SQUELCH

Robot Radio

A walkie-talkie a three-year-old can use to ask a robot anything.

researchupdated June 22, 2026with Alasdair, 3

The problem

Alasdair loves asking "the robot" questions. I'm not handing a three-year-old my phone or a tablet — too fragile, too much rabbit-hole, too many ways for it to go sideways. He needs a durable, single-purpose, screen-free device that does exactly one thing: push the button, ask, hear an answer back.

For whom: Alasdair (3) — and whatever guardrails keep me comfortable with a small kid talking to a large model.

Constraints

  • One button. Push to talk. No screen, no menus, no setup.
  • Rugged enough to be dropped, thrown, and left in the yard.
  • No general-purpose phone, tablet, or open internet in his hands.
  • Push-to-talk only — no always-listening microphone living in the house.
  • Answers spoken back: short, calm, and age-appropriate.
  • Parent in the loop: content guardrails, a reviewable transcript, time limits, and an off switch on the base.
  • The handheld is dumb; the base station does all the thinking.

What the research found

A real over-the-air walkie-talkie is the wrong tool — legally
An automated base station that generates spoken replies is a prohibited automatic / unattended station on FRS (47 CFR §95.347, with no carve-out) and on GMRS (auto-control is allowed only for repeaters, §95.1747 — an AI that originates its own replies is not a repeater). Ham radio is the only over-the-air path, and only for data-mode interrogate-and-respond, not toddler voice. The "serious walkie-talkies" premise is the one to avoid.
Wi-Fi / Part 15 sidesteps all of it
Unlicensed Part 15 devices have no control-operator rule and no automatic-operation ban — those live only in the licensed services. Make the handset a Wi-Fi push-to-talk device and keep the AI on the wired side: it never creates a regulated station, feels exactly like a walkie-talkie to a three-year-old, and is cheaper and lower-latency than a radio bridge. The residual real concerns are child privacy (COPPA) and product safety, not spectrum.
Anthropic is the reasoning step only
Claude's API is text in, text out — speech-to-text and text-to-speech are separate pieces on either side. Use Claude Haiku 4.5 for the kid Q&A (fast and cheap), stream its reply sentence-by-sentence into the voice so the first words play in about a second, and log every exchange.
Understanding a three-year-old is the hard part
A toddler sits below the floor of every English child-speech dataset — roughly a 40% word error rate on spontaneous speech. This, not the hardware or the model, is the real risk. Mitigate with push-to-talk (never open voice-detection), a few expected topics, and a Claude prompt told to read garbled input charitably — and prove it for ~$0 before buying anything.

Approach

The cleanest build keeps the "radio" off the regulated airwaves entirely. He carries a rugged push-to-talk handset that talks to the house over Wi-Fi — not a licensed channel — and the thinking happens on a small always-on machine (a Raspberry Pi). Hold the button, ask; the Pi turns speech to text, sends it to Claude with a system prompt written for a three-year-old, turns the reply into a warm voice, and plays it back. To him it's just a walkie-talkie that answers. A true two-way-radio version is legal only on licensed ham as a data mode — not the toddler-voice use case — so Wi-Fi is both the lawful path and the better one.

How it works

  1. 1

    Push to talk

    He holds the button on the handset and asks. The button — not voice detection — defines when he is talking.

  2. 2

    Capture

    The handset streams his audio to the Pi over Wi-Fi while the button is held.

  3. 3

    Speech → text

    whisper.cpp on the Pi (or a cloud transcriber for accuracy) turns it to text — forgiving of three-year-old diction.

  4. 4

    Claude

    Claude Haiku 4.5 with a kid-safety system prompt; short answers, streamed sentence by sentence.

  5. 5

    Text → speech

    A warm voice (local Piper, or a cloud voice for extra warmth) reads each sentence as it arrives.

  6. 6

    Play back

    Audio streams back to the handset and plays — first words in about a second.

  7. 7

    Log it

    Every question and answer is saved so I can review what he asked and how it answered.

Parts we're considering

  • Handset (his)

    An ESP32-S3 board with a microphone, a small speaker/amp, and one big push-to-talk button — streams audio over Wi-Fi.

    Reference design: atomic14/esp32-walkie-talkie. ~$20-40 for two.

  • Bridge machine

    Raspberry Pi 5 (8GB) — always on, runs the whole pipeline on CPU.

    ~$80. No GPU or AI-HAT needed for a voice-only pipeline.

  • Speech-to-text

    whisper.cpp base.en locally, or a cloud transcriber (Deepgram) if toddler accuracy demands it.

    The slowest local link, and where the accuracy risk lives.

  • The model

    Claude Haiku 4.5 via the Anthropic API, streamed.

    Reasoning only — speech-to-text and text-to-speech are separate.

  • Text-to-speech

    Local Piper (free, private, sub-second) or a cloud voice (ElevenLabs Flash) for extra warmth.

  • Radio-bridge alternative

    If an air-link "radio feel" matters: a license-free 2.4GHz wireless intercom with the AI wired into its base port — never a regulated FRS/GMRS automated station.

    More cost and latency than Wi-Fi; only for the aesthetic.

Open questions

Will it understand him?
Toddler diction is make-or-break. Before buying anything, record 20-30 clips of Alasdair and run them through whisper.cpp and a cloud transcriber. If the error rate is unusable, narrow it to a few expected topics with a forgiving prompt — or shelve it until he is older.
Does it feel instant?
Aim for one to two seconds from letting go of the button to the first words. Stream Claude sentence-by-sentence into the voice and keep transcription fast. Prove the feel on a laptop before committing to Pi-local speech-to-text.
Kid-safety prompt + review
Short answers, simple words, refuse or redirect anything scary or off-limits, never ask for personal info, hard length cap — plus a saved transcript and time limits. Dry-run the prompt in text first.
Privacy (COPPA) and what leaves the house
A child's voice going to a cloud transcriber, model, or voice is the real privacy question now that spectrum is settled. Lean local-first (whisper.cpp + Piper) so little or nothing leaves the Pi, and decide deliberately what is allowed to.
Interruptions
He will talk over the answer. Decide whether holding the button cuts the current reply — probably yes, the simplest and most natural behavior.
Where the base lives + Wi-Fi range
An always-on Pi, and solid Wi-Fi out to wherever he actually plays — the yard included.

Next steps

  1. Record 20-30 clips of Alasdair actually talking and run them through whisper.cpp and a cloud transcriber — the cheapest way to learn whether the whole idea is viable.
  2. Build the ~$0 laptop prototype: spacebar push-to-talk → speech-to-text → Claude Haiku (streamed) → streamed voice. Tune it until it feels instant.
  3. Write and dry-run the kid-safety system prompt in plain text before any hardware.
  4. Only then buy the Pi 5 + two ESP32-S3 handsets and port the working pipeline onto them.