Skip to content

Hardware guide & bill of materials

Part Recommended Notes
Board Raspberry Pi 5, 8 GB Pi 4 (4 GB+) also works, a bit slower on the LLM
Storage 32 GB+ microSD, or NVMe via HAT models are ~1–3 GB total
Microphone USB mic, or ReSpeaker 2-Mic HAT far-field arrays improve wake-word later
Speaker USB speaker, 3.5 mm, or I²S (MAX98357A) I²S sounds best in a case
Power Official 27 W USB-C PD (Pi 5) undervoltage throttles inference
Cooling Active cooler / fan the LLM will use all cores
Case 3D-printed (optional) keep mic ports unobstructed

Choosing models by board

Board whisper LLM (Ollama)
Pi 5 (8 GB) base.en / small.en llama3.2:3b or qwen2.5:3b
Pi 5 (4 GB) base.en llama3.2:1b
Pi 4 (4 GB) tiny.en / base.en llama3.2:1b

Smaller models = faster replies. Start small, scale up if latency is acceptable.

Audio setup

  1. Plug in the mic and speaker.
  2. List devices: arecord -l and aplay -l.
  3. Note the card/device (e.g. card 1 → plughw:1,0) and set device in your config.
  4. Test capture: arecord -d 3 -f S16_LE -r 16000 -c 1 test.wav && aplay test.wav.

Case & assembly

  • Any Pi 5/4 case with an active cooler and openings for the mic/speaker works.
  • If you use an I²S DAC/mic HAT, pick a case with GPIO access.
  • Keep the microphone away from the speaker to reduce echo (or add a small baffle).

Printable case STLs and a wiring diagram are a great community contribution — PRs welcome to add them under hardware/.