Optimum Speech Classification and Its Application to Adaptive Noise Cancellation.
MASSACHUSETTS INST OF TECH LEXINGTON LINCOLN LAB
Pagination or Media Count:
The problem of determining whether a given interval of a speech signal should be classified as voiced speech, unvoiced speech or silence is formulated as a test of statistical hypotheses. A robust detector is obtained by modelling the speech and the acoustic background noise signals as correlated Gaussian random processes. The methods of statistical decision theory are applied to these models to synthesize an optimum, minimum probability of error, classifier. The optimum classifier is an estimator-correlator receiver which is well approximated using a linear phase high pass filter in the unvoiced channel and a linear phase low pass filter in the voiced channel. A clutter filter appears in the reference channel which tries to eliminate as much noise as possible before forming the unvoiced and voiced correlations. The statistics of the noise are learned during the silent intervals which makes the classifier adaptive to time-varying noise statistics.
- Voice Communications