JP2001092489A - Continuous voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To additionally register unknown words without decrease in presumption accuracy of parameters concerning words registered in a stochastic language model. SOLUTION: Parameters to be allocated to unknown words for each class are prepared, and a parameter corresponding to the word class of an unknown word to be registered is determined. A stochastic language model to be additionally registered is created by an unknown word N-gram creating part 6 by using the determined parameter, and is additionally registered in the existing stochastic language model.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous speech recognition apparatus for performing speech recognition using a probabilistic language model, and more particularly, to a continuous speech recognition apparatus having a function of additionally registering an unknown word in an existing probabilistic language model. The present invention relates to a voice recognition device and a recording medium.

[0002]

2. Description of the Related Art Conventionally, the following method is known as a method of registering an unknown word in a probabilistic language model used in this type of speech recognition apparatus.

(A) A method using a class language model (Brow
n, PFet.al., "Class-Based n-gramModels of Natura
l Language ", Computaional Linguistics, vol. 18, N
o.4, pp.467-479,1992) This method uses N-gra that stochastically expresses the concatenation of words.
m model (Ng including unigram and biggram)
The ram model is described in, for example, "Speech Recognition by Probabilistic Model", Seiichi Nakagawa, IEICE, pp. 146-64. 1
09) using a word class.

[0004] Here, the word class refers to a set of words classified according to some criterion, for example, a part of speech (noun,
(Verb, adjective). bigram
, The probabilities (b
igm probability) p (wi | wi-1) is the probability p (ci | ci-1) of the concatenation of each word in the class and the probability p (wi | ci-1) of the word appearing from the class. It is approximated by the product (see the following equation).

[0005]

(Equation 1)

[0006] Registration of an unknown word can be realized by additionally registering as a new element in a word class in which the unknown word is classified. For example, if the unknown word u belongs to the word class Cu, the bigram probability p (u | wi-1), p
(wi | u) is the probability p (Ci | C
u) and p (Cu | Ci-1) are represented by the following equations.

[0007]

(Equation 2)

Here, it is necessary to determine the probability p (u | Cu) that the unknown word u appears from the word class Cu.
Prepare a large amount of text data including unknown words separately and estimate from the appearance frequency ratio of unknown word u and the appearance frequency of all words belonging to word class Cu, or appear from class of word w belonging to Cu A method has been proposed that assumes the probabilities p (w | Cu) to be equal probabilities (Asadi, A., et. Al .: "Automatic
Modeling for Adding New Words to Large-vocabulary
Continuous Speech Recognition System ", ICASSP-91
Proc. Vol. 1, IEEE, 1991, pp. 305-308).

[0009]

According to the above-mentioned conventional unknown word registration method, when a large amount of data is already prepared for estimating the parameters of the N-gram model with high estimation accuracy, approximation by word class is performed. On the contrary, there is still room for improvement in that the parameter estimation accuracy is reduced.

[0010] Therefore, an object of the present invention is to provide a parameter relating to an unknown word, for example, a parameter relating to an unknown word without reducing the estimation accuracy of the parameter relating to a word registered in the probabilistic language model.
Bigram probability p (u | w), p (w) of unknown word u and known word (registered word) w
| u), and a continuous speech recognition device and a recording medium capable of additionally registering an unknown word in a stochastic language model.

[0011]

In order to achieve the above object, according to the present invention, an input speech is recognized using a known word registered in a stochastic language model and its parameters. A continuous speech recognition device having a function of adding an unknown word to the probabilistic language model; and input means for inputting the unknown word and a word class to which the unknown word belongs, and a parameter assigned to the unknown word for each word class. A first information processing unit that obtains a parameter corresponding to a word class of an unknown word input from the input unit from the storage unit; and a storage unit that stores the parameter and the obtained parameter and the input unit. Second information processing means for creating a probabilistic language model for registration using the input unknown word and additionally registering the probabilistic language model in the probabilistic language model It is characterized in.

According to a second aspect of the present invention, in the continuous speech recognition apparatus according to the first aspect, a classifying means for classifying a known word and its parameters registered in the stochastic language model for each word class, An arithmetic processing unit for acquiring a parameter stored in the storage unit in accordance with a predetermined arithmetic expression based on a parameter of a known word classified for each class.

According to a third aspect of the present invention, in the continuous speech recognition apparatus according to the second aspect, an average value of parameters of known words is obtained by the predetermined arithmetic expression.

According to a fourth aspect of the present invention, in the continuous speech recognition apparatus according to the first aspect, the parameter is unigr.
am probabilities, back-off coefficients, front words in the word set, rear words in the word sets, and bigram probabilities of the word sets.

