Automatic speech recognition (ASR) is the engineering term for software that converts spoken language into text. It is used interchangeably with speech-to-text and "speech recognition" across research papers, model names, and API documentation. An ASR system maps an audio signal to the most likely sequence of words in a given language, and its output is scored the same way regardless of the name a vendor prints on it.
From HMMs to end-to-end
ASR has been a research field for more than 60 years. For decades the standard was a hidden Markov model paired with a Gaussian mixture model, later upgraded to a hybrid of HMMs and deep neural networks.
Around 2016 the field shifted to end-to-end models trained to map audio straight to text, using techniques like connectionist temporal classification and attention. Self-supervised pretraining, where a model like wav2vec learns from unlabeled audio before fine-tuning, pushed accuracy further still.
The Whisper model family made accurate multilingual ASR broadly available in 2022. This progression is why error rates that sat near 20 percent in the early 2010s now land in the single digits on clean speech.
How automatic speech recognition works
An ASR system turns each slice of audio into acoustic features, then predicts which words those features represent. Classical systems paired an acoustic model, which maps sound to sound units, with a language model, which scores likely word sequences so the output reads as real language.
Modern systems fold both jobs into a single neural network that learns the mapping directly, then decode the best path through the candidate words and add punctuation and formatting. Accuracy is measured by word error rate, the share of words wrong through substitutions, insertions, and deletions.
Why the ASR terminology matters
When a vendor advertises a "speech-to-text API," a "transcription API," or "speech recognition," they describe the same task, measured the same way. "ASR" is the term you will see in model names, papers, and technical docs, while "speech-to-text" tends to dominate product marketing.
Knowing they are synonyms keeps you from over-reading naming differences that carry no real distinction, and it helps when you cross-reference a provider's marketing page against its developer documentation.
Frequently asked questions
What is automatic speech recognition (ASR)? ASR is technology that converts spoken words into written text by analyzing the audio waveform and predicting the word sequence. It is the same capability marketed as speech-to-text or voice-to-text.
How does automatic speech recognition work? ASR extracts numerical features from short frames of audio, runs them through an acoustic and language model or a single end-to-end network, decodes the most likely words, then formats the result. Modern systems learn this mapping from large volumes of transcribed audio.
What is the difference between ASR and speech-to-text? There is no functional difference. Both name the audio-to-text task and are measured by word error rate. ASR is the research and engineering label; speech-to-text is the common product label.
How accurate is automatic speech recognition? Top models now reach single-digit word error rates on clean, clear speech. Accuracy drops with heavy accents, background noise, overlapping speakers, and specialized vocabulary, which is why benchmarks vary widely by audio type.