should really start experimenting in English (which would have been fair) but according to this it should be ok in Dutch

The example sentences translated to Dutch are all met with: Sorry, I dont understand, or Sorry I cant find a device with the name (and then follows the complete command, not even the name…)

Admitted, the install went fine, updated et all, config flow was almost perfect (had to lookup the host address in the Router myself, but other than that, kudos for that experience)

Because I have exposed all devices I need in Assistant for years to my GH’s, I figured I could just jump in and replace all our commands and throw them at OK Nabu. They are all there in the overview of exposed devices to Voice Assistants.

The assistant hears me fine, the listen sound is responding perfectly, so thats cool too. Even from rather long range

The disaster is when we are actually asking it to do something (in Dutch).
even most basic things like ‘Staat de voordeur open’ , aren’t recognized. It returns ‘Sorry, ik kan open niet vinden.’…
‘Stel Wandverlichting in op 80% helderheid’ it says, ‘Sorry, ik kan geen apparaat vinden met de naam Stel Wandverlichting in’…

I can go on with a list of other replies that are too ridiculous to even imagine it could respond with those… Seems there is no AI at all in this device. Even Google Home is way smarter (and we often are amazed with its stupidity and inability to execute basic commands)

Long rant short, before I return this device, is there anything I should have done to get this Voice PE to a somewhat experimental level of functionality?

Because right now it’s just a bad sounding albeit nice looking device that will probably be neglected over time. Very soon…

[pixi]

And your configuration is?

[Edwin_D]

Has anyone ever asked that question with respect to Alexa, Google or Siri? Those just mostly work for the basic stuff. My experience was mostly the same. Luckily I found out before buying Voice PE.

[Mariusthvdb (Marius)]

What do you mean? I’ve simply installed the device successfully and my assistant exposed entities like I wrote above

[pixi]

I run it locally, and within whisper i can configure the model used, and the beam size wich have impact on the performance and understanding.

It also provides a log, where you can see what voice pe thinks you are saying, using this you can improve your set up.

I have no experience with the cloud part, but i would expect that some more or less the same config options are available.

[AshaiRey]

Have you configured the voice pipeline so that it will use Dutch?

I am using Dutch too and i have only minor problems. Most of them are spelling issues.

Voice assistant troubleshooting

[Mariusthvdb (Marius)]

i have not set any other settings no, I figured lets try. it out of the box, with all of my locale settings automated.

Will check the options you both mention and report back, thanks.
But given the screen above, at lest the Language setting is picked up, and of course the Assist responds in Dutch, so that part should be good to go

Regarding Piper/Wyoming: why do I need speech to text to text to speech, if only commanding the Voice PE to control my exposed entities?

Those are only required if one needs to get more into reponse variables and templated base on that, or use other advanced options?

Right now, I just want to test it controlling and reporting about my exposed entities.

[AshaiRey]

@Mariusthvdb

Language setting is picked up, and of course the Assist responds in Dutch, so that part should be good to go

No it isn’t
Please understand that this has little to nothing to do with the language assist listen too. For example, on a system with english locale you can set the voice assist pipeline to norsk and it will understand norsk and not english (or badly)
Do not assume things and talk bad about it without reading the guides about a proper setup.

[Nick4 (Nick)]

Look

• have you checked “Home Assistant Cloud” as default
• if you open up that ‘pipeline’ you see the detailed settings, are those correct
• in the previous step, with the overview, there are the 3 dots where you can select ‘Debug’.
  Here it shows what is captured/interpreted as text

Also, in check the settings under ‘TTS’.

If some commands don’t work, try with Assist TTS the same command if the system even is able to perform that command.

[Mariusthvdb (Marius)]

Yes all of those are in Dutch.
Dont assume people dont understand…

@AshaiRey

Do not assume things and talk bad about it without reading the guides about a proper setup.

Not ‘talking bad’.

and I did read, and more importantly, follow the correct setup procedure.

It’s just that there is no real Ai there, and the commands need to fit the few pre-defined sentences.

Sentences that deviate from those will fail.

This boils down to requiring the user to fit the programming…
that is not what I had hoped for, nor expect from a serious Assistant.

Compared to the already failing GH’s, that is even worse, and nothing more than an experiment at best.
Of course that fits the name Preview. So no hard feelings

A cool one, and Id love to contribute to that experiment. hence the buy.

Otoh, there is only so much time in ones schedule, and given the other priorities I have with HA, those wont allow me to contribute in this field a lot.

So the PE is in the mail right now.

thanks for having a look at my post and taking the time to reply, I appreciate that a lot.

[AshaiRey]

Thank you for your fair answer and I am sad you are disappointed but that can happen too.
Any user review is welcome.

For readers that following this thread further this HA voice assistant box has nothing at all to do with AI. It is just plain voice recognision. If you need a device acting more into the field of AI youn need tech like a LLM hosted locally or as a service from the cloud. However that is in a complete different ballpark

[NathanCu (Nathan Curtis)]

@AshaiRey

HA voice assistant box has nothing at all to do with AI.

Truth. It is a device only backed by a pipeline. The default pipe is deterministic and supports predetermined actions through scripts that’s ALL. They just happen to have delivered basic voice intents for some very basic stuff. As an English speaker that list is woefully underpowered! Not devs fault, takes a lot of work…) but then they all need translation… Making it even harder on non English speakers.