According to a fifth aspect of the present invention, a computer has a function of recognizing an input speech using a known word registered in a stochastic language model and its parameters, and adding an unknown word to the stochastic language model. Is realized by executing a program, and in a recording medium on which the program is recorded, the program comprises: an input step of inputting an unknown word and a word class to which the unknown word belongs; A first information processing step of obtaining a parameter corresponding to the word class of the unknown word input in the input step from the storage means, and a parameter assigned to the unknown word in the input step. And creating a probabilistic language model for registration using the unknown words input in the input step. And it is characterized in that comprising a second information processing step of additionally registered in the probabilistic language model.

According to a sixth aspect of the present invention, in the recording medium according to the fifth aspect, the program comprises a classifying step of classifying a known word and its parameters registered in the probabilistic language model for each word class; An arithmetic processing step of acquiring a parameter stored in the storage means in accordance with a predetermined arithmetic expression based on a parameter of a known word classified for each word class.

According to a seventh aspect of the present invention, in the recording medium according to the sixth aspect, an average value of a parameter of a known word is obtained by the predetermined arithmetic expression.

According to an eighth aspect of the present invention, in the recording medium according to the fifth aspect, the parameters include a unigram probability, a back-off coefficient, and b of a preceding word in the word set.
It is characterized by including an gram probability and a bigram probability of a word behind in the word pair.

[0019]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 shows the functional configuration of the unknown word registration unit of the continuous speech recognition device.

Reference numeral 1 denotes a registered (hereinafter referred to as "known") probabilistic language model. In this embodiment, an N-gram model is used as an example.

The N-gram model 1 has unigra
As the data of m, a known word (entry of a unigram) w, a unigram probability p (w) of a known word, and a backoff coefficient a (w) of a known word are described. Also, the N-gram model 1 includes
As bigram data, two sets of known words (bigram entry) w1w2, and a bigram probability p (w2 | w1) of the set of known words are described.

Reference numeral 3 denotes a word classifier, and the N-gram
Classify models into word classes. As the word classifier 3, "Japanese morphological analysis system" ChaSen "ver.2.0" developed by Matsumoto et al. (Refer to its instruction manual, NAIST-IS-TR99008)
However, any tool that can be classified into word classes from word expressions can be used.

An N-gram parameter estimating unit 4 estimates an unknown word N-gram parameter list 5 described later based on each word class G (w) created by the word classifier 3.

Reference numeral 6 denotes an unknown word N-gram creating unit, which outputs an N-gram model 7 relating to the unknown word based on the unknown word list 2 describing data relating to the unknown word and the unknown word N-gram parameter list 5. .

The unknown word list 2 describes: • Notation of unknown word u • Word class G (u) to which unknown word u belongs.

The unknown word model 7 is used as data: a unigram probability p (u) of the unknown word u; a backoff coefficient a (u) of the unknown word; a known word w1 preceding the unknown word; bigram probability p (u | w1)-a known word w2 succeeding the unknown word u-has a bigram probability p (w2 | u) of the unknown word u and the known word w2. Estimating these data and adding them to the N-gram model 1 has a new feature of the present embodiment.

Reference numeral 8 denotes an N-gram model synthesizing unit, which synthesizes an N-gram model 1 for a known word and an N-gram model (for registration) for an unknown word to add an unknown word N
-Output the gram model.

Next, a continuous speech recognition apparatus having an unknown word model registration unit having the above-described functional configuration will be described.

FIG. 2 shows the system configuration of the continuous speech recognition apparatus. In FIG. 2, the CPU 100 executes a speech recognition program to perform a continuous speech recognition process, and executes an unknown word registration program described later to register an unknown word in a probabilistic language model. The system memory 110 is a CPU
The input / output data for the information processing performed by 100 is temporarily stored. Hard disk storage device (abbreviated as HDD) 1
Reference numeral 30 stores a probabilistic language model (N-gram model 1) used for continuous speech recognition for storage.
Further, a program for the continuous speech recognition and the unknown word registration is also stored in the HDD 130. These programs are executed by a keyboard or mouse (not shown) instructed to execute.
After loading from 130 to the system memory 110, the program specified by the CPU 100 is executed.

An input interface (I / O) 120 performs A / D conversion of a voice signal input from a microphone and converts a digital voice signal into a C signal for continuous voice recognition.
Deliver to PU100.

The above unknown word list 2 is a document creation function (word processor) of an external device or a continuous speech recognition device.
After that, it is input to the continuous speech recognition device directly or via communication or a portable recording medium (a floppy disk or the like), and is stored in the HDD 130 in advance.

