Devices and methods for speech recognition of vocabulary words with simultaneous detection and verification

• US 5,832,430 A
• Filed: 12/08/1995
• Issued: 11/03/1998
• Est. Priority Date: 12/29/1994
• Status: Expired due to Term

• Current Assignee: Alcatel-Lucent USA, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Lleida, Eduardo, Rose, Richard Cameron
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US08/569,471
• Time in Patent Office: 1,061 Days
• Field of Search: 395/2.4, 395/245, 395/2.6, 395/2.61, 395/2.63, 395/2.64
• US Class Current: 704/256
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

Devices and methods for speech recognition enable simultaneous word hypothesis detection and verification in a one-pass procedure that provides for different segmentations of the speech input. A confidence measure of a target hypothesis for a known word is determined according to a recursion formula that operates on parameters of a target models and alternate models of known words, a language model and a lexicon, and feature vectors of the speech input in a likelihood ratio decoder. The confidence measure is processed to determine an accept/reject signal for the target hypothesis that is output with a target hypothesis signal. The recursion formula is based on hidden Markov models with a single optimum state sequence and may take the form of a modified Viterbi algorithm.

Claims

1. A method for determining a confidence measure of the presence of a known word in an utterance, comprising the steps of:

inputting a signal including signal components representing the utterance into a likelihood ratio decoder;

inputting target model signals representing a target model and alternate model signals representing an alternate model into the likelihood ratio decoder;

inputting language model signals representing a language model and lexicon signals representing a lexicon into the likelihood ratio decoder;

processing the signal including signal components, the target model signals, the alternate model signals, the language model signals, and the lexicon signals in one pass in the likelihood ratio decoder by applying a recursion formula in iterative steps to determine a confidence measure signal that represents a confidence measure of the identification of a particular known word in the utterance.

2. The method of claim 1, wherein said processing step includes obtaining a maximum of a recursively iterated quantity at each of said iterative steps.

3. The method of claim 1, further comprising the steps of:

inputting the confidence measure signal to a decision rule device;

comparing the confidence measure signal to the decision threshold signal in the decision rule device; and

outputting a signal representative of the presence or absence of the known word in the utterance based on the comparison of the confidence measure signal and the decision threshold signal from the decision rule device.

4. The method of claim 3, further comprising the step of outputting a signal representative of a target model hypothesis associated with a known word and the confidence measure signal from said likelihood ratio decoder.

5. The method claim 1, wherein said step of inputting target model signals and alternate model signals includes inputting signals representing hidden Markov models (HMMs) having a number of states N_c and N_a respectively,

wherein said step of inputting a signal including signal components includes inputting signal components that are representative of feature vectors Y={y₁, y₂, . . . y_T } corresponding to times 1, 2, . . . ,T respectively, and

wherein said step of inputting target model signals and alternate model signals includes inputting signals representing the following parameters related to said HMMs,

7. The method of claim 1, wherein said target model and said alternate model are hidden Markov models (HMMs), each having a number of states N,

wherein said step of inputting a signal including signal components includes inputting signal components that are representative of feature vector Y={y₁, y₂, . . . y_T } corresponding to times 1,2, . . . ,T respectively, and

wherein said step of inputting target model signals and alternate model signals includes inputting signals representing the following parameters related to said HMMs,

6. A method for determining a confidence measure of the presence of a known word in an utterance, comprising the steps of:

inputting a signal including signal components representing the utterance and signal components that are representative of feature vectors Y={y₁, y₂, . . . y_T } corresponding to times 1, 2, . . . ,T respectively into a likelihood ratio decoder;

inputting target model signals representing a target model including signals representing hidden Markov models (HMMs) having a number of states N_c and alternate model signals representing an alternate model into the likelihood ratio decoder, said step of inputting target model signals and alternate model signals including inputting signals representing the following parameters related to said HMMs,

space="preserve" listing-type="equation">π

.sub.n.sup.C, a.sub.in.sup.c, π

.sub.m.sup.a, and a.sub.jm.sup.a for 1≦

n≦

N.sub.c, 1≦

M<

N.sub.a, 1≦

i≦

N.sub.c, 1≦

j≦

N.sub.a, and t=1,2, . . . ,T,

where

π

_n^c =the initial state probability of the state n for the model λ

^c,

a_in^c =the transition probability from the state i to the state n for the model λ

