Telephones you can talk to

A speaker-dependent speech recognizer for connected sequences of digits that uses a time-delayed neural network (TDNN) is described. The advantage of using a TDNN is that, at any given instant, acoustic features of the speech signal for a past interval of time up to the present can be observed simultaneously. Thus, once the system is trained, no detection of boundaries between words is necessary for the recognition. Instead of parsing the speech signal into words, or even further into phonemes, and then recognizing each word or phoneme by some process, the TDNN analyzes the signal in a continuous manner, word recognition units being activated as their corresponding words are completed in the input speech signal. A system of this type will most likely be incorporated in the very near future into a voice-activated telephone.<>