Speech latency in AI is the total delay from user speech to system response, involving STT, LLM processing, and TTS stages. Keeping this under 500–800 ms is key to creating natural, seamless interactions.