Voice AI applications require combining speech recognition, LLM reasoning, and text-to-speech synthesis. Architecture: microphone input -> VAD (Voice Activity Detection) -> Whisper STT -> GPT-4o -> TTS -> audio output. Implementation: 1) Speech capture: use pyaudio for real-time audio capture, webrtcvad for voice activity detection to avoid processing silence. 2) Whisper STT: openai.audio.transcriptions.create(model="whisper-1", file=audio_file, language="en") - returns text. For real-time, use faster-whisper (4x faster, same accuracy). 3) GPT-4o processing: maintain conversation history, add system prompt for voice context ("You are a voice assistant. Keep responses under 50 words as they will be spoken aloud. No markdown."). 4) TTS output: openai.audio.speech.create(model="tts-1", voice="alloy", input=response_text) - returns audio. Stream to speakers in real-time. 5) Wake word detection: use Porcupine (Picovoice) for always-on keyword detection without sending audio to cloud. Production: implement interrupt capability (user can speak while TTS is playing), handle network errors gracefully, add explicit end-of-speech detection. Latency target: <500ms from end of speech to first audio output.