^c,

π

_m^a =the initial state probability of the state m for the model λ

^a, and

a_jm^a =the transition probability from the state j to the state m for the model λ

^a ;

inputting language model signals representing a language model including signals representing hidden Markov models (HMMs) having a number of states N_a and lexicon signals representing a lexicon into the likelihood ratio decoder;

processing the signal including signal components, the target model signals, the alternate model signals, the language model signals, and the lexicon signals in the likelihood ratio decoder by applying a recursion formula in iterative steps to determine a confidence measure signal that represents a confidence measure of the identification of a particular known word in the utterance, said processing step including the steps of;

determining in the likelihood ratio decoder signals representing the following parameters,

b_n^c (y→

_t)=the observation density for the feature vector y_t for the state n and the model λ

^c,

b_j^a (y→

_t)=the observation density for the feature vector y_t for the state j and the model λ

^a, and

determining the confidence measure signal based on δ

_t (n,m) determined according to the following recursion formula;

##EQU15##

8. A method for determining a confidence measure of the presence of a known word in an utterance, comprising the steps of:

inputting a signal including signal components representing the utterance into a likelihood ratio decoder, including signal components that are representative of feature vector Y={y₁, y₂, . . . y_T } corresponding to times 1,2, . . . ,T respectively;

inputting target model signals representing a target model and alternate model signals representing an alternate model into the likelihood ratio decoder, said target model and said alternate model being hidden Markov models (HMMs), each having a number of states N, said inputting including inputting signals representing the following parameters related to said HMMs,

space="preserve" listing-type="equation">π

.sub.i.sup.C,a.sub.ij.sup.c,π

.sub.i.sup.a, and a.sub.ij.sup.a for 1≦

i≦

N and t=1,2, . . . ,T,

where

π

_j^c =the initial state probability of the state j for the model λ

^c,

a_ij^c =the transition probability from the state i to the state j for the model λ

^c,

π

_j^a =the initial state probability of the state j for the model λ

^a ;

a_ij^a =the transition probability from the state i to the state j for the model λ

^a ;

inputting language model signals representing a language model and lexicon signals representing a lexicon into the likelihood ratio decoder;

processing the signal including signal components, the target model signals, the alternate model signals, the language model signals, and the lexicon signals in the likelihood ratio decoder by applying a recursion formula in iterative steps to determine a confidence measure signal that represents a confidence measure of the identification of a particular known word in the utterance, said processing step including the steps of;

determining in the likelihood ratio decoder signal representing the following parameters,

b_i^c (y→

_t)=observation density for the feature vector y_t for the state i and the model λ

^a,

b_j^a (y→

_t)=the observation density for the feature vector y_t for the state j and the model λ

^a, and

determining the confidence measure signal based on δ

_t (n,m) determined according to the following recursion formula;

##EQU16##

9. A speech recognizer for determining a confidence measure of the presence of a known word in an utterance, comprising:

means for determining a signal including signal components representing the utterance;

means for determining target model signals representing a target model and alternate model signals representing an alternate model;

means for determining language model signals representing a language model and lexicon signals representing a lexicon;

means for processing the signal including signal components, the target model signals, the alternate model signals, the language model signals, and the lexicon signals in one pass by applying a recursion formula in iterative steps to determine a confidence measure signal that represents a confidence measure of the identification of a particular known word in the utterance.

10. The speech recognizer of claim 9, wherein said means for processing includes means for obtaining a maximum of a recursively iterated quantity at each of the iterative steps.

11. The speech recognizer of claim 9, further comprising:

means for determining a decision threshold signal;

means for comparing the confidence measure signal to the decision threshold signal; and

means for outputting a signal representative of the presence or absence of the known word in the utterance based on the comparison of the confidence measure signal and the decision threshold signal.

12. The speech recognizer of claim 11, further comprising means for outputting a signal representative of a target model hypothesis associated with a known word and the confidence measure signal.

13. The speech recognizer of claim 9, wherein said means for determining target model signals and alternate model signals determines signals representing hidden Markov models (HMMs) having a number of states N_c and N_a respectively,

wherein said means for determining a signal including signal components determines signal components that are representative of feature vectors Y={y₁, y₂, . . . y_T } corresponding to times 1, 2, . . . ,T respectively, and

wherein said means for determining target model signals and alternate model signals determines signals representing the following parameters related to the HMMs,

15. The speech recognizer of claim 9, wherein the target model and the alternate model are hidden Markov models (HMMs), each having a number of states N,

wherein said means for determining a signal including signal components determines signal components that are representative of feature vector Y={y₁, y₂, . . . y_T } corresponding to times 1,2, . . . ,T respectively, and

wherein said means for determining target model signals and alternate model signals determines signals representing the following parameters related to said HMMs,

17. The speech recognizer of claim 9, wherein said means for processing comprises a likelihood ratio decoder that determines the confidence measure as a likelihood ratio.

