Voice commands that work without Google or Alexa

This is for anyone who’s already a little uneasy that Siri hears a baby cry and Google listens to our mail. And who wants a real voice dialogue with an AI without handing the recording to Big Tech.

What hurts

We’ve gotten used to dictating instead of typing. It’s faster, more convenient, and sometimes the only option (behind the wheel, in the kitchen, with a child in your arms).

But three things nag:

First — the voice goes to Big Tech. When we dictate “a letter to my wife about the hospital” to Siri, that audio recording passes through Apple’s servers. When we tell Google Assistant “log a meeting with the lawyer” — it goes to a Google data center. This isn’t “they could theoretically pass it on” — it’s a fact under the terms of service we all signed without reading.

Second — our language often suffers. Most voice systems are optimized for English. Other languages — Ukrainian, Polish, Czech, other less common ones — exist, but work worse: they recognize less accurately, don’t understand accents, mangle specialized terms. English speakers notice this less; everyone else, daily.

Third — it’s only input, not a conversation partner. Siri/Google mostly “recognize voice → search Google” or “recognize voice → run a command.” It’s not a conversation with an AI that knows us. It’s an assistant that hears us but doesn’t understand us.

A home lab offers a different option: voice → text → AI → voice back, all of it on our home server, never leaving the apartment.

Where we’re heading — and what already works

Let me say it upfront: the full setup is still on my to-do list, not a daily driver. There are parts that already work, and parts I’m testing or planning. I’ll write honestly about where we are.

What already works every day:

• A Telegram bot that takes voice messages and returns text. Simple, open, written in a couple of hours (as I described in the first post of the series). I use it regularly — to quickly jot an idea, a reminder, a draft email. The voice is processed on my server, it doesn’t go to Telegram’s services.
• Access to Claude via a web interface from my phone (from the previous post). I can type questions and get answers. It’s not voice, but it’s the foundation the voice will sit on.

What I’m testing / preparing:

• A full voice dialogue with the AI (voice in → voice reply back). Technically it’s assembled, but I’m still working out the nuances: latency, the quality of the synthesized voice, integration with the phone through a convenient interface, reliability.
• A wake word (“Hey, [name]”) instead of pressing a button. That’s a separate story; for now I keep an explicit button — so the server doesn’t “listen constantly.”

What’s still ahead:

• Stability “like Siri” — no reboots, no lag, with a decent voice
• Integration into the family’s routine (not just mine)
• Documentation simple enough that my wife can use it without me explaining

If you have concrete questions about a specific part of this setup — better to take it as a separate conversation, because there are a lot of situational decisions here.

Why this isn’t Siri or Google

Two key differences — conceptual, no matter what percentage of the setup we’ve assembled:

1. The voice goes nowhere third-party. I dictate → the server listens → the server answers. Between those steps — nothing. No “processing the voice to improve quality.” No “anonymous recording to train a model.”

2. The AI knows me. It’s not “log a meeting” → “meeting logged.” It’s a full Claude with memory of my projects, my clients, my family. I can ask: “What’s the next step on the project with Oleksandr?” — and it answers concretely, because it knows who that is and what we’re doing.

That’s an entirely different level of usefulness. Not voice input — a voice conversation partner who knows us.

The quality we need to look at honestly

Speech recognition (voice → text). Modern open-source models (Whisper) work very well. English — practically perfect. Less common languages (Ukrainian, Polish, Czech, and the like) — also good, but names and surnames sometimes need correcting, and technical terminology comes through with a sprinkle of English. All of this improves from model to model.

Voice synthesis (text → voice). Here it’s still middling. The voice comes out understandable but mechanical — it’s not Siri’s bright female voice. It’s more like “Google Translate five years ago.” For functional replies — good enough. For long texts — faster to read with your eyes.

Bottom line: input works very well, output is middling. For most of us that’s enough, because we read faster than we listen anyway. If we critically need a “pleasant voice” — for now this setup won’t match Apple/Google.

Scenarios where this setup becomes useful

Let me cover where it genuinely comes in handy (for me — the Telegram variant already now, the full one as it gets ready):

1. Kitchen / morning routine. Hands full, need to add a task, set a reminder, check something. Press the mic button on the phone, dictate, get a short answer. More convenient than searching and typing.

2. Behind the wheel. Not for AI conversations about philosophy — that’s dangerous. But for quick “add a meeting,” “remind me to call tomorrow,” “save this idea” — perfect. The voice doesn’t go to Google Drive.

3. On a walk with a child. We have minutes when our hands are busy with the stroller but our head is free. We can talk through an idea, get quick feedback, save it to notes.

4. In the bath (yes, seriously). Sounds odd, but it’s the best time for thinking. We set the phone on the shelf, talk, get answers through the speaker. The voice does NOT go to Apple/Google, just to our server in the next room.

The common thread — these are scenarios where the hands are busy but the head is free. Not “replace the keyboard,” but complement it where it’s out of reach.

How to set it up (the general outline)

I won’t give concrete commands — that’s a separate guide. The general logic is this:

1. On the server we run three Docker containers: speech recognition, the “glue” between them and the AI, voice synthesis
2. Recognition is Whisper (open, free)
3. Synthesis is Piper (open, free, with voices for many languages)
4. On the phone — a browser with a bookmark to the server, or a simple web page: “press a button — speak”
5. We connect it all to our Claude or another AI already running on the server

Roughly: 1-2 evenings of setup, then a few more evenings of polishing (especially the convenient interface on the phone). Set it up once — after that it runs itself.

A detailed step-by-step guide — a separate post, once my full setup is stable and I can describe everything with concrete numbers. If you’d like to go through the setup together on your specific scenario, we can do a one-off discovery conversation.

Limitations worth knowing

What this thing doesn’t solve:

• It depends on our server. If it’s off / frozen — no voice. Unlike Siri, which “always works.” You have to set the server to restart automatically — a separate small detail
• Latency of 2-4 seconds. Between saying something and the AI answering — a pause. Siri is half a second faster. For an adult it’s OK; for a child, noticeable
• It needs Wi-Fi or Tailscale. Without a private network — it doesn’t work. On a plane without Wi-Fi it won’t work either (though Siri isn’t great there either)
• The voice output is mechanical. Not the level of Siri/Alexa. For functional commands — good enough. For long dialogues — faster to read with your eyes

All of this may improve over time — open source catches up fast. But as of today, it’s worth knowing what we’re getting into.

What’s next

Now we have the beginning of a voice dialogue with an AI that doesn’t spy. In the next post — how to make that same server serve not just us but the whole family: separate accounts for the kids and wife, shared folders for photos, movies without YouTube algorithms.

I’ll say it upfront — as with the voice, this is only half set up for me. I’ll tell you where I’m heading and what already works, without sugarcoating.

Common mistakes

1. Trying to replace Siri/Google entirely. No need. They have their niche (quick commands, music, calls). A home voice AI has a different niche (longer dialogues, our privacy, our language, our AI with memory). Use both, each for its purpose
2. Waiting for the perfect voice. For now, local synthesis is middling. What exists is enough for work, not enough for an audiobook. Don’t get blocked on this
3. Running all of this on the phone, instead of the server. The phone heats up, the battery drains, the models are large. The server just sits there, running. All the heavy lifting — on it
4. Installing Whisper directly into the system, not via Docker. In a month everything gets tangled. As I’ve written — get used to Docker from the start
5. Not trying at all, because “it’s all so technical.” Running Whisper in Docker — 5 minutes and one command. A simple Telegram bot with voice-to-text — a couple of hours. Integration with the phone — a separate evening. It’s less than setting up a new printer