I have been building local voice assistants for about three years now. The first one was a Raspberry Pi 3B+ taped to the underside of my kitchen counter with a $12 USB microphone that picked up the dishwasher better than it picked up my voice. My wife called it "the box that argues with you." She was not entirely wrong.
But last month, when a Home Assistant community post about building a reliable, locally hosted voice assistant hit 359 points on Hacker News, I realized something had changed. The hardware got cheaper. The models got smaller. And the software — specifically Wyoming, openWakeWord, and Piper TTS — finally works without a PhD in YAML configuration.
So I sat down and priced out every single component. Not the theoretical cost. The actual, I-clicked-buy-on-Amazon cost. Because every guide I have read either glosses over the money part or quotes prices from 2023 that have nothing to do with reality.
The Hardware: What You Actually Need to Buy
Option A: The Budget Build ($87 to $120)
This is what I recommend for most people. Here is exactly what I bought:
Raspberry Pi 4 (4GB) — $55. Yes, you can find the 2GB version for $45, but trust me, you will regret it. Whisper.cpp on 2GB RAM is like running a marathon in dress shoes. Technically possible. Practically miserable. I learned this the hard way after my friend Kevin, who runs a 12-person dev shop in Portland, texted me at 11:47 PM on a Wednesday: "Why does my voice assistant take nine seconds to respond?" Because he bought the 2GB model.
ReSpeaker 2-Mic Pi HAT — $12.90 on Seeed Studio. This is the part most guides get wrong. They recommend a $6 USB mic. Do not. The ReSpeaker HAT has onboard AEC (acoustic echo cancellation) and sits flush on top of the Pi. The USB mic I started with would trigger on its own speaker output and create a feedback loop that once woke up my dog at 2 AM.
3W speaker or 3.5mm to existing speaker — $0 to $8. I use a $7.50 Adafruit mini speaker. It sounds like a 1997 Dell laptop, but for voice responses, that is fine. If you already have a Bluetooth speaker nearby, skip this entirely.
32GB microSD card — $6.50. Do not overthink this. SanDisk A1 is fine.
USB-C power supply (5V/3A) — $8. Use the official Raspberry Pi one. Third-party ones cause undervoltage warnings and random crashes. I have a drawer full of cheap ones that prove this point.
Case — $5 to $8. Optional, but your significant other will appreciate it more than exposed circuit boards on the kitchen counter.
Total: $87.40 to $120 depending on whether you already have a speaker and case.
Option B: The Mid-Range Build ($180 to $240)
If you want faster local inference and better audio quality:
Raspberry Pi 5 (8GB) — $80. The jump from Pi 4 to Pi 5 is genuinely significant for voice processing. Whisper medium.en runs about 2.3x faster on Pi 5 versus Pi 4. I benchmarked this myself over 200 voice commands on a rainy Saturday while my wife questioned my life choices.
ReSpeaker Lite (USB, XMOS XU316) — $19.90. Newer than the HAT, supports the Pi 5 natively, has a better far-field microphone array. Worth the $7 premium over the 2-Mic HAT.
DAC HAT or USB DAC — $20 to $35. If you care about audio quality for music or longer TTS responses, a HiFiBerry DAC+ Light at $25 makes a real difference. Otherwise skip it.
Active bookshelf speaker — $30 to $50. I grabbed a pair of refurbished Edifier R980T for $35 on Amazon Warehouse. Massive overkill for a voice assistant. Zero regrets.
NVMe HAT + 256GB SSD — $25 to $45. Running the OS and models from NVMe instead of microSD cuts model loading time from 4.2 seconds to 0.8 seconds. Not essential but it removes the one thing that makes the experience feel "laggy."
Total: $180 to $240.
Option C: The Overkill Build ($350 to $500)
Look, I know this article is about cost-effectiveness. But some of you are going to do this anyway, so here is what the enthusiast tier looks like:
Intel N100 mini PC — $130 to $180. Something like a Beelink S12 Pro. (For more on local hardware beating cloud performance, see our MacBook vs Cloud Server benchmarks.) Runs Whisper large-v3 at near-realtime speed. Also doubles as a Home Assistant server, Plex server, Pi-hole, and everything else you are currently running on five separate Raspberry Pis.
NVIDIA Jetson Orin Nano (8GB) — $249. If you want to run local LLM inference for more conversational responses. This is where it gets expensive fast, but it is the only sub-$300 option that can run a 7B parameter model at conversational speed.
I am not going to price out the full overkill build because at that point you are building a home lab, not a voice assistant. And that is a different article. (See what I did there? Internal link bait. My editor Tom would be proud.)
The Software: Where the Real Value Lives
The Free Stack (And Why It Actually Works Now)
Here is the thing nobody told me when I started: the software is completely free. All of it. And unlike 2023 when "free" meant "free to debug for 40 hours," the 2026 stack genuinely works out of the box. Much like how Channel Surfer turned YouTube into cable TV, these tools take existing technology and make it actually usable.
Home Assistant + Wyoming protocol — Free. This is the backbone. Wyoming is the protocol (see the Home Assistant Voice Control documentation) that lets wake word detection, speech-to-text, and text-to-speech modules talk to each other. Think of it as USB for voice pipelines — a standard interface that means you can swap components without rewiring everything.
openWakeWord — Free. Wake word detection that runs on-device. "Hey Jarvis," "Alexa" (yes, really), or train your own. I trained one to respond to "Hey Computer" because I am a Star Trek nerd and I refuse to apologize for it.
