Utterance verification using word based minimum verification error training for recognizing a keyboard string

• US 5,717,826 A
• Filed: 08/11/1995
• Issued: 02/10/1998
• Est. Priority Date: 08/11/1995
• Status: Expired due to Term

• Current Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: LoCicero, Joseph Lawrence, Szeszko, Grzegorz, Setlur, Anand Rangaswamy, Sukkar, Rafid Antoon
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): MATTSON, ROBERT
• Application Number:
  US08/514,034
• Time in Patent Office: 914 Days
• Field of Search: 395/2.45, 395/2.49, 395/2.6-2.61, 395/2.64, 395/2.65
• US Class Current: 704/252
• CPC Class Codes: G10L 15/144 Training of HMMs
• CPC Class Codes: G10L 2015/088 Word spotting

Abstract

A speech recognition method and apparatus which has a first stage to provide keyword hypotheses and a second stage to provide testing of those hypotheses by utterance verification. The utterance verification used has three separate models for each word: one keyword verification model, one misrecognition verification model, and one non-keyword verification model. Further, all three are developed independently of the recognizer keyword models. Because of this independence, the three verification models can be iteratively trained using existing speech data bases to jointly provide a minimum amount of verification errors.

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 and a likelihood score for said word;

developing a plurality of at least three verification scores for said probable keyword;

developing a keyword verification confidence score by combining said plurality of word verification scores into word likelihood ratio for said keyword;

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 models used for developing said plurality of verification scores are independent of the models used in the HMM recognizer, and are determined by a discriminative training procedure.

3. A method in accordance with claim 2, wherein said discriminative training procedure includes the step of iteratively training all models used for developing said plurality of verification scores for a keyword to provide minimum verification errors.

4. 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 signal 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 word based verification model scores for said probable keyword;

means for determining a confidence score by combining said plurality of word based verification scores of said word; and

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

5. A method for utterance verification of a speech recognized word hypothesis to verify keywords, comprising the steps of:

defining a plurality of word based verification HMMs, which are determined discriminatively;

defining a plurality of observation vectors corresponding to a word hypothesis as determined by HMM segmentation;

testing said observation vectors corresponding to said word hypothesis against said plurality of word based verification HMMs including a keyword based verification HMM, a misrecognition word based HMM and a non-keyword speech based HMM; and

combining the results of hypothesis testing by said plurality of word based verification HMMs to determine if a threshold has been passed, if it has been passed said word is verified, otherwise it is rejected.

6. A method for utterance verification of a string of speech recognized words each having a speech recognized word hypothesis to verify a keyword string, comprising the steps of:

defining a plurality of word based verification HMMs, which are determined discriminatively;

defining a plurality of observation vectors corresponding to each word hypothesis of the string as determined by HMM segmentation;

testing said observation vectors corresponding to each said word hypothesis against said plurality of word based verification HMMs including a keyword verification HMM, a misrecognition HMM and a non-keyword HMM to obtain confidence scores;

combining the likelihood scores of the hypothesis testing by said plurality of word based verification HMMs for each word of said string and forwarding said confidence scores to a combiner; and

combining said confidence scores of each word in said string; and

, if a threshold has been passed, said string is verified;

otherwise, it is rejected.

7. A method for training verification HMMs for providing word based minimum verification error;

comprising the steps of;

determining parameters of a verification model set, V_q(n), for each of the keywords in the recognizer vocabulary set;

defining for each word, w_q(n), a word verification likelihood ratio, which is a function of a keyword verification likelihood, a misrecognized word verification likelihood, and a non-keyword verification likelihood;

discriminatively training said word verification likelihood to make said keyword verification likelihood large compared to said misrecognized word verification likelihood and said non-keyword verification likelihood when w_q(n), is recognized correctly, to make said misrecognized word verification likelihood large compared to said keyword verification likelihood, when w_q(n), is misrecognized and to make said non-keyword verification likelihood large compared to said keyword verification likelihood when the input speech does not contain any keyword and w_q(n), is recognized; and

adjusting parameters of V_q(n) to minimize a log of an inverse of the word verification likelihood ratio.

8. A method according to claim 7, wherein said training is done using a generalized probabilistic descent technique.