JPH0155476B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a word speech recognition method that recognizes words by comparing input speech with a word dictionary containing phonetic notation.

(Structure of the conventional example and its problems) FIG. 1 is a diagram for explaining the method of the first conventional example, and is shown as a functional block diagram of an apparatus that implements the method. In the figure, 1 is a parameter extraction part, 2 is a probability density calculation part, 3 is a phoneme recognition part, 4 is a word recognition part, 5 is a phoneme standard pattern part, and 6
7 is a word dictionary section, and 7 is a confusion matrix section. The phoneme standard pattern section 5 stores a standard phoneme pattern that represents the distribution of various parameters of each phoneme in the form of an average value for each phoneme and a covariance matrix between the various parameters. is created in advance through preliminary experiments and the like. The word dictionary section 6 stores a word dictionary in which all words to be recognized are expressed in phonemes, and the word dictionary includes, for example, the word "satsuporo",
"Asahikawa" is written as SAQPORO, ASAHIKAWA, etc. The confusion matrix section 7 contains probabilities of how various phonemes used in dictionary notations will be recognized in actual phoneme recognition, for example, the probability that A will be recognized as A is 85%,
A conflation matrix is stored that shows a value such that the probability that .

Next, the operation of the above conventional example will be explained. The parameter extraction unit 1 analyzes input audio every 10 mS frame, extracts parameters, and creates a parameter time series. Next, the probability density calculation unit 2 compares the obtained parameters with the phoneme standard pattern for each frame, and calculates the probability density that the value of the parameter is generated from each phoneme. Then, the phoneme recognition unit 3
In this step, speech is recognized into the phoneme with the maximum probability density calculated above, and the phoneme recognition result for each frame is combined with the result of phoneme segmentation (dividing the input speech into sections for each phoneme) to generate a recognized phoneme sequence. is created. The word recognition unit 4 calculates the degree of similarity between the recognized phoneme sequence and each dictionary item in the word dictionary unit 6 using a conflation matrix.
The word (dictionary entry) with the highest degree of similarity was selected as the recognized word. Note that the calculation in the probability density calculation unit 2 is based on the following formula. The n parameters in one frame are expressed as a vector 〓. Also,
If the average value of these parameters of a certain phoneme X is expressed as μ _X and the covariance matrix is expressed as Σ _X , then the probability density _φ It is expressed as However, the distribution shape is assumed to be Gaussian distribution.

However, in the above conventional example, the value of the probability density, that is, information on the certainty of phoneme recognition, is lost at the stage of phoneme recognition, and this information is not reflected in word recognition (specifically, similarity calculation The disadvantage was that the accuracy was low.

Next, a second conventional example will be described with reference to FIG. In FIG. 2, a phoneme standard pattern section 14 stores a phoneme standard pattern that represents the average value of various parameters of each phoneme and is prepared in advance through preliminary experiments or the like. The parameter extraction section 11 and word dictionary section 15 are similar to those of the first conventional example shown in FIG. The operation of the second conventional example will be explained. First, as in the first conventional example, the parameter extraction section 11 obtains a parameter time series of input speech. Then, the distance calculation unit 12 calculates the distance between each frame and the standard pattern. Let N parameters in a certain frame be C ₁ , C ₂ , ... C _N , and the average values of these parameters for a certain phoneme X be μ _X1 , μ _X2 ,
...If _{μ XN} , _{then the} distance _d It is calculated by Next, the word recognition unit 13 calculates the degree of similarity for each dictionary item. At the time of this similarity calculation, phoneme segmentation is performed according to the dictionary phoneme sequence that constitutes the dictionary item, and this is done according to the following formula. The likelihood l between a phoneme and a segmented interval corresponding to that phoneme is calculated. Similarity is determined as the average likelihood of each phoneme in the dictionary entry.

Let the phoneme be X, and the frame numbers at the start and end of the segmented section corresponding to this X are
The likelihood l is calculated by the following formula, where N _s and _Ne are taken, and the values of each parameter in the n-th frame are _Co1 , _Co2 , . . . _CoN .

