ABSTRACT
Current performance of speech synthesizers and recognizers makes them already extremely useful for a variety of practical tasks, and they are now deployed in many applications. In both synthesis and recognition, the gap between human and machine performance widens as the conditions become more difficult, for example involving spontaneous speech, emotional speech or noisy environmental conditions. The search for representations and methods that facilitate greater manipulation of speech characteristics within a waveform-based concatenative framework will probably continue to be a focus for speech synthesis research for several years to come. An advantage of automatic techniques is that they can be applied to derive synthesis parameters for any talker of any language or dialect, given enough labelled speech data to train the system. Current Automatic Speech Recognition performance can be very impressive, even for tasks involving very large vocabularies.
