ASR systems don’t predict words directly from audio. Instead, they use phonemes, graphemes, or subword units as intermediate representations. This guide explains how these units work and how they affect accuracy, latency, and real-time ASR design.