Marvin — A Voice Assistant on Custom Hardware

Marvin is a client/server voice assistant running on a purpose-built device. The client — a Raspberry Pi 3 Model B+ with a 7-microphone array, 12 addressable RGB LEDs, stereo speakers and a battery in a 3D-printed casing — detects speech, responds to a wake word, and streams audio to the server. The server runs the pipeline: speaker diarization and identification via embeddings, speech-to-text, an LLM for the response, and text-to-speech, serving multiple clients in real time.

Final project for the Artificial Intelligence for Science and Technology course, MSc in Artificial Intelligence (University of Milano-Bicocca, A.Y. 2024/2025), with Andrea Yachaya.

The 3D-printed client: hexagonal microphone array on top, Raspberry Pi and audio/battery boards stacked below. CAD renders from the project presentation.

Architecture

Data flow: the client captures audio on wake word and sends it to the server, which runs diarization, STT, LLM inference and TTS, then returns audio for playback. Source: project presentation.

• Client (Final_Project/client/): wake-word detection and voice-activity detection on-device, LED feedback sequences, audio capture and playback, socket communication with the server.
• Server (Final_Project/server/): speaker diarization + enrollment via embeddings (enroll_speaker.py), speech-to-text, LLM inference, text-to-speech (audio_processing.py, server.py); handles multiple concurrent clients.

Hardware

Labelled assembly: Raspberry Pi 3 Model B+, 3300 mAh lithium battery with charger, 5 V regulator, 2×5 W speaker amplifier, MEMS microphones and addressable RGB LEDs, ~115 mm across. Source: project presentation.

Microphones are read over I²S; the RGB LEDs are driven over I²C. Power comes from a lithium battery through a boost regulator sized for the Pi.

How to run

pip install -r Final_Project/requirements.txt
python Final_Project/server/main.py     # on the server machine
python Final_Project/client/main.py     # on the Raspberry Pi client

Wake-word training lives in Final_Project/training/; enroll a speaker with Final_Project/server/enroll_speaker.py.

Presentation

Full slides: Morello_Yachaya_Presentation.pdf — Mirko Morello, Andrea Yachaya, 2025.