Automatic Speech Recognition (ASR)

Automatic Speech Recognition (ASR) is the technology that converts spoken audio into machine-readable text by mapping acoustic signals to linguistic units. A modern ASR pipeline captures waveforms, computes Mel spectrograms, and feeds them to a neural acoustic model—often a Transformer like Whisper, Conformer, or wav2vec 2.0—to predict token sequences. In end-to-end systems such as Whisper the model emits text directly; in hybrid systems an external language model takes part in decoding through beam search, shallow fusion, or n-best rescoring. Post-processors add capitalization and punctuation, while speaker diarization is a separate task that operates on the audio signal, not on the transcript. Advanced systems deliver real-time streaming with word-level timestamps and multilingual support, and can be fine-tuned on in-domain audio to reduce word-error rate (WER), though the size of the gain depends on the base model, the domain, and audio quality. ASR powers captions, voice assistants, call-center bots, and Retrieval-Augmented Generation pipelines that ground LLM responses in spoken data. Key metrics include WER, real-time factor (RTF), and latency; challenges—accent variance, background noise, privacy—are mitigated by noise-robust training, on-device inference, and encryption.

Related terms

• Automatic speech recognition AI
• Define text to speech
• Speech AI
• Speech synthesizers use to determine context before outputting
• Speech-to-Text

Related services: Agentic AI consulting.

← Back to the AI Glossary