[LLM] # Available models: Any DeepInfra-supported model # Examples: openai/gpt-oss-20b, meta-llama/Llama-3.3-70B-Instruct-Turbo, meta-llama/Llama-3.2-3B-Instruct model = openai/gpt-oss-20b # Gating model for intervention decisions (smaller/faster model recommended) gating_model = meta-llama/Llama-3.2-3B-Instruct [STT] # Available providers: speechmatics, deepgram, deepgram-flux # - speechmatics: Speechmatics with SMART_TURN detection # - deepgram: Deepgram Nova-2 with endpoint detection # - deepgram-flux: Deepgram Flux with built-in turn detection (recommended) provider = deepgram-flux [TTS] # Available providers: elevenlabs, qwen3 provider = qwen3 # Qwen3-TTS Configuration (only used if provider = qwen3) # Available models: Qwen/Qwen3-TTS-12Hz-0.6B-Base, Qwen/Qwen3-TTS-12Hz-1.7B-Base qwen3_model = Qwen/Qwen3-TTS-12Hz-0.6B-Base # Available devices: mps (Mac), cuda (NVIDIA), cpu qwen3_device = mps # Reference audio file for voice cloning (relative to project root) qwen3_ref_audio = assets/audio/tars-clean-compressed.mp3 [Emotional] # Enable real-time emotional state monitoring via video enabled = true # How often to sample video frames (in seconds) sampling_interval = 3.0 # How many consecutive negative states before intervention intervention_threshold = 2 [Connection] # Transport mode: "robot" (aiortc WebRTC to RPi) or "browser" (SmallWebRTC for browser) mode = robot # Raspberry Pi WebRTC server URL (Tailscale or local network IP) rpi_url = http://100.115.193.41:8001 # Auto-connect to RPi on startup (only for robot mode) auto_connect = true # Delay between reconnection attempts (seconds) reconnect_delay = 5 # Maximum reconnection attempts (0 = infinite) max_reconnect_attempts = 0 [Display] # Enable TARS Raspberry Pi display integration (HTTP commands) enabled = true # URL of TARS display API (Tailscale or local network IP) tars_url = http://100.115.193.41:8001