18. The speech recognizer of claim 9, wherein said means for determining a signal including signal components comprises a feature extractor that determines the signal components as representative of feature vectors of the utterance.

19. The speech recognizer of claim 9, wherein said means for determining target model signals and alternate model signals comprises a memory that stores parameters of the target model and the alternate model.

20. The speech recognizer of claim 9, wherein said means for determining language model signals and lexicon signals comprises a memory that stores parameters of the language model and the lexicon model.

14. A speech recognizer for determining a confidence measure of the presence of a known word in an utterance, comprising:

means for determining a signal including signal components representing the utterance, said means determining signal components that are representative of feature vectors Y={y₁, y₂, . . . y_T } corresponding to times 1, 2, . . . ,T respectively;

means for determining target model signals representing a target model and alternate model signals representing an alternate model, said means determining signals representing hidden Markov models (HMMs) having a number of states N_c and N_a respectively, said means determining signals representing the following parameters related to the HMMs,

space="preserve" listing-type="equation">π

.sub.n.sup.C, a.sub.in.sup.c,π

.sub.m.sup.a, and a.sub.jm.sup.a for 1≦

n≦

N.sub.c, 1≦

M<

N.sub.a, 1≦

i≦

N.sub.c, 1≦

j≦

N.sub.a, and t=1,2, . . . ,T,

where

π

_n^c =the initial state probability of the state n for the model λ

^c,

a_in^c =the transition probability from the state i to the state n for the model λ

^c,

π

_m^a =the initial state probability of the state m for the model λ

^a, and

a_jm^a =the transition probability from the state j to the state m for the model λ

^a ;

means for determining language model signals representing a language model and lexicon signals representing a lexicon;

means for processing the signal including signal components, the target model signals, the alternate model signals, the language model signals, and the lexicon signals by applying a recursion formula in iterative steps to determine a confidence measure signal that represents a confidence measure of the identification of a particular known word in the utterance, said means for processing comprising;

means for determining signals representing the following parameters,

b_n^c (y→

_t)=the observation density for the feature vector y_t for the state n and the model λ

^c,

b_j^a (y→

_t)=the observation density for the feature vector y_t for the state j and the model λ

^a, and

means for determining said confidence measure signal based on δ

_t (n,m) determined according to the following recursion formula;

##EQU17##

16. A speech recognizer recognizer for determining a confidence measure of the presence of a known word in an utterance, comprising:

means for determining a signal including signal components representing the utterance, said means determining signal components that are representative of feature vector Y={y₁, y₂, . . . y_T } corresponding to times 1,2, . . . ,T respectively;

means for determining target model signals representing a target model and alternate model signals representing an alternate model, the target model and the alternate model being hidden Markov models (HMMs), each having a number of states N, said means determining signals representing the following parameters related to said HMMs,

space="preserve" listing-type="equation">π

.sub.i.sup.C,a.sub.ij.sup.c,π

.sub.i.sup.a, and a.sub.ij.sup.a for 1≦

i≦

N and t=1,2, . . . ,T,

where

π

_j^c =the initial state probability of the state j for the model λ

^c,

a_ij^c =the transition probability from the state i to the state j for the model λ

^c,

π

_j^a =the initial state probability of the state j for the model λ

^a,

a_ij^c =the transition probability from the state i to the state j for the model λ

^a ;

means for determining language model signals representing a language model and lexicon signals representing a lexicon;

means for processing the signal including signal components, the target model signals, the alternate model signals, the language model signals, and the lexicon signals by applying a recursion formula in iterative steps to determine a confidence measure signal that represents a confidence measure of the identification of a particular known word in the utterance said means for processing comprising;

means for determining signals representing the following parameters,

b_i^c (y→

_t)=the observation density for the feature vector y_t for the state i and the model λ

^a,

b_j^a (y→

_t)=the observation density for the feature vector y_t for the state j and the model λ

^a, and

means for determining said confidence measure signal based on δ

_t (n,m) determined according to the following recursion formula;

##EQU18##