Voice

What it does

Three boxes, not one

Tap the robot and talk. It works out what you said, decides what to answer, and says it out loud — and you can see all three steps happen.

A single speech-to-speech model would be faster. We keep the three apart on purpose: when the robot mishears you — and it will, hilariously — you can see exactly which step went wrong. That is the whole lesson.

It also turns its body toward whoever spoke, because the four microphones can hear direction, and it moves its head in time with its own voice.

Greek first

It really does speak Greek

Most voice assistants treat Greek as an afterthought. These parts don't: the ears score 4.2% word error rate on Greek where Whisper scores 13.8%, and they transcribe English words in English even in the middle of a Greek sentence — which is exactly how a nine-year-old actually talks.

The stock voices are English speakers who read Greek with an accent. Greek voices from the ElevenLabs Voice Library are flagged in the dropdown — or keep the accent, if that turns out to be funnier.

The microphone

Off until someone reaches out

By default the robot hears nothing at all until a hand taps it — the same boop that made it giggle in Build 02. One tap opens one listening window, which closes when you stop talking.

That is a choice, not a limitation. A child can see when the robot is listening, and they are the one who decides. There is an always-on mode in the settings; think about it before you turn it on, and then talk about it together. That conversation is the bonus part of this build.

Install & set up

Put it on your robot

1. On the robot dashboard (reachy-mini.local:8000) → Apps → Install, paste this Space's URL.
2. Start it, then open the settings page at reachy-mini.local:8042.
3. Paste your OpenAI and ElevenLabs keys, pick a voice, and write the robot's personality.
4. Boop the robot and say hello.

You'll need: a Reachy Mini, Wi-Fi with internet, an OpenAI key, and an ElevenLabs key (optional — without it the robot still talks, in an English accent). Talking costs well under a cent per exchange, and the app enforces a daily spending limit you set, so it can't run away from you.