Vocabulary independent discriminative utterance verification for non-keyword rejection in subword based speech recognition

• US 5,675,706 A
• Filed: 03/31/1995
• Issued: 10/07/1997
• Est. Priority Date: 03/31/1995
• Status: Expired due to Term

• Current Assignee: Alcatel-Lucent USA, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Lee, Chin-Hui, Sukkar, Rafid Antoon
• Primary Examiner(s): TSANG, FAN S
• Application Number:
  US08/414,243
• Time in Patent Office: 921 Days
• Field of Search: 395/2.6, 395/2.63-2.66, 398/2.61
• US Class Current: 704/256
• CPC Class Codes: G10L 15/04 Segmentation; Word boundary detection
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 2015/088 Word spotting

Abstract

A verification system to determine unknown input speech contains a recognized keyword or consists of speech or other sounds that do not contain any of the keywords. The verification system is designed to operate on the subword level, so that the verification process is advantageously vocabulary independent. Such a vocabulary-independent verifier is achieved by a two-stage verification process comprising subword level verification followed by string level verification. The subword level verification stage verifies each subword segment in the input speech as determined by an Hidden Markov Model recognizer to determine if that segment consists of the sound corresponding to the subword that the HMM recognizer assigned to that segment. The string level verification stage combines the results of the subword level verification to make the rejection decision for the whole keyword. Advantageously, the training of this two-stage verifier is independent of the specific vocabulary set implying that when the vocabulary set is update or changed the verifier need not be retrained and can still be reliably verifying the new set of keywords.

Claims

1. A method for use in a speech recognition system to verify whether input speech signals comprising digitized speech represents a probable keyword, said probable keyword being determined by a speech recognizer, said method comprising the steps of:

processing said digitized speech into recognizer observation vectors;

processing said recognizer observation vectors in a Hidden Markov Model (HMM) keyword recognizer, said HMM keyword recognizer having output signals representing said probable keyword, one or more subword segments of said probable keyword, and a likelihood score for each of said one or more subword segments;

developing a plurality of subword verification likelihood ratio scores for said one or more subword segments of said probable keyword by generating two probability density functions (PDFs) for each subword, wherein a first of said PDFs represents a PDF of subword verification scores when said subword is correctly recognized and a second of said PDFs represents a PDF of subword verification scores when said subword is not correctly recognized;

developing a keyword verification confidence score from said plurality of subword verification likelihood ratio scores for each of said one or more subword segments;

verifying whether said probable keyword is present in said input speech signals by comparing said keyword verification confidence score to a threshold; and

delivering as an output said probable keyword if said threshold test is met, and delivering as an output an indication that no keyword is detected if said threshold test is not met.

2. A method in accordance with claim 1 wherein said subword verification score comprises a difference between a subword verification model score and an anti-subword verification model score.

3. A method in accordance with claim 2 wherein said subword verification model score comprises a weighted sum of the components of a subword verification model feature set.

4. A method in accordance with claim 3 wherein said weights in said subword verification score are determined by a discriminative training procedure.

5. A method in accordance with claim 2 wherein anti-subword verification model score comprises a weighted sum of components of an anti-subword verification model feature set.

6. A method in accordance with claim 5 wherein said weights in said anti-subword verification score are determined by a discriminative training procedure.

7. A method in accordance with claim 3 wherein said subword verification model feature set is computed for each of said one or more subword segments.

8. A method in accordance with claim 5 wherein said anti-subword verification model feature set is computed for each said subword segment.

9. A method in accordance with claim 7 wherein said subword verification model feature set for a given subword segment is determined by processing said subword segment through a corresponding HMM model and computing HMM state centroids of the recognizer observation vector, HMM state durations, and an HMM likelihood score.

10. A method in accordance with claim 6 wherein said anti-subword verification model feature set for a given segment is determined by processing said subword segments through an anti-subword HMM model of the recognized subword for that segment and computing HMM state centroids of the recognizer observation vector, HMM state durations, and an HMM likelihood score.

11. A method in accordance with claim 10 wherein anti-subword HMM models are generated specifically for said verification task, and, for a given subword, an anti-subword model is trained to model sounds other than those modeled by said subword HMM model.

12. A method in accordance with claim 2 wherein the weights in said subword verification score are determined by a discriminative training procedure.

13. A method in accordance with claim 1 wherein keyword verification confidence score is generated by taking the logarithm of the product of the subword verification likelihood ratio scores of all of the subwords comprising the keyword;

wherein said keyword verification confidence score represents a logarithmic likelihood difference for said keyword.

14. A keyword detection apparatus that determines whether a digitized speech signal includes one of a preselected plurality of keywords, said apparatus comprising:

means for receiving input signals representing digitized speech and developing a plurality of signals representing feature vectors of said digitized speech;

means responsive to said input signals and said signals representing feature vectors of said digitized speech for developing output signals representing a probable keyword, one or more subword segments of said probable keyword, and one or more likelihood scores for each of said speech segments;

means for developing a plurality of subword and anti-subword model scores for one or more subword segments of said probable keyword, wherein said means for developing a plurality of subword and anti-subword model scores includes means for generating two probability density functions for each subword, wherein a first of said probability density functions represents a probability density function of subword verification scores when said subword is correctly recognized and a second of said probability density functions represents a probability density function of subword verification scores when said subword is not correctly recognized;

means for determining a confidence score by combining said plurality of subword and anti-subword scores of said one or more subword segments; and

means for comparing said confidence score against a threshold value for determining whether the probable keyword is present in said input signals.