When the user instructs CPU 100 to execute the program for registering an unknown word using a keyboard or a mouse, the program for registering an unknown word is loaded from HDD 130 to system memory 110, and CPU 100 is executed.
Is executed by

In the unknown word registration program, processing for realizing the functions of the components shown in FIG. 1 is defined. FIG. 3 shows details of processing for realizing the function of the word classifier 3 in this program. FIG. 4 shows details of a process for realizing the function of the N-gram parameter estimation unit 4. FIG. 5 shows details of a process for realizing the function of the unknown word N-gram creation unit 6.

The operation of the system shown in FIG. 2 will be described with reference to the flowcharts shown in FIGS.

When the execution of the unknown word registration program is started, the CPU 100 activates the classification program stored in the HDD 130 and is described in the existing N-gram model (1) stored in the HDD 130. un
The combination w1w2 of the known word w in the gram data and the known word in the bigram data is delivered to the classification program. When the CPU 100 executes the classification program and classifies the delivered data w and w1w2 into three word classes G (w), G (w1) and G (w2), the classification result is stored in the system memory 110. By writing to the predetermined area,
The classification result is passed to the unknown word registration program.

More specifically, the CPU 100
After reading the leading data of the N-gram model 1 on the system memory 1 into the system memory 1, the data is stored in the unigram.
m is determined (step S10 in FIG. 3).
→ S20). When the read data is unigram data, the known word w, the known word unigram probability p (w), and the known word back-off coefficient a (w) are extracted from the read data and stored in the system memory 110. The uni
The data is written in the storage area allocated for the program (step S20 → S30). Thereafter, the procedure goes to step S60 through the process of determining whether or not the data classification has been completed in step S60.
10, returning the next data to the N-gram of the HDD 130.
Take out from Model 1.

On the other hand, if the read data is unigram
If it is not the data of (1), it is next determined whether or not the data is bigram. The read data is bigram
In the case of the data of (1), the leading known word w1, the trailing known word w2, and the bigram probability p (w2 | w1) are extracted from the data, and the extracted known words are written to the assigned storage area of the system memory 110 ( Step S20 → S40 →
S50). Thereafter, the procedure returns to step S10 via the process of judging the end of the data classification in step S60, and returns the next data to the N-gram model 1 of the HDD 130.
Remove from

If the read data is not unigram or bigram data, then the procedure returns to step S10 via the process of judging the end of the data classification in step S60, and returns the next data. Take out from the N-gram model 1 of the HDD 130.

The data classification process described above (step S1)
0 to S60), all the data described in the N-gramu model can be classified. When the classification of the last data is completed, a YES determination is obtained in the end presence / absence determination processing of step S60, and thus the processing procedure of FIG. 3 ends. Thereafter, N in FIG.
-The gram parameter estimation process is started.

The CPU 100 sets all ungram probabilities and all backoff coefficients of the word class G (w) classified (created) in the word classification process of FIG.

[0042]

[Outside 1]

And extract the average un for the word class
igram probability

[0044]

[Outside 2]

Back-off coefficient averaged for word class G (w)

[0046]

[Outside 3]

Is calculated (step S100 in FIG. 4).
In addition,

[0048]

[Outside 4]

Represents the average value of f (x) with respect to x.

Similarly, the CPU 100 determines the word class G (w1),
For G (w2), the bigram probability averaged for the word class G (w1) of the preceding known word w1

[0051]

[Outside 5]

The bigram probability averaged for the word class G (w2) of the known word w2 on the back side

[0053]

[Outside 6]

Is also calculated (step S11 in FIG. 4).
0).

If the known word does not belong to one specific word class but can belong to a plurality of word classes, it is excluded from the above calculation for calculating the average.

Since a well-known processing procedure for detecting the duplicate known word may be executed, a detailed description is omitted here.

The result of this calculation is the following unknown word N-gram
It is stored in the HDD 130 (storage means of the present invention) in the form of a parameter list (processing corresponding to reference numeral 5 in FIG. 1, step S120 in FIG. 4).

The unknown word N-gram parameter list 5 may have the following form, for example, having a word class as an array element so that each word class can be searched as a key.

[0059]

[Outside 7]

Next, the CPU 100 shifts to the processing procedure of FIG. 5 and uses the word class described in the unknown word list 2 stored in the HDD 130 as a key to set the unknown word Ng.
The data stored in the ram parameter list 5 is searched (step S200).

The word class of the data of the unigram and the data of the bigram extracted by this search is replaced with the notation of the unknown word. The replacement result is stored in HDD 130 as an N-gram model relating to the unknown word.

In the above processing, for example, the unknown word u
Search for the word class G (u) of

[0063]

[Outside 8]

When is extracted, the word class G (u) is replaced with the unknown word u.

Next, each average value calculated for the word class G (u) in the average value calculated in step S110 described above.

[0066]

[Outside 9]

Is the N-gram for each unknown word u.
The data of the model is stored in the N-gram model 7 relating to the unknown word on the HDD 130 in the following form as model data (FIG. 5).
Steps S210 to S230).

