How speech recognition works: What Wikipedia says

A speech recogniser turns audio into numerical features, uses an acoustic model to map them to sound units.

A speech recogniser turns audio into numerical features, uses an acoustic model to map them to sound units, and uses a language model to pick the most plausible word sequence. Modern systems use deep neural networks, often end-to-end, and accuracy is rated by word error rate.

What the sources say

  • Wikipedia, Speech recognition. Speech recognition systems have an acoustic model that maps audio to sound units and a language model that scores which word sequences are plausible; hidden Markov models were dominant from the 1980s, deep neural networks made a major breakthrough in 2010, and end-to-end systems that learn all components at once appeared from about 2014.
  • Wikipedia, Speech recognition. Speech recognition accuracy is usually rated with word error rate (WER): the recognised text is aligned with a reference transcript and substitutions, deletions and insertions are counted and divided by the number of reference words.
  • Wikipedia, Speech recognition. In 2017 Microsoft researchers reached human parity in transcribing conversational speech on the Switchboard benchmark.

All read on 22 October 2026.

What Autrans does here

Autrans publishes word error rates only from named public test sets, each tied to the kind of speech and the corpus it was measured on.

Questions people ask

How speech recognition works: what do the sources say?

A speech recogniser turns audio into numerical features, uses an acoustic model to map them to sound units, and uses a language model to pick the most plausible word sequence. Modern systems use deep neural networks, often end-to-end, and accuracy is rated by word error rate.

Where do these facts come from?

They come from Wikipedia, Speech recognition, read on 22 October 2026. Each bullet above links to the page it rests on, so you can read the original wording.

What does Autrans do about this?

Autrans publishes word error rates only from named public test sets, each tied to the kind of speech and the corpus it was measured on.

Related

Start transcribing free

Send Autrans into your next meeting, use it in the room, or upload a recording. Your first 30 minutes are free, given once, and no credit card is needed to start.

Last checked 22 October 2026. How we measure accuracy. Spotted a mistake? Write to hello@autrans.online.