Testing the assumptions of linear prediction analysis in normal vowels

Area of Science:

• Speech processing
• Dynamical systems theory
• Acoustic phonetics

Background:

• Gaussian linear prediction analysis is widely used in speech technology.
• Current speech analysis often assumes linear and Gaussian dynamics.
• Recent research explores nonlinear and non-Gaussian models for speech.

Purpose of the Study:

• To test the validity of Gaussian linear prediction analysis for speech dynamics.
• To provide experimental evidence for nonlinearity and/or non-Gaussianity in speech signals.
• To investigate the success of linear Gaussian models as approximations.

Main Methods:

• Development of an improved surrogate data testing method.
• Application of the test to U.S. English vowels from male and female speakers.
• Circumvention of known surrogate data testing problems through calibrated statistics and experimental protocols.

Main Results:

• Experimental evidence shows Gaussian linear prediction analysis cannot extract all dynamical structure from speech time series.
• Robust evidence undermines the validity of purely linear techniques.
• Findings support the presence of dynamical nonlinearity and/or non-Gaussianity in speech.

Conclusions:

• Linear Gaussian models are insufficient for fully capturing speech dynamics.
• Nonlinear and non-Gaussian dynamics are likely present in real speech.
• Hybrid linear/nonlinear/non-Gaussian models are recommended for accurate speech modeling.