For years, the deal with a smart speaker was simple: say "turn off the kitchen light" and the kitchen light turned off. Then Google replaced Google Assistant with Gemini, Amazon replaced Alexa with Alexa+, and a lot of homes that had worked fine for years started arguing with their owners.
The complaints repeat from house to house: a "goodnight" routine that no longer runs, a command that lands on the wrong device or in the wrong room, a speaker that answers a question nobody asked, an alarm that will not cancel, a reply that takes ten seconds. Owners who built a whole home around one ecosystem, with a dozen cameras and a couple of thermostats, are openly asking whether it is time to leave.
The usual answer from the smart home community is Home Assistant and its own voice assistant, Assist. This article explains what went wrong with the big assistants, how a Home Assistant voice setup really works, why so many people try it and give up, and which setup we would recommend if you want to move away from Alexa or Google without making your family miserable.
What Went Wrong With Gemini for Home and Alexa+
Both companies swapped a predictable, rule-based assistant for a large language model (LLM). The new assistants hold a conversation far better, and they are worse at the one thing a smart home needs most: doing the same thing every time.
Gemini for Home
- Broken routines. After the upgrade, voice-triggered routines stopped responding to their phrase for some owners, and older ones returned errors such as "that device does not exist". 9to5Google's own "Goodnight" routine stopped working because Gemini answered the phrase instead of running it. Duplicating the routine fixed it for some people, not for everyone.
- The wrong device, the wrong room. Owners report alarms scheduled in the wrong room, items added to the wrong shopping list, the wrong playlist, and commands that have to be repeated several times before they work.
- Ignored commands. In an Android Authority reader poll of more than 3,000 votes, close to half described Google Home as unreliable and 13% as completely broken. One owner summed up the mood: they want lights that turn on and off, not a gimmick. Others on Reddit describe a "leaving" routine that needs three tries, a speaker that claims it has no data from a Nest thermostat, and a TV volume pushed to the maximum.
- Language and countries. Spanish support only arrived in March 2026, and early access outside the US is still by invitation, country by country. Owners who speak a regional variety of a language often find the assistant understands them worse than before.
Alexa+
- Slow answers. Reviewers measured smart home commands and simple questions taking up to 15 seconds.
- Devices it cannot control. The Verge's smart home reviewer could not get Alexa+ to run her bathroom fan, and it insisted it could not make coffee while connected to a smart coffee maker.
- Alarms. On the Hard Fork podcast, Alexa+ was described as having lost the ability to reliably set and cancel alarms.
- Owner reports. In Reddit's Echo community, owners also describe alarm volume dropping by itself, lights turning on with nobody asking, and a conversation started on one Echo that the next Echo in the house knows nothing about.
The Verge put the underlying problem in one line: LLMs are not designed to be predictable, and predictability is exactly what you want when you control your home.
Why Home Assistant's Voice Assistant Is Different
Home Assistant does not try to be a better Alexa. It gives you the parts of a voice assistant and lets you choose where each one runs: in your house, in a private cloud you pay for, or in an AI provider you pick. Nothing changes overnight because a company decided to ship a new model.
It also does not tie you to one speaker. Any device that can listen and answer can be an Assist satellite, from the project's own speaker to a board you put together yourself (the hardware options are further down). What decides whether it works well is how the steps below are set up.
The Four Steps Every Voice Command Goes Through
Every voice assistant, Alexa and Google included, does four things in a row. If any one of them fails, the whole command fails, and the speaker gets the blame.
- Wake word. The device listens all the time but does nothing until it hears its name. Only then does it send audio anywhere.
- Speech to text. Your voice becomes text.
- Understanding. A "conversation agent" works out what you want from that text and does it.
- Text to speech. The assistant answers you out loud.
In Home Assistant, each step has local and cloud options, and the wrong combination is what makes people give up.
Step 1: The Wake Word
The official wake words are trained on English pronunciation. They work with other accents, but with background noise and modest microphones they can miss. If the speaker does not light up when you talk, the problem is not your setup further down; it is the microphone and the wake word.
Step 2: Speech to Text, Where Most Setups Break
There are three ways to turn your voice into text, and most people only know one of them:
- Whisper (local). It tries to transcribe anything you say. Home Assistant's own docs say it takes around 8 seconds per command on a Raspberry Pi 4 and recommend it only if you have powerful hardware at home. With the small models a Raspberry Pi can run, and in languages other than English, it mishears names constantly. In the video above, a lamp named in Spanish came back as gibberish on the first try and the second.
- Speech-to-Phrase (local). A newer engine from the Home Assistant team that only recognizes the commands it already knows and the names of the devices you expose. Because the list is closed, it is extremely fast, under a second even on a Home Assistant Green or a Raspberry Pi 4, and it gets your device names right. The catch: it only understands that subset of commands. "Turn on the desk lamp" works; "make the desk lamp a bit warmer" does not.
