Automatic Speech Recognition (ASR), also called speech-to-text, converts spoken language into written text using machine-learning models trained on large speech datasets. Modern ASR uses deep neural networks (and increasingly end-to-end transformer models) to map audio directly to text, handling accents, noise, and natural speech. Quality is measured by word error rate (WER) and, for voice agents, by streaming latency.
ASR is the first stage in almost every voice AI pipeline: it turns the caller's audio into text that a language model can reason over before a response is spoken back with TTS. Accurate, low-latency streaming ASR is what makes real-time voice agents feel responsive. Teams evaluate ASR on WER, language and accent coverage, real-time streaming, and support for domain-specific vocabulary.