Event-synchronous analysis for connected-speech recognition
David Peter Morgan · 1988
The motivation for event-synchronous speech analysis originates from linear system theory where the speech-source transfer function is excited by an impulse-like driving function. In speech processing, the impulse response obtained from this linear system contains both semantic information and the vocal tract transfer function. Typically, an estimate of the transfer function is obtained via the spectrum by assuming a short-time stationary signal within some analysis window. However, this spectrum is often distorted by the periodic effects which occur when multiple (pitch) impulses are included in the analysis window. One method to remove these effects would be to deconvolve the excitation function from the speech signal to obtain the transfer function. The more attractive approach is to locate and identify the excitation function and synchronize the analysis frame with it. Event-synchronous analysis differs from pitch-synchronous analysis in that there are many events useful for speech recognition which are not pitch excited. In addition, event-synchronous analysis locates the important boundaries between speech events, such as voiced to unvoiced and silence to burst transitions. In asynchronous processing, an analysis frame which contains portions of two adjacent but dissimilar speech events is often so ambiguous as to distort or mask the important phonetic features of both events. Thus event-syncronous processing is employed to obtain an accurate spectral estimate and in turn enhance the estimate of the vocal-tract transfer function. Among the issues which have been addressed in implementing an event-synchronous recognition system are those of developing robust event (pitch, burst, etc.) detectors, synchronous-analysis methodologies, more meaningful feature sets, and dynamic programming algorithms for nonlinear time alignment. An advantage of event-synchronous processing is that the improved representation of the transfer function creates an opportunity for developing more precise feature sets. However, because the input signal is analyzed using variable time intervals, the feature vectors produced are synchronized to speech events but are nonlinear in time. Therefore any pattern matching or dynamic programming algorithm must contain some mechanism to compensate for this nonlinear relationship between feature vector index and time.