Conversational AI applications aim to replicate the fluidity and intelligence of human conversations, with latency being a crucial factor in achieving this goal. Such applications involve four primary components: speech-to-text, turn-taking, text processing with large language models (LLMs), and text-to-speech, each contributing to overall latency. Minimizing latency is essential to maintain the realism of interactions, as each component's delay can accumulate, disrupting the conversational flow. Automatic Speech Recognition (ASR) converts audio to text, with latency determined by the time between speech end and text generation completion. Turn-taking relies on Voice Activity Detectors to ensure natural conversation flow without unnecessary interruptions. Text processing with LLMs generates responses, with latency influenced by model choice, prompt length, and knowledge base size, while text-to-speech translates processed text into audible speech, with recent advances significantly reducing delay. Additional factors like network latency, function calling, and telephony can further affect response times. Companies like ElevenLabs are focused on optimizing each component to achieve seamless, realistic conversations by targeting sub-second latency and leveraging state-of-the-art models.