Automatic speech recognition (ASR)

Automatic speech recognition, or ASR, converts spoken audio into written text. It is the input layer for most speech-based analytics: a call center transcribing support calls, a voice assistant parsing a spoken command, or a media company captioning video all rely on an ASR model as the first step before any language understanding happens.

Modern ASR systems are trained end to end on large amounts of paired audio and transcript data, learning to map acoustic patterns directly to text, in contrast to older pipelines that separated acoustic modeling, pronunciation lookup, and language modeling into distinct stages. Accuracy is reported using word error rate, and typically depends heavily on audio quality, background noise, accent, and domain-specific vocabulary, so a model trained on general speech can perform noticeably worse on, say, medical terminology or a noisy warehouse floor.

ASR matters as the foundation for speech analytics, voice search, natural language query interfaces, and captioning, and it is usually paired with speaker diarization to attribute text to individual speakers. A common pitfall is judging a vendor's ASR accuracy claim without checking what domain and audio conditions it was measured under; a low reported error rate on clean, single-speaker studio audio says little about performance on a noisy multi-speaker phone call.