Method for speech processing involving whole-utterance modeling

• US 6,961,703 B1
• Filed: 09/13/2000
• Issued: 11/01/2005
• Est. Priority Date: 09/13/2000
• Status: Expired due to Fees

• Current Assignee: Harris Corporation (L3Harris Technologies, Inc.)
• Original Assignee: ITT Manufacturing Enterprises, Inc. (ITT, Inc.)
• Inventors: Higgins, Alan Lawrence, Bahler, Lawrence George
• Primary Examiner(s): Ometz, David L.
• Assistant Examiner(s): Wozniak, James S.
• Application Number:
  US09/660,635
• Time in Patent Office: 1,875 Days
• Field of Search: 704/256, 704/246, 704/243, 704/254, 704/255, 704/247, 704249-251
• US Class Current: 704/249
• CPC Class Codes: G06F 18/21322 Rendering the within-class scatter matrix non-singular
• CPC Class Codes: G10L 17/02 Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
• CPC Class Codes: G10L 19/0204 using subband decomposition
• CPC Class Codes: G10L 19/022 Blocking, i.e. grouping of samples in time; Choice of analysis windows; Overlap factoring

Abstract

A speech verification process involves comparison of enrollment and test speech data and an improved method of comparing the data is disclosed, wherein segmented frames of speech are analyzed jointly, rather than independently. The enrollment and test speech are both subjected to a feature extraction process to derive fixed-length feature vectors, and the feature vectors are compared, using a linear discriminant analysis and having no dependence upon the order of the words spoken or the speaking rate. The discriminant analysis is made possible, despite a relatively high dimensionality of the feature vectors, by a mathematical procedure provided for finding an eigenvector to simultaneously diagonalize the between-speaker and between-channel covariances of the enrollment and test data.

Claims

1. A method of speaker verification by matching a claimed speaker with a known speaker, including the steps of processing spoken input enrollment speech data and test speech data, generating respective match scores therefrom, and determining whether the test speech data corresponds with the enrollment speech data, the method comprising:

forming enrollment speech data as a first plurality of pair-phrases using a set of words, the set of words consisting of a predetermined number of words, wherein the set of words are words between one to nine and at least one bridging word “

ti”

;

forming test speech data as a second plurality of pair-phrases from the same set of words, the second plurality of pair-phrases different from the first plurality of pair-phrases;

converting, by a Baum-Welch algorithm, the first plurality of pair-phrases into a first set of adapted HMM word models;

converting, by the Baum-Welch algorithm, the second plurality of pair-phrases into a second set of adapted HMM word models;

ordering the first set of adapted HMM word models into a first sequence;

ordering the second set of adapted HMM word models into a second sequence, the second sequence and the first sequence having the same order and the same predetermined number of words; and

comparing the first and second sets of adapted HMM word models using a weighted Euclidean distance.

2. A method of speaker verification according to claim 1, wherein the predetermined number of words comprise five words, namely, “

four”

, “

six”

, “

seven”

, “

nine”

, and “

ti”

.

3. A method of speaker verification according to claim 1, wherein enrollment and test feature vectors are created by concatenating state-mean vectors of the first and second sets of the adapted HMM word models.

4. The method according to claim 3 including the further step of:

comparing said enrollment feature vector obtained from said enrollment with the test feature vector obtained from a speech test to determine the identity of a test speaker voice.

5. The method according to claim 3 wherein said enrollment feature vector has a total dimensionality of 1568.

6. The method according to claim 3 further including the step of:

forming said enrollment feature vector for each known speaker using the difference in vectors between a first and second speaker channel.

7. The method according to claim 6 wherein each speaker vector approximates speaker speech with white noise channel differences.