For whatever it's worth, I've seen AT&T experiment with voice recognition for authentication. One of the Achiles' Heels of the system is you can collect a lot o

For whatever it's worth, I've seen AT&T experiment with voice recognition for authentication. One of the Achiles' Heels of the system is you can collect a lot of the samples you need to bypass it from peoples' voicemail greetings. Especially if like for some jobs, they're required to re-record the greeting every day, i.e. "Today is Monday, June 17th, and I'm in the office..." That being said, linear predictive codecs used in mobile/lower bandwidth IP networks have damaged the sound quality of the PSTN quite significantly (some are even vocoderized). If you're just dealing with, say, g.711 in wireline networks (when this was written, aside from g.722 in some ISDN networks, g.711 was used almost exclusively; GSM was very much a niche in its infancy), you might be able to squeak by with enough data to reliably recognize people.