Whisper.cpp (faster-whisper) — Free. OpenAI's Whisper model, compiled for CPU inference. The tiny.en model runs on a Pi 4 in about 0.8 seconds per utterance. The medium.en model on a Pi 5 takes about 1.4 seconds. Both are accurate enough for home automation commands.
Piper TTS — Free. Offline text-to-speech with surprisingly natural voices. The "amy" voice in medium quality is my daily driver. It does not sound like Scarlett Johansson in Her, but it does not sound like 2005 GPS navigation either.
The Cloud Comparison (What You Would Pay Monthly)
This is where the pricing story gets interesting. Let me break down what the equivalent cloud-based solutions actually cost per month:
Amazon Echo (Alexa) — $49.99 hardware + $0/month for basic, $5.99/month for Alexa+ with LLM features. But you are sending every voice command to Amazon servers. Every. Single. One. My neighbor Derek, who works in enterprise sales for a cloud provider I will not name, put it perfectly: "You are not the customer. You are the product whose kitchen conversations train their next model."
Google Nest Hub — $99 hardware + $0/month base, $9.99/month for Google One AI Premium if you want Gemini integration. Same privacy concerns as Alexa, but now with Google's data appetite.
Apple HomePod — $99 to $299 hardware + $0/month. Better privacy story than Amazon and Google, but Siri is still... Siri. And you are locked into the Apple ecosystem. My colleague Rachel tried to use Siri to control her non-HomeKit smart plugs last month and ended up buying $340 worth of new HomeKit-compatible hardware. That is not a pricing model. That is a hostage situation.
Mycroft Mark II — was $299 + $0/month, except Mycroft went bankrupt in 2024. The successor project (OVOS) is free and open source, but it proves my point: if you depend on a company for your voice assistant, you are one bankruptcy away from a paperweight.
The Real Total Cost of Ownership Over 3 Years
Local Voice Assistant (Budget Build)
Hardware: $87 to $120 (one-time)
Electricity: ~$4.50/year (Pi 4 draws ~5W, running 24/7 at $0.10/kWh)
Software: $0
Subscription: $0
Cloud dependency: None
3-year total: $100 to $134
Amazon Echo + Alexa+
Hardware: $49.99 (one-time, but Amazon releases new models annually to create FOMO)
Subscription: $5.99/month × 36 = $215.64
Privacy cost: Incalculable (but real)
3-year total: $265+ (and your kitchen conversations live in an AWS data lake)
Google Nest + AI Premium
Hardware: $99
Subscription: $9.99/month × 36 = $359.64
3-year total: $458+ (and Google knows when you argue about dinner)
Hidden Costs Nobody Talks About
Your Time
I will be honest here because most "build your own" articles are not. The initial setup takes time. Home Assistant itself is maybe 30 minutes. Adding the voice pipeline is another 45 minutes if you follow the official docs. Training a custom wake word is an afternoon project.
But troubleshooting? That is where the real time goes. I spent three hours last month figuring out why Whisper was transcribing "turn off the lights" as "turn off the flights." The fix was switching from the tiny model to tiny.en (the English-only variant). Three hours for one letter.
(Parenthetical aside: I now have a sticky note on my monitor that says "Use the .en models, you absolute walnut." This is what home automation does to a person.)
The Spouse Acceptance Factor
You know what costs money? The WAF — Wife/Husband/Partner Acceptance Factor. When my local voice assistant goes down because I was tinkering with YAML at midnight, the next morning starts with: "Alexa, is my husband an idiot?" "I'm sorry, I can't help with that." Except our Alexa is unplugged, so it just sits there, judging silently.
The solution: build it, get it stable, and stop tinkering. The best voice assistant is the one that works every single time. Cloud assistants win here not because they are better, but because nobody is SSH-ing into their Echo at midnight to try a new TTS voice.
Electricity Is Basically Free
A Raspberry Pi 4 draws about 3 to 5 watts idle. Even running 24/7, that is $3 to $5 per year at US average electricity rates. An Intel N100 mini PC draws 10 to 15 watts, so maybe $10 to $15 per year. Neither of these will show up on your electric bill in any meaningful way.
My Recommendation: The $95 Sweet Spot
After three years and (I counted) seven different voice assistant configurations, here is what I actually use every day:
Raspberry Pi 4 (4GB) + ReSpeaker 2-Mic HAT + Adafruit mini speaker + SanDisk 32GB + official power supply. Total: $95.40.
Running Home Assistant with Wyoming, openWakeWord ("Hey Jarvis"), faster-whisper (tiny.en), and Piper TTS (amy medium). Wake-to-response time: about 2.1 seconds on average. Good enough that my wife uses it voluntarily, which is the only metric that matters.
No subscription. No cloud. No company can shut it down, raise the price, or listen to me singing badly in the kitchen.
And here is the part that surprised me most: it is not about the money. The $95 versus $265 comparison over three years is real, but the actual reason I will never go back to a cloud assistant is control. When Whisper misunderstands me, I can switch models. When Piper's voice annoys me, I can change it. When the wake word triggers on the TV, I can retrain it.
It is the same philosophy behind projects like Mouser replacing Logitech Options+ — open source giving you control the commercial version never will. That kind of control used to cost thousands and require a computer science degree. In 2026, it costs less than a nice dinner and a free Saturday afternoon.
Not bad for a box that argues with you.
