Scale transform in speech analysis

Contextual invariant-integration features for improved speaker-independent speech recognition

“…Thus, the magnitudes can be seen to be scaling invariant. The scale cepstrum, which has the same invariance property, was also introduced in the work of Umesh et al [47]. The scale transformation is a special case of the Mellin transformation.…”

“…To begin with, Cohen [5] introduced the scale transformation which was further investigated for its applicability in the field of ASR by Umesh et al [47]. Its use in ASR is motivated by the relationship given in Eq.…”

Frequency-warping in speech

“…In our experiments, we c hose vowels that were relatively stationary over 768 samples and the middle 256 samples were used in the computation of the formant e n v elopes. The procedure used to compute the formant e n v elopes are described in detail in [6]. Briey, each frame of speech (consisting of 256 samples) is segmented into Q overlapping subframes, and each subframe is hamming windowed.…”

“…forming X(e f ), the speaker-specic scale constant, , is purely a function of the translation parameter in the log-warped domain. This may be easily seen by considering X1(f) = X(e f ) (5) X2(f) = X(e f ) = X(e f+log ) = X1(f + log ) (6) Therefore, if we h a v e t w o formant e n v elopes that are related by a pure scaling constant, that is independent of frequency but is dependent o n the pair of speakers, then in the logwarped domain, the envelopes are the same except for a translation factor dependent o n .…”

“…Recently, w e h a v e proposed the use of scale-transform based cepstral coecients as acoustic features in speech recognition [6]. The scale-transform based cepstrum is motivated by speaker-normalization techniques.…”