Back to glossary

Term

Automatic Speech Recognition (ASR)

Automatic Speech Recognition (ASR), also called Speech-to-Text (STT), is the automatic conversion of spoken language into written text by an acoustic model.

Automatic Speech Recognition — explained in more detail

ASR stands for Automatic Speech Recognition and is synonymous with Speech-to-Text (STT). Classic pipelines run through several stages: audio capture, preprocessing (resampling, normalization, optional noise reduction), feature extraction into a mel spectrogram, an acoustic model for phoneme probabilities and a language model that turns those into the final text. Modern end-to-end architectures such as Whisper, NVIDIA Parakeet or Conformer collapse feature extraction and decoding into a single transformer stack — the conceptual stages still exist, but they are no longer separate modules.

Example / In practice

In practice, cloud APIs (Google Cloud Speech, AWS Transcribe, Azure Speech, Deepgram, AssemblyAI) sit next to open models (Whisper, Distil-Whisper, Parakeet, Canary). Use cases range from voice input in editors and CLIs to dictation, captioning, meeting transcripts, call-center analytics and voice control. ASR quality is mainly measured by Word Error Rate (WER); for languages without clear word boundaries (e.g. Chinese, Japanese) the Character Error Rate (CER) is used as well. WER depends heavily on audio quality, speaker accent and domain vocabulary — an ASR system does not produce a “correct” transcript but a probability-weighted prediction.

TTS (Text-to-Speech) is the reverse direction — it synthesizes speech from text. Speaker diarization answers “who is speaking when” and usually runs as a separate model in parallel or after ASR. NLU (Natural Language Understanding) consumes the ASR output and extracts intent and entities from the already-transcribed text — ASR itself only produces the words, not the meaning.

Entdecke mehr