- Home Assistant Cloud. The paid subscription runs speech recognition on a private Microsoft Azure setup, and Home Assistant says every supported language works well with it. Only the audio goes to the cloud: the control of your home still happens locally.
Step 3: Understanding What You Asked
Out of the box, Assist understands you by matching your words against known sentences: "turn on", "turn off", "set to", followed by the name of a device, an area or a floor. It is fully local and answers in milliseconds, but it is not a mind reader.
- Names matter. If the light in your living room is called "Bulb 2 Xiaomi", saying "turn on the living room light" finds nothing. Home Assistant's guide recommends a predictable pattern (area, then a description, then the device type) and warns that a vendor name like "Tuya Light Controller 0E54B1" is hard to say and hard to match.
- Aliases fix the rest. Add the names your family actually uses as aliases, on devices, areas and floors. This is the single biggest improvement most homes can make.
- Expose only what you use. Every entity is opt-in. Fewer exposed devices means fewer wrong matches and, with an AI agent, cheaper requests.
- Phrasing is fixed. In the video, "cancel the timers" worked and "turn off all the timers" did not, because only the first was a known sentence.
Step 4: The Answer
Piper, the local voice, is fast and runs on a Raspberry Pi, but it sounds robotic. Home Assistant Cloud and other providers sound far more natural. Since 2025 both can start speaking as soon as the first words of the reply are ready: in Home Assistant's own benchmark, the time before the assistant starts talking dropped from over 5 seconds to about half a second.
Add an AI, Without Giving Up Control
The fixed-sentence agent is fast and reliable, and it cannot handle "sorry, I made a mistake" or "my head hurts, adjust the light". That is where an AI agent comes in, and Home Assistant lets you connect one as the conversation agent: OpenAI, Google, Anthropic, or a model running on your own hardware.
The setting that makes it work is the local-first fallback. With it on, Assist handles every command it understands by itself, instantly and for free, and only sends the rest to the AI. In the video, "turn off the right bulb" was handled locally, while "my head hurts, adjust the right bulb" went to the AI, which set it to white at 50%, and "I changed my mind, go back to the previous color" worked too. Combined requests ("make one lamp pink and the other green") are split into commands Home Assistant understands.
There is an important difference from Alexa+ and Gemini, and it is about who sets the rules:
- You choose the provider, and you can change it tomorrow.
- You choose what it can see. The AI can only act on the entities you expose to Assist.
- You write its instructions: how it should answer, in which language, with how much detail.
- You can give each speaker a different assistant, or skip the AI entirely in some rooms.
Running a cloud model for home control is cheap. The author of the video estimates it at cents a month, not euros, with the local-first fallback on. OpenAI still recommends watching your usage and setting limits, which is good advice for any paid provider.
A Fully Local AI
If no request should leave your house, you can run the model yourself. Home Assistant says recent local models that use 8 GB of video memory or more are nearly keeping up with cloud models. In the video, a small open model running on an AMD Ryzen AI Max machine understood "the right bulb is too cold, my head hurts, I want something warmer" and acted on it, a little slower than the cloud but perfectly usable.
That takes real hardware: a mini PC with a powerful integrated GPU and a lot of shared memory, or a machine with a dedicated graphics card. The same box can run Home Assistant itself and Whisper with a larger model.
MS-S1 Max (64GB + 2TB)
4.8(13)$2599.00
A9 Mega (64GB + 1TB)
4.3(245)$2399.00
Which Speaker to Use
The engines decide how well Assist understands you; the speaker decides whether it hears you at all. Today there are two good starting points.
Home Assistant Voice: The Official One
The Home Assistant Voice (sold as the Voice Preview Edition) is the project's own speaker and the easiest way in. It has two microphones with an XMOS audio chip for echo cancellation and noise removal, a hardware switch that physically cuts power to the microphones, and a 3.5 mm output for a proper speaker, because the built-in one is meant for voice replies, not music. It answers to "Okay Nabu", "Hey Jarvis" or "Hey Mycroft", detected on the device itself.
Nabu Casa calls it a preview on purpose: in its own words, there is still more to do before it is ready for every home in every country. Its weak point is the one Alexa and Google spent a decade solving: hearing you across a noisy room. Two microphones struggle with the TV on, music playing or a family dinner, and the built-in speaker is small.
ReSpeaker XVF3800: Better Ears, More Work
New boards are starting to close that gap. Seeed Studio's ReSpeaker XVF3800 uses four microphones in a circle and a dedicated audio chip with beamforming (it focuses on the direction your voice comes from), echo cancellation, AI-based noise suppression, de-reverberation and automatic gain, rated for pickup up to about 5 meters. The version with a XIAO ESP32-S3 board joins Home Assistant over Wi-Fi through ESPHome, and it now comes in a ready-made case, so there is nothing to 3D print. No speaker comes in the box: you plug one into its 3.5 mm jack or its connector for a 5 W speaker. You still flash a firmware and an ESPHome configuration that Seeed publishes, so it remains a project for tinkerers, but it is the most promising way today to give Home Assistant ears that work across a room.
