TARS-Mini

A desk-sized voice assistant modeled after TARS from Interstellar. Runs on a Raspberry Pi 3B. Listens for a wake word, transcribes speech, calls Claude, speaks the reply through a Bluetooth speaker.

Persona: 90% honesty, 75% humor. Short responses. No pleasantries.

---

Architecture

Two modes — pick one:

Standalone (recommended for Pi 4/5)

Everything runs on the Pi itself.

Mic → openWakeWord → Silero VAD → faster-whisper → Claude → Piper TTS → BT speaker

File: tars_v2.py

Client/Server (recommended for Pi 3B)

The Pi is a thin audio bridge. All AI runs on a more powerful computer on the same LAN.

Pi mic → TCP → Server (Whisper + Claude + Piper) → TCP → Pi speaker

Files: tars_pi.py (Pi) + tars_server.py (server)

---

Hardware

• Raspberry Pi 3B or newer
• I2S MEMS mic or USB mic
• Bluetooth speaker (tested: Sony SRS-XB23 via BlueALSA SCO)

---

Quick Start

1. Set up the Pi

# Clone and run setup
git clone https://github.com/2016judea/tars-mini
cd tars-mini
chmod +x setup.sh && ./setup.sh

setup.sh installs system packages, creates a venv, installs Python deps, and downloads Piper.

2. Get a wake word model

Option A — Train "Hey TARS" (recommended): Use the openWakeWord Colab notebook. Copy the .tflite here and set WAKE_WORD_MODEL in tars_v2.py.

Option B — Use a pre-built placeholder:

python -c "import openwakeword; openwakeword.utils.download_models()"

Then set WAKE_WORD_MODEL = Path("hey_jarvis.tflite") temporarily.

3. Download the Piper voice model

# Already done by setup.sh, but if needed manually:
wget https://github.com/rhasspy/piper/releases/download/v0.0.2/voice-en-us-lessac-medium.tar.gz
tar -xzf voice-en-us-lessac-medium.tar.gz

4. Set your API key

export ANTHROPIC_API_KEY=sk-ant-...

5. Run

source .venv/bin/activate
python tars_v2.py

---

Client/Server Setup

On the server (Windows/Mac/Linux):

pip install -r requirements_server.txt
# Place piper binary next to tars_server.py
export ANTHROPIC_API_KEY=sk-ant-...
python tars_server.py

On the Pi:

Edit SERVER_HOST in tars_pi.py to your server's LAN IP, then:

pip install -r requirements_pi.txt
python tars_pi.py

Deploy from your computer to the Pi:

# Configure TARS_HOST if needed
export TARS_HOST=pi@192.168.0.20
./deploy.sh

---

Configuration

All constants are at the top of each script. Key ones:

Variable	Default	Description
WAKE_WORD_MODEL	hey_jarvis.tflite	Path to .tflite wake word model
WAKE_WORD_THRESHOLD	0.5	Raise to reduce false triggers
WHISPER_MODEL_SIZE	tiny	base is better if Pi has headroom
ANTHROPIC_MODEL	claude-sonnet-4-6	Claude model
MAX_TOKENS	256	Short replies = lower TTS latency
ALSA_OUTPUT_DEVICE	BlueALSA SCO	Change to default for 3.5mm jack
MIC_DEVICE_INDEX	None (system default)	Set to int for a specific device

Persona dials (_honesty, _humor) can be changed at runtime by saying "set honesty to 70" — TARS parses SETTING:X:N sentinels in its own output.

---

Test Without Hardware

export ANTHROPIC_API_KEY=sk-ant-...
python test_prompt.py

Interactive REPL — no mic, no speaker, no wake word. Good for tuning the persona.

---

Files

File	Purpose
tars_v2.py	Standalone Pi build — full pipeline
tars_server.py	Server-side AI pipeline
tars_pi.py	Pi audio bridge (client/server mode)
test_prompt.py	Local persona REPL, no hardware
setup.sh	One-time Pi setup
deploy.sh	Sync to Pi + restart
requirements.txt	Standalone Pi deps
requirements_pi.txt	Thin client Pi deps
requirements_server.txt	Server deps

---

Dependencies

• openWakeWord — wake word detection
• Silero VAD — end-of-utterance detection
• faster-whisper — ASR
• Anthropic Python SDK — LLM
• Piper — TTS (persistent process, not per-sentence respawn)
• BlueALSA — Bluetooth audio on Pi (SCO profile)