[0068]

[Outside 10]

The CPU 100 adds and inserts the N-gram model 7 for unknown words (for registration) created in this way into the N-gram model 1 for known words (corresponding to the N-gram model in FIG. 1). Thereby, the N-gram model 9 to which the unknown word is added is created.

1) In the above-described embodiment, when an unknown word is additionally registered in the probabilistic language model, the averaging operation of the parameters relating to the known word is performed. However, the averaging operation is performed at any time before the registration of the unknown word. Just do it. After the averaging operation, the created unknown word parameter list 5 is stored in the HDD 130,
It may be used when registering unknown words.

2) As a storage medium for recording a program for registering unknown words, in addition to the above-mentioned HDD 130,
An IC memory such as a ROM and a RAM, and a portable recording medium such as a floppy disk and a magneto-optical disk can be used.

3) The above embodiment is merely an example for describing the present invention, and various modifications can be made in accordance with the technical concept described in the claims. However, even if such a modification is made, the modification is within the scope of the patent of the present application.

[0073]

As described above, according to the present invention, parameters to be assigned to unknown words are required for each word class. A probabilistic language model for additional registration is determined by using a parameter corresponding to the unknown word to be registered.
It is created by the am creating unit 6. For this reason, in the present invention, it is not necessary to make any change to the created stochastic language model of the known word, and the estimation accuracy of the stochastic language model related to the known word is not sacrificed unlike the related art.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a system configuration according to an embodiment of the present invention.

FIG. 2 is a block diagram illustrating a hardware configuration according to the embodiment of the present invention.

FIG. 3 is a flowchart illustrating processing executed by a CPU 100;

FIG. 4 is a flowchart illustrating processing executed by a CPU 100;

FIG. 5 is a flowchart illustrating processing executed by a CPU 100;

[Description of Signs] 3 Word Classifier 4 N-gram Parameter Estimation Unit 6 Unknown Word N-gram Creation Unit 8 N-gram Synthesis Unit

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Akio Ando 1-10-11 Kinuta, Setagaya-ku, Tokyo Japan Broadcasting Corporation Japan Broadcasting Research Institute F-term (reference) 5B091 AA15 CB12 5D015 GG03 HH23

Claims (8)

[Claims] 1. A continuous speech recognition apparatus having a function of recognizing an input speech using a known word registered in a probabilistic language model and its parameters and adding an unknown word to the probabilistic language model. Input means for inputting an unknown word and a word class to which the unknown word belongs; storage means for storing a parameter to be assigned to the unknown word for each word class; and word class of the unknown word input from the input means A first information processing means for acquiring a parameter corresponding to the above from the storage means, and creating a probabilistic language model for registration using the acquired parameters and the unknown word inputted from the input means, A continuous speech recognition apparatus, comprising: a second information processing means for additionally registering a stochastic language model. 2. A continuous speech recognition apparatus according to claim 1, wherein said known words registered in said probabilistic language model and their parameters are classified for each word class; A continuous speech recognition apparatus, further comprising: an arithmetic processing unit that acquires a parameter stored in the storage unit in accordance with a predetermined arithmetic expression based on the parameter of the known word. 3. The continuous speech recognition device according to claim 2, wherein an average value of parameters of known words is obtained by the predetermined operation formula. 4. The continuous speech recognition device according to claim 1, wherein the parameters include a unigram probability, a back-off coefficient, a front word in the word set, and a rear word in the word set. A word and a bigr of the word pair
A continuous speech recognition device comprising an am probability. 5. A computer executes a program having a function of recognizing an input speech using a known word registered in a probabilistic language model and its parameters, and adding an unknown word to the probabilistic language model. And a recording medium that records the program,
An input step of inputting an unknown word and a word class to which the unknown word belongs; and a storage means of the computer storing parameters to be assigned to the unknown word for each word class. A first information processing step of acquiring a parameter corresponding to the word class of the obtained unknown word from the storage means, and a probabilistic process for registration using the obtained parameter and the unknown word input in the input step. A second information processing step of creating a language model and additionally registering the language model in the probabilistic language model. 6. The recording medium according to claim 5, wherein the program comprises: a classification step of classifying a known word and its parameters registered in the probabilistic language model for each word class; A calculation processing step of acquiring a parameter stored in the storage unit in accordance with a predetermined calculation expression based on the classified known word parameters. 7. The recording medium according to claim 6, wherein an average value of a parameter of a known word is obtained by the predetermined arithmetic expression. 8. The recording medium according to claim 5, wherein the parameters include a unigram probability, a back-off coefficient, a bigram probability of a front word in the word set, and a rear side of the word set. The bigram of the word
A recording medium characterized by including a probability.