Speech recognition and synthesis
Speech recognition makes voice an input to systems and makes audio searchable or accessible. Combined with language processing, transcribed text can trigger a search or command; a synthesis step can then speak the answer.
Text reviewed on
← Back to the 7 pillars
When it emerged
Speech recognition research predates today's assistants. IBM presented its Shoebox prototype in 1961: it recognized digits and a small set of spoken commands. It demonstrated a limited vocabulary, not open conversation. Later methods improved the handling of continuous speech and voice variation.
What it is
Speech recognition turns spoken audio into text. Speech synthesis goes the other way, from text to audio. These audio tasks can be combined with NLP but have their own challenges involving signal, noise, accents and voice.
How it works
A microphone turns speech into a digital signal. A system extracts audio features, estimates likely sequences of sounds and words and produces a transcript. It may work after a recording ends or while someone speaks, returning interim results. Speech synthesis runs in the opposite direction, from text to audio; identifying the speaker is a separate task.
What it is for
Speech recognition makes voice an input to systems and makes audio searchable or accessible. Combined with language processing, transcribed text can trigger a search or command; a synthesis step can then speak the answer.
Where it is used
Captions, meeting transcripts, dictation, phone support and voice commands are applications. To produce captions, a service receives audio and returns timed text segments, then a person reviews names and uncertain passages.
Limits and safeguards
Recognizing words, identifying the speaker and generating a voice are different tasks. Noise, accents, poor microphones and specialized vocabulary increase errors. A transcript may drop a negation or confuse names; consequential decisions should check the original audio, and recordings require privacy care.