Does voice data leave the robot?

Yes. After the wake word, the 3-second command recording is sent to Groq for recognition, and the reply is generated there. Check Groq's terms if recordings are sensitive.

What commands does it understand, and what are the limits?

English commands for seven moves — forward, backward, strafe left or right, turn left or right, stop — plus continuous mode and speed changes. The command window is 3 seconds after the wake word. It does not navigate or avoid obstacles.

How far away can I stand?

Stand within about 1 m and say "Hey Jarvis" clearly. If it never wakes, check the container log to confirm the reSpeaker Flex was selected as the microphone.

How fast does it respond?

The first request after start can take 2–3 seconds, and later ones are faster. Response time depends on the Groq API and your network.

Can I use my own controller instead of voice?

Yes. The XIAO ESP32S3 takes one-character commands over USB serial at 115200 baud, with x as emergency stop. Any program that writes to that port can drive the chassis.

Can I change the voice, wake word, language or model?

Six voices are selectable in the device settings. The wake word, recognition language and model names are container settings. Other languages need a recognition and voice model that supports them.

What does it cost to run?

A Groq API key is required, and the free tier covers casual use.

Voice-Controlled LeKiwi Robot

Drive a three-wheeled LeKiwi robot by voice. Say "Hey Jarvis" and a command in English; the robot replies and moves forward, back, sideways or turns. Speech recognition, the language model and the voice reply run on Groq's cloud API, so the robot needs internet.

SenseCraftIntermediate30minAcoustic & Audio AI Source Code

reSpeaker Flex XVF3800 — the microphone array that hears "Hey Jarvis" on the robot

Concept render of the voice-controlled LeKiwi

Reusable Site Templates

A small mobile robot that people steer by speaking, with no remote and no app.

STEM teaching — one build covers servos, a microcontroller, voice and a language model.
Robot prototyping — a voice-driven base for delivery, inspection or companion robot ideas.
Events and demos — visitors talk to the robot and watch it move.
Hands-free control research — try voice driving for assistive use.

Not a fit: sites without internet, languages other than English, autonomous navigation, and data that must not leave the site.

How it works

Voice-controlled LeKiwi device composition: reSpeaker Flex connects over USB to a Pi 5, which calls the Groq cloud API over HTTPS, plays replies through a speaker over 3.5mm / USB, and sends moves over USB to a XIAO ESP32S3 that drives the LeKiwi chassis over the servo bus

Five things on the robot: the LeKiwi chassis with three wheel servos, a XIAO ESP32S3 that drives the servos, a Pi 5 (8GB) as the voice brain, a reSpeaker Flex microphone array and a speaker. The Pi sends each spoken command to Groq's cloud API and passes the resulting move to the XIAO over USB.

LeKiwi kit — a three-wheeled omnidirectional chassis on three STS3215 servos. It moves forward, back, sideways and turns on the spot.
XIAO ESP32S3 — the motor controller. It turns one-letter serial commands into wheel motion and has an emergency stop.

Setup is done in the deployment guide: assemble the chassis, assign servo IDs 1–3 one at a time, then flash the motor firmware. The servos need external power, not only USB.

A Pi 5 (8GB) on the robot runs one container: wake word, recording, the serial link and playback.

Step	Where it runs
Wake word	On the Pi
Speech recognition (Whisper)	Groq cloud
Picking the reply and the move (Llama 3.1 8B)	Groq cloud
Voice reply (Orpheus, English)	Groq cloud
Wheel control	XIAO ESP32S3

You need a Groq API key; the free tier is enough for casual use.

The data interface is listed under What You Get below.

How well it works

What the LeKiwi + XIAO ESP32S3 + Pi 5 + reSpeaker Flex configuration does after deployment.

What you can use

Wake word, a spoken English command, a spoken reply, then the move.
7 moves: forward, backward, strafe left and right, turn left and right, stop.
Short nudges by default; "keep going" switches to continuous motion until "stop".
"Increase speed" and "decrease speed" adjust the nudge; six voices are selectable in the device settings.
Questions unrelated to driving get a spoken answer and no motion.

Limits

Needs internet on every command. Recognition, the language model and the voice reply all call Groq.
English only by default.
The command window is a fixed 3 seconds after the wake word.
The first request after start can take 2–3 seconds to answer.

What you get out of it

The voice brain talks to the motor controller over USB serial at 115200 baud, one character per command.

Character	Move
`w` / `s`	Forward / backward
`q` / `e`	Strafe left / right
`a` / `d`	Turn left / right
`x`	Emergency stop

Lower-case letters give a short nudge; upper-case letters keep moving until stop. Any program that can write to the serial port can drive the chassis, so the voice brain can be replaced with your own controller.

Porting It to Your Own System

The reusable unit is wake word → cloud speech and language model → one move → single-character serial command.

Layer	What porting costs you
XIAO motor firmware and the serial command set	reuse as-is
Wake word and microphone selection on the Pi	reuse as-is
Move list and reply style in the system prompt	edit for new commands
Cloud recognition, model and voice	swap for a local speech service if the site is offline
Sensors and autonomy	add yourself — the design only executes spoken moves

Where this shape does not fit

Offline or privacy-bound sites — every command goes to a cloud API. See the Local Voice Service reference design for on-device speech.
Navigation — there is no map, obstacle sensing or path planning.

How to deploy

Check the setup and its product list, then follow the deployment guide.

Deployment Options

Voice Control Your LeKiwi

Assemble the robot, flash motor firmware, and deploy the voice AI brain to Raspberry Pi.

6 steps · ~30min

Open the deployment guide →

Robot Chassis

lekiwi

Motor Controller

xiao_esp32s3

Voice Brain

Pi 5 (8GB)

Microphone

respeaker_flex

FAQ

No. Speech recognition, the language model and the voice reply all run on Groq's cloud API. Only the wake word and wheel control run on the robot.

Didn't find your answer?

We Are Glad to Be Your Hardware Partner !