|Stochastic Systems Group|
A blind algorithm for recovering articulator positions from acoustics
Los Alamos National Lab
MIMICRI is a signal processing algorithm that has been shown to blindly infer and invert memoryless nonlinear functions of unobservable bandlimited signals, such as the mapping from the unobservable positions of the speech articulators to observable speech sounds. This blind inversion can be accomplished because the bandwidth of signals almost always increases when transformed by a nonlinear function. Thus, if we transform the observable signals to have the same pass-band as the unobservable signals, then we are within an affine transform of the unobservable signals. We review results of using MIMICRI on toy problems and on human speech data. We note that MIMICRI requires that the user specifies two parameters, the dimensionality and pass-band of the unobservable signals, but the user may not know the best values to use. We show how to use cross-validation techniques with MIMICRI to help estimate parameters that previously needed to be specified. An unexpected consequence of this work is that it helps separate signals that have different frequency characteristics. For example, since lip motions tend to be slower than tongue motions, we may be able to separate the acoustic effects of lips from the acoustic effects of the tongue, or possibly add noise robustness to speech recognition.
Problems with this site should be emailed to email@example.com