I totally get this gripe (and honestly why I don’t even run local first on my assist setup.) It’s llm only now. Anything less feels too limiting now. But…

@AshaiRey

that is in a complete different ballpark

Even more true! Yep if you’re doing anything beyond dabbling with LLM that means paid models or local hosting, probably both. Get ready for a significant hardware purchase and hours sucked down the productivity drain learning… But then you’re in for a different type of rabbit hole to chase.

Yeah Marius it’s far… Far away from plug and play i remind people it’sa HA Voice PE not HA Voice every day. The PE, while better every day is nowhere near just drop one in and go.

…yet.itll get there.

[Mariusthvdb (Marius)]

suppose the marketing for the device could be toned down a bit…
I mean, it’s all about setting expectations, and your assessment above should be in bold.

currently, I does nothing more than my M5 Echo…, (well its sound is a tad better, but even that falls short to the Google mini)

That is why my response to the lines in the Music Assistant headlines was one of surprise also

ofc the evangelists ( ) reply that I could add an externally amped speaker to the PE…
just like we can add external LLM crunchers.

(tbh, Ive never understood why we would use the HA instance to control our music library.
Most of us already have dedicated hardware for that, which is way more adequate, and, more importantly, doesnt weigh on the HA system.)

Sure, we could do all that, but then why buy the PE in the first place.

I wont re-iterate nor proceed, and would love to close my topic here, as I don’t wish to let my disappointment affect others.

[NathanCu (Nathan Curtis)]

I for one have fired all of my other devices for that. And I absolutely DO use voice and it puts ZERO load on my ha box. Voice is processed off box.

Friday send yacht rock to the living room. Ok now move that to the hot tub deck and toss in some Jimmy Buffet for the afternoon. Don’t have to find a phone or a dashboard. I holler it as I walk out the back door.

Totally kicks the living snot out of Alexa or GH or any of them really. BUT. I also have the added benefits of working on the setup for nearly 20 months straight now and beating the living hell out of the voice pipeline to the point of breakage to find out the weaknesses.

I know you’re a huge dash guy it’s just a different UX one that for me took over a dashboard a year ago.

Theres a few things going on at once.

The device. Yep underpowered speakers. There’s 3.5 and a project to get TOSLIKK off the device hook it to something else if sound quality is your issue. After a ground loop isolator it sounds really good hooked to a Denon AVR.

Speech to phrase… Is ok I guess. But very limited if you don’t have good translation (this is your issue I suspect, also the MA music control scripts (Fes did them) are AWESOME but… In English.

LLM overcomes most of that. If it’s setup (no just exposing the ents is not enough it needs a curated prompt.) Still needs good scripts (funny enough the llm can translate and use a script form another language in most cases) but will ALWAYS be more work than most people are willing to do because… Why pay for an infrastructure and all that the time to customize kit when you don’t have to pay GH or Amz??? (hint: you’re the product, I’m tired of being product)

In 2025 people have been trained oh bring it home turn it on and magic. Sorry the vpe isn’t that yet. It’s currently more like, bring home a pet and feed it and care for it and put in hard work and…

For me well worth it to kick AMZ out.

[Mariusthvdb (Marius)]

@NathanCu

Friday send yacht rock to the living room. Ok now move that to the hot tub deck and toss in some Jimmy Buffet for the afternoon. Don’t have to find a phone or a dashboard. I holler it as I walk out the back door.

sweet!

@NathanCu

For me well worth it to kick AMZ out.

Haha, just to be clear, when I mentioned dedicated hardware/software for a music library, I wasn’t talking about something like a Spotify playlist running on a GH.

And I also wasn’t referring to programming-style setups like the ones you do on Fridays and that sort of thing.

I totally see the appeal there, though — those ideas actually sound pretty cool.

That said, my music playback needs sit a bit above that. I’m using a dedicated DAC, and I generally avoid Toslink optical — except for the TV, which doesn’t have any other digital output. For proper listening, I’m running AES/EBU balanced I/O, or BNC when it really matters.

All of that’s kind of beside the point now, but hopefully it explains my reservations a bit better…

[NathanCu (Nathan Curtis)]

Totally I get you. You’re an audiophile.

Theres still ways but yeah you’re controlling gear that will take a LOT of custom work rn because the scripts probably need I’d say a year-ish of work? You need translations.

In your case I’d probably look at it this way. For now just work on controlling your gear (do it with dashboards) think of the llm like a human using those tools. Build scripts that act like buttons and knobs and fill in the blanks. The same scripts and automations you could trigger from a button, that do the things your dashboards do. Start small with primitives then build them into bigger ones. And suddenly the voice agent has a critical mass of tools and does weird stuff. And get someone to translate the ma scripts… They’re that good.

Over 2026 we won’t see GPU prices come down quickly due to ram pricing so you’ll have another year to go before that sa-weet LLM rig to drive it becomes reasonably priced… You have time.

[AshaiRey]

Fyi, the intents are already in Dutch. Basic stuff all work but you need to have your setup in order because they make use of labels and other tags. Like, Turn (on|off) the light in {area}

[Mariusthvdb (Marius)]

Believe me, (or not…) my setup is in order…and those Dutch intents are as Nathan said.