Speech Beyond 10 milliseconds (Temporal Filtering in Feature Domain)

Some speech analysis techniques used in automatic speech recognition utilize temporal processing of short-term speech feature vectors to derive a final feature set. We argue that such a strategy in in priciple consistent with properties of human hearing and could lead to improved performance of the recognizer in adverse environments. 1 Short term analysis of speech Steady configurations of vocal tract are rare and carry only a little of linguistic information. Linguistic information is in changes of the speech signal. The nonstationary nature of speech was the reason behind an introduction of a short-term analysis of speech. Such an analysis treats short-term (about 10 - 20 ms) segments of speech as independent samples from different and unrelated stationary processes. The short-term analysis technique allows for use of techniques which assume stationarity of the signal such as Fourier analysis of linear predictive analysis for processing of nonstationary signals such as speech. The d..