Neural networks for voiced/unvoiced speech classification

Abstract

This paper describes the results of designing, training, and testing a neural network for the voiced/unvoiced speech classification problem. A feed-forward multilayer backpropagation network was used with 6 input, 10 intemal, and 2 output nodes -for a binary decision. The six feames are common and easily computed. Training was done with 72 frames from two speakers; testing was done with 479 frames from four speakers; a total of 2 errors (0.4%) occurred. Thus a small neural network performs well on the V / W problem. 1, Review of Previous ResearchVoiced/Unvoiced (VRnr) classifiers can been grouped into two general categories [SL77]: 1) classifiers which determine the VRnr content of a segment of speech as a byproduct of an attempt to determine the primary pitch of the segment, and 2) classifiers which determine the V / W content of a segment of speech by examining one or more speech signal features which are known to be correlated with the V/W distinction. This paper describes a neural network to do V/UV classification using the latter approach. Our motivation was to produce an accurate V / W classifier for high-quality analysis-synthesis using linear predictive coding.Multi-feature techniques have some drawbacks, and it is important to be aware of the practical limitations imposed by these methods. One problem with the techniques presented below is their lack of immunity to nonstationary noise effects [AR76, CBSO]. This problem is largely due to the fact that these methods do not incorporate some form of continuous adaptation to the environment being sampled-they involve training a V/UV classifier in a fixed noise environment, to operate in that same environment. For a discussion of V / W classification techniques designed to operate in the presence of varying noise conditions, see [CB80, KH87, BG87, KS791. In addition, the tradeoff between time and accuracy is important in many applications. In particular, it may not be practical to use a large number of features in real-time applications, especially if the features used are computationally complex. The ongoing development of digital signal processing hardware should cause a continuous reconsideration of such ?This work was supported in part by NSF Grant MIP-8705454. U. S. Army Research Office -Durham Contract DAAG29-85-K-0191.tradeoffs. In the analysis below, the techniques considered are applicable when high accuracy is desired, computation time is not a severe constraint, and noise effects can be determined in advance.Two issues are involved in building a V/W classifier of this general class. First, it is necessary to determine a set of speech features adequate to the V/UV decision task, and second, it is necessary to find an explicit rule for making the V / W decision based upon the values of the features. Although these two aspects cannot be completely isolated from one another, it is useful to view them as separate parts of the problem. (Sometimes, as in [SBSO], the set of features used is tailored to the classification method.) Separation allows a more sy...

A statistical model for robust integration of narrowband cues in speech

“…In our view, these probabilistic graphical models provide a way to study feature detection that combines the benefits of prior knowledge and learning from examples. They enable researchers to bridge the divide between models based on expert engineering (Holmes, 1998) and those derived by automatic methods (Bendiksen & Steiglitz, 1990).…”

Real‐time implementation and performance evaluation of speech classifiers in speech analysis‐synthesis

“…Several methods of speech classification such as the autocorrelation function (ACF), short‐time energy of the signal (E), average magnitude difference function (AMDF), zero crossing rate (ZCR), cepstrum, discrete wavelet transform (DWT), and so on, that make use of acoustic features, have been reported in the literature [1,4–8]. A hybrid approach of speech classification such as the hidden Markov models, Gaussian mixture model or neural network (NN) model, that uses more than one feature, has been also reported [9‐21].…”

“…Bagavathi and Padma [13] presented a fuzzy c‐implies clustering method for classifying voiced and unvoiced activity using MFCC as features and achieved 91.5% classification accuracy. Bendiksen and Steiglitz [14] used an NN as a classifier for voiced/unvoiced speech classification. They extracted six features, namely, the rms energy of signal, the rms energy of the pre‐emphasized signal, the normalized autocorrelation coefficient of the signal at unit sample delay, the normalized autocorrelation coefficient of the pre‐emphasized signal at unit sample delay, the ratio of the signal energy above 4000 Hz to the signal energy below 2000 Hz, and the product of the signal energy above 4000 Hz to the signal energy below 2000 Hz.…”

Speech processing

“…Switched coders first try to classify the short-time speech frames into a small number of classes, and then use an appropriate coding scheme for each class. Here, the ANNs can be used to classify the frames, for example, as voiced, unvoiced, or silence (Bendiksen and Steiglitz 1990, Cohn 1991, Ghiselli-Crippa and El-Jaroudi 1991.…”