What problem does it solve? Building a JARVIS-style voice assistant requires wiring together wake word detection, voice activity detection, speech-to-text, LLM reasoning, and text-to-speech — each with its own latency pitfalls, echo problems, and microphone quirks. This Skill provides a verified architecture and working code patterns for assembling that full audio pipeline, including hard-won fixes for common failures like echo self-retriggering and silent microphones. ## Core Features & Use Cases - Full Pipeline Architecture: Wake word (openWakeWord) → VAD (Silero or RMS energy gate) → STT (faster-whisper) → LLM (OpenRouter or Hermes CLI) → TTS (Kokoro or Piper) → playback, with a defined state machine (SLEEP, LISTENING, PROCESSING, SPEAKING, FOLLOW_UP). - Engine Selection Guidance: 2026 comparisons of STT engines (faster-whisper, Nemotron Streaming, Parakeet TDT) and TTS engines (Kokoro, Piper, Chatterbox, Step Audio EditX) with latency, license, and quality trade-offs. - Diagnostics & Pitfall Fixes: A mic self-test script for the "never hears me" problem, echo cancellation via mic draining, persistent output streams to avoid PortAudio open/close latency, and Windows .bat/PYTHONPATH gotchas. - Use Case: Ask the agent to build a single-file jarvis.py that listens for "Hey Jarvis", transcribes your command with Whisper tiny, gets a reply from an LLM, and speaks it aloud through a British male voice — all running locally on CPU. ## Quick Start Ask the agent to build a minimal single-file voice assistant that wakes on "Hey Jarvis", transcribes speech with faster-whisper, answers via an LLM, and speaks the reply with Kokoro TTS.
