Method for recognizing a keyword in speech

Details Method for recognizing a keyword in speech Method for recognizing a keyword in speech

2000-02-25

2003-01-07

Dorvil, Richemond (Department: 2654)

Data processing: speech signal processing, linguistics, language

Speech signal processing

Recognition

C704S240000, C704S242000

Reexamination Certificate

active

06505156

ABSTRACT:

BACKGROUND OF THE INVENTION
The invention is directed to a method for recognizing a keyword in spoken language.
A modelling of the complete spoken expression has hitherto always been required in the recognition of a keyword in spoken language. The person skilled in the art is familiar with essentially two methods:
M. Weintraub, “Keyword-spotting using SRI's DECIPHER large-vocabulary speech-recognition system”, Proc. IEEE ICASSP. Vol. 2, 1993, pp. 463-466 discloses a method for the recognition of a keyword that employs a speech recognition unit with a large vocabulary. The attempt is thereby made to completely recognize the spoken language. Subsequently, the recognized words are investigated for potentially existing keywords. This method is complex and affected with errors because of the large vocabulary and because of the problems in the modelling of spontaneous vocal expressions and noises, i.e. part of the voice signal that cannot be unambiguously allocated to a word.
Another method employs specific filler models (also: garbage models) in order to model parts of expressions that do not belong to the vocabulary of the keywords (what are referred to as OOV parts; OOV=out of vocabulary). Such a speech recognition unit is described in H. Boulard, B. D'hoore and J.-M. Boite, “Optimizing recognition and rejection performance in wordspotting systems”, Proc. IEEE ICASSP, vol. 1, 1994, pages 373-376, and comprises the keywords as well as a filler model or a plurality of filler models. One difficulty is to design or train a suitable filler model that contrasts well with the modelled keywords, i.e. exhibits high discrimination with respect to the keyword models.
Further, hidden Markov models (HMMs) are known from L. R. Rabiner, B. H. Juang, “An Introduction to Hidden Markov Models”, IEEE ASSP Magazine, 1986, pp. 4-16, or A. Hauenstein, “Optimierung von Algirthmen und Entwurf eines Prozessors für die automatische Spracherkennung”, Doctoral Dissertation at the Chair for Integrated Circuits of the Technical University, Munich, Jul. 19, 1993, pp. 13-35. It is also known from Rabiner et al or Hauenstein to define a best path with the Viterbi algorithm.
Hidden Markov models (HMMs) serve the purpose of describing discrete stochastic processes (also called Markov processes). In the field of speech recognition, hidden Markov models serve, among other things, for building up a word lexicon in which the word models constructed of sub-units are listed. Formally, a hidden Markov model is described by:
&lgr;=(
A, B
,
&pgr;
)  (0-1)
with a quadratic status transition matrix A that contains status transition probabilities A
ij
:
 
A={A
ij
} with
i,j=
1, . . . ,
N
  (0-2)
and an emission matrix B that comprises emission probabilities B
ik
:
B={B
ik
} with
i=
1, . . . ,
N; k=
1, . . . ,
M
  (0-3)
An n-dimensional vector
&pgr;
serves for initialization, an occurrence probability of the N statusses for the point in time t=1 defined:
&pgr;={&pgr;
i
}=P
(
s
(1)=
s
i
)  (0-4)
In general,
P
(
s
(
t
)=
q
t
)  (0-5)
thereby indicates the probability that the Markov chain
s={s
(1),
s
(2),
s
(3), . . . ,
s
(
t
), . . . }  (0-6)
is in status q
t
at time t. The Markov chain s thereby comprises a value range
s
(
t
)&egr;{
s
1
,s
2
, . . . ,s
N
}  (0-7)
whereby this value range contains a finite set of N statusses. The status in which the Markov process is at time t is called q
t
.
The emission probability B
ik
derives from the occurrence of a specific symbol &sgr;
k
in the status s
i
as
B
ik
=P
(&sgr;
k
|q
t
=s
i
)  (0-8)
whereby a character stock &Sgr; having the size M comprises symbols &sgr;
k
(with k=1 . . . M) according to
&Sgr;={&sgr;
1
,&sgr;
2
, . . . ,&sgr;
M
}(0-9)
A status space of hidden Markov models derives in that every status of the hidden Markov model can have a predetermined set of successor statusses: itself, the next status, the next but one status, etc. The status space with all possible transitions is referred to as trellis. Given hidden Markov models of the order 1, a past lying more than one time step in the past is irrelevant.
The Viterbi algorithm is based on the idea that, when one is locally on an optimum path in the status space (trellis), this is always a component part of a global optimum path. Due to the order 1 of the hidden Markov models, only the best predecessor of a status is to be considered, since the poorer predecessors have received a poorer evaluation in advance. This means that the optimum path can be sought recursively time step by time step beginning from the first point in time, in that all possible continuations of the path are identified for each time step and only the best continuation is selected.
A respective modelling of the OOV parts is required given the methods described in Weintraub and Boulard et al. In the former instance of Weintraub, the words of the expression must be explicitly present in the vocabulary of the recognition unit; in the latter instance of Boulard et al, all OOV words and OOV noises are presented by specific filler models.
SUMMARY OF THE INVENTION
The object of the invention is comprised in specifying a method that enables the recognition of a keyword in spoken language, whereby the above-described disadvantages are avoided.
According to the method of the invention for recognizing a keyword in spoken language, the keyword is represented by a sequence of statuses W of hidden Markov models. The spoken language are sampled with a predetermined rate and a feature vector O
t
is produced at every sampling time t for a voice signal from the spoken language belonging to the sampling time t. The sequence O of feature vectors O
t
are imaged onto the sequence of the statuses with a Viterbi algorithm, whereby a local confidence standard is calculated on the basis of an emission standard at a status. With the Viterbi algorithm, a global confidence standard is supplied. The keyword in the spoken language is recognized when the following applies:
A method for recognizing a keyword in spoken language, comprising the steps of representing the keyword by a sequence of statuses W of hidden Markov models; sampling the spoken language with a predetermined rate and providing a feature vector O
t
at every sampling time t for a voice signal from the spoken language belonging to the sampling time t; imaging a sequence O of feature vectors O
t
onto the sequence of statuses with a Viterbi algorithm, whereby a local confidence standard is calculated on the basis of an emission standard at a status; with the Viterbi algorithm supplying a global confidence standard; recognizing the keyword in the spoken language when the following applies C(W, O)<T,
where
C( ) denotes the confidence standard,
W denotes the keyword, presented as a sequence of statuses,
O denotes the sequence of feature vectors O
t
,
T denotes a predetermined threshold.
Otherwise, the keyword in the spoken language is not recognized.
One advantage of the invention is comprised that a keyword is recognized within the spoken language without the expression having to be modelled overall. As a result thereof, a clearly reduced expense derives in the implementation and, accordingly, a higher-performance (faster) method. By employing the (global) confidence standard C as the underlying decoding principle, the acoustic modelling within the decoding procedure is limited to the keywords.
One development is that a new path through the status space of the hidden Markov models in a first status of the sequence of statusses W begins at each sampling time t. As a result thereof, it is assumed at every sampling time that a beginning of a keyword is contained in the spoken language. On the basis of the confidence standard, feature vectors resulting from following sampling times are imaged onto those statusses of the keyword repr

Affiliated with

Also associated with

Rating

Method for recognizing a keyword in speech does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Method for recognizing a keyword in speech, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for recognizing a keyword in speech will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3055189