University System of Maryland

Carol Espy-Wilson

Srikanth Vishnubhotla

Amit Juneja

An apparatus includes a function module, a strength module, and a filter module. The function module compares an input signal, which has a component, to a first delayed version of the input signal and a second delayed version of the input signal to produce a multi-dimensional model. The strength module calculates a strength of each extremum from a plurality of extrema of the multi-dimensional model based on a value of at least one opposite extremum of the multi-dimensional model. The strength module then identifies a first extremum from the plurality of extrema, which is associated with a pitch of the component of the input signal, that has the strength greater than the strength of the remaining extrema. The filter module extracts the pitch of the component from the input signal based on the strength of the first extremum.

Multiple pitch extraction by strength calculation from extrema

In some embodiments, a processor-readable medium stores code representing instructions to cause a processor to receive an input signal having a first component and a second component. An estimate of the first component of the input signal is calculated based on an estimate of a pitch of the first component of the input signal. An estimate of the input signal is calculated based on the estimate of the first component of the input signal and an estimate of the second component of the input signal. The estimate of the first component of the input signal is modified based on a scaling function to produce a reconstructed first component of the input signal. The scaling function is a function of at least one of the input signal, the estimate of the first component of the input signal, the estimate of the second component of the input signal, or a residual signal.

Systems and methods for speech extraction

Systems and methods for multiple pitch tracking using a multidimensional function and strength values

A probabilistic framework for acoustic-phonetic automatic speech recognition organizes a set of phonetic features into a hierarchy consisting of a broad manner feature sub-hierarchy and a fine phonetic feature sub-hierarchy. Each phonetic feature of said hierarchy corresponds to a set of acoustic correlates and each broad manner feature of said broad manner feature sub-hierarchy is further associated with a corresponding set of acoustic landmarks. A pattern recognizer is trained from a knowledge base of phonetic features and corresponding acoustic correlates. Acoustic correlates are extracted from a speech signal and are presented to the pattern recognizer. Acoustic landmarks are identified and located from broad manner classes classified by the pattern recognizer. Fine phonetic features are determined by the pattern recognizer at and around the acoustic landmarks. The determination of fine phonetic features may be constrained by a pronunciation model. The most probable feature bundles corresponding to words and sentences are those that maximize the joint a posteriori probability of the fine phonetic features and corresponding acoustic landmarks. When the hierarchy is organized as a binary tree, binary classifiers such as Support Vector Machines can be used in the pattern classifier and the outputs thereof can be converted probability measures which, in turn may be used in the computation of the aforementioned joint probability of fine phonetic features and corresponding landmarks.

Growing community of inventors

Washington, DC, United States of America

Carol Espy-Wilson