Voice recognition analyzes human speech and converts it to text or commands. It powers smartphone assistants (Siri, Google Assistant, Alexa), voice guidance input, call transcription, and IVR voice menus.
Accuracy improved dramatically once deep learning was adopted in the 2010s. Training models on large volumes of speech to learn the patterns of sound sequences brought practical accuracy even to Japanese, a language full of homophones and written without spaces between words, and support for dialects and accents advanced as well. Business uses include automatically drafting meeting minutes and analyzing call center conversations in real time.
For telephony, voicemail transcription is the best-known application. iPhone's "Live Voicemail" shows a caller's message as live text while the phone is still ringing so you can decide whether to pick up; Apple lists Japanese (Japan) among the supported languages. Google Pixel's "Call Screening" also uses speech recognition to find out why someone is calling, but according to Google, calls can be screened automatically only in Australia, Canada, Ireland, the United Kingdom and the United States, while Japan supports manual screening, where you start the screening yourself.
Misuse is a growing concern. Voice cloning is built on speech synthesis: a text-to-speech model learns a speaker's vocal characteristics from a small sample and then produces a convincing imitation of that voice. It is a different line of technology from recognition, which turns speech into text; recognition sits on the verification side, in voice authentication. Because voiceprint authentication can be defeated by a cloned voice, identity checks that combine several factors rather than relying on the voice alone have become important.