How does a computer turn speech into text?

Speech recognition converts recorded speech into an estimated written transcript. A microphone first turns changing air pressure into an electrical signal. The computer records measurements of that signal at closely spaced times. These digital samples describe sound; they do not already contain written letters or spaces between words.

A recognition model has learned relationships between sound patterns and text from training data. It processes the sound over time and compares possible word sequences. Modern systems can learn several parts of this process together, rather than following a single fixed list of separate sound-to-letter rules.

Context helps because different words can sound alike. “Weather” and “whether” are homophones: they can have the same pronunciation. In “The weather is sunny,” the surrounding words support one spelling. In “I wonder whether it will rain,” they support the other. The illustration gives the two candidates invented ranking points to show this change. The points are not measurements from a real recognizer.

An output is still an estimate. Background noise, overlapping voices, unfamiliar names, accents and the system’s training can affect recognition. Some systems return alternatives or a confidence estimate, and a developing transcript may change as more speech arrives. A high ranking does not guarantee that the chosen words match what was said. Producing text is also a different task from establishing the meaning or truth of the speech.