Toward speaker-independent recognition-by-synthesis

A system has been designed which abstracts terminal-analog synthesis parameters as sufficient description of the speech signal to be used for talker-independent speech recognition. These parameters are considered descriptors of the articulatory process, obeying the constraints of articulation. These parameters were employed by a very simple statistical-clustering recognition scheme to test their sufficiency for recognition of a small but very difficult vocabulary from a number of different talkers. Results indicate that is, indeed, a very promising approach to speaker-independent recognition by synthesis.