What About Old Echo Devices?
There are community projects that unlock older Echo devices. The bootloaders of the first Echo Show 5, the first Echo Spot and the second-generation Echo Dot have been unlocked, and the models with a screen can run LineageOS, an open version of Android, which turns them into small Home Assistant dashboards. It only applies to old models, the process is long, and the builds are not production quality. Fun for a weekend; not a plan for the whole house.
Music Assistant
Music is often the last reason a home keeps its Echo or Nest speakers. Music Assistant is a free, open-source system from the Open Home Foundation that puts local files and streaming services (Spotify, Apple Music, YouTube Music, TIDAL, Qobuz, Deezer and more) in one library and plays it on the speakers you already own: Sonos, AirPlay, Google Cast, Home Assistant players and even Alexa devices.
Since Home Assistant 2025.6, asking Assist to play something on a speaker or in a room works out of the box. Anything beyond that (shuffle, repeat, stop, grouping, naming an album's artist) still needs community blueprints, custom sentences or an AI agent, and Music Assistant's own docs say Home Assistant does not plan to add more media actions. It works, and it takes some setup.
Three Ways to Set It Up
- Basic local. Speech-to-Phrase, the built-in agent and Piper, on a Home Assistant Green or any small box. Everything stays at home and it is very fast, as long as you stick to clear commands and well-named devices.
- Powerful local. Whisper with a larger model and a local AI, on a mini PC with a strong GPU. Fully private and conversational, at the cost of the hardware.
- Hybrid (our recommendation for most homes). Home Assistant Cloud for speech to text and text to speech, an AI agent with the local-first fallback on, and Home Assistant controlling everything locally. The simple commands never leave your network; only the ones the built-in agent cannot understand reach the AI you chose.
Home Assistant Green
4.4(916)
Mini S12 Pro (N100 16GB + 500GB)
4.5(659)$339.00SER8 (24GB + 1TB)
4.3(176)$799.00
Can You Live Without Alexa Today?
For most people, honestly, not yet. If you want to plug something in and forget about it, Home Assistant Voice is not plug and play: it takes time to name your devices, pick the right engines and test. And its microphones still do not match a commercial speaker in a noisy room.
If you enjoy tuning your home, it can already do more than Alexa or Google: an AI you choose, with instructions you write, acting only on what you allow, and a basic local mode that keeps working with the internet down. The price of that control is your time.
A few tips that make the move easier:
- Do not force it on the family. Leave the Echo or Nest where people like it, add a Home Assistant speaker next to it, and let the better experience win them over.
- Fix names before blaming the microphone. Most "it does not understand me" problems are a device name or a missing alias.
- Start with Speech-to-Phrase or Home Assistant Cloud, not Whisper on a Raspberry Pi.
- Turn on the local-first fallback when you add an AI, so simple commands stay instant and free.
- Keep a physical switch or a button for the essentials. Whatever assistant you use, a light should never depend on one.
Conclusion
Alexa+ and Gemini for Home made the big assistants better at conversation and worse at the job most homes bought them for. Home Assistant Voice flips the trade-off: it is less polished and needs more of your time, but every piece is yours to choose, and nothing breaks because a company changed its model overnight.
Sources
- ¿Podemos VIVIR SIN ALEXA en 2026? Home Assistant Voice a prueba (Un loco y su tecnología, YouTube)
- Gemini for Home reportedly breaks automations (9to5Google)
- Google Home continued problems (Android Authority)
- Think twice about upgrading to Gemini for Home (TechRadar)
- Google Gemini for Home adds Spanish support and launches in Mexico (Thurrott)
- Gemini for Google Home launches in 16 new countries (Tech Advisor)
- Amazon's AI Alexa is a half-working mess (Futurism, quoting The Verge and Hard Fork)
- Home Assistant Voice Preview Edition (Home Assistant)
- Voice control (Home Assistant)
- Installing a local Assist pipeline (Home Assistant)
- Best practices with Assist (Home Assistant)
- Speech-to-Phrase brings voice home: Voice chapter 9 (Home Assistant)
- Voice chapter 10 (Home Assistant)
- Building the AI-powered local smart home (Home Assistant)
- OpenAI integration (Home Assistant)
- Getting started with ReSpeaker XVF3800 (Seeed Studio)
- Smart Home Voice Control with Home Assistant with reSpeaker XVF3800 (Seeed Studio)
- Older Echo Show 5, Echo Spot and Echo Dot rooted with LineageOS ROMs (AFTVnews)
- Music Assistant
- Music Assistant voice control (Music Assistant)