Here, the range of i in the summension of the denominator in distance division differs depending on the phoneme X. For example, when X is the phoneme A (a), the range of i is 5.
The vowels are A, E, I, O, and U. The degree of similarity obtained in the above manner was determined for each dictionary item and used as a recognized word. In this second conventional example, since the phoneme standard pattern has only the average value of the parameters, the degree of variation in the parameter values is not reflected in the distance calculation using the formula, and therefore the degree of similarity is considered to have low accuracy. There was a drawback.

(Objective of the Invention) The present invention is intended to eliminate the drawbacks of the above-mentioned conventional example, and aims to improve the accuracy of similarity calculation, thereby improving the word recognition rate.

(Structure of the Invention) In order to achieve the above object, the present invention compares a parameter time series obtained from input speech with a phoneme standard pattern and a word dictionary, performs phoneme segmentation for each dictionary item, and Calculate the probability density that a phoneme indicates the value of its parameter, and use it directly as
By using it for similarity calculation, it has the effect of improving the accuracy of similarity calculation.

(Description of Embodiment) An embodiment of the present invention will be described below with reference to FIG. 3. In FIG. 3, 21 is a parameter extraction unit, 22 is a probability density calculation unit, 23 is a word recognition unit that performs segmentation, likelihood calculation, and similarity calculation for each phoneme, and 24 is a parameter extraction unit for each phoneme. The distribution is expressed as the average value for each phoneme ( _μX ),
and a phoneme standard pattern section that stores a phoneme standard pattern expressed in the form of a covariance matrix ( _Σ It is.
The parameter extraction section 21, phoneme standard pattern section 24, and word dictionary section 25 are the same as the corresponding parts in FIG. 1, respectively.

The operation of this embodiment will be explained. First, the parameter extraction unit 21 obtains parameters for each frame from the input audio, and then the probability rate density calculation unit 2
In step 2, the probability density that the value of the parameter is obtained from each phoneme standard pattern is calculated. The process up to this point is the same as the first conventional example. Next, the word recognition unit 23
The degree of similarity is calculated for each dictionary item.When calculating this degree of similarity, phoneme segmentation is performed according to the dictionary phoneme sequence that makes up the dictionary item, and the phoneme and its phoneme are segmented according to the following formula. The likelihood l of the segmented interval corresponding to
is calculated, and the similarity is determined as the average of the likelihoods of each phoneme in that dictionary entry. Here, the phoneme is
Let N _s and N _e be the frame numbers at the start and end of the segmented section corresponding to X,
When the value of each parameter in the n-th frame is 〓 _o , the probability density φ is the likelihood l
is defined by the following formula.

Here, the range of i in the summension of the denominator in probability density division differs depending on the phoneme X. For example, when X is the phoneme A, the range of i is 5 vowels, A, E, I,,U. The similarity obtained from the above is calculated for each dictionary item,
The dictionary item with the maximum similarity is selected as a recognized word.

In this embodiment, phoneme segmentation is performed for each dictionary entry, that is, unlike the first conventional example, accurate segmentation is performed without adding or dropping phonemes, and then each phoneme is segmented. By performing the likelihood calculation shown in the formula between the music and the corresponding speech interval, the probability density, that is, the certainty that the speech interval is uttered from that phoneme, is directly reflected in the degree of similarity. This method has the advantage that similarity can be obtained with high accuracy.

(Effects of the Invention) The present invention has the above-described configuration, and provides the following effects. The input speech is segmented according to the dictionary phoneme sequence of each dictionary entry, and the likelihood between each phoneme in the dictionary and the corresponding segmented speech interval is calculated directly using probability density. By defining similarity as the average value of likelihoods, highly accurate similarity values can be obtained.

[Brief explanation of drawings]

1 and 2 are functional block diagrams for explaining a conventional word speech recognition method, and FIG. 3 is a functional block diagram for explaining an embodiment of the word speech recognition method according to the present invention. 21... Parameter extraction section, 22... Probability density calculation section, 23... Word recognition section, 24... Phoneme standard pattern section, 25... Word dictionary section.

Claims (1)

[Claims] 1 In a word speech recognition method that performs word recognition of input speech using a word dictionary in which words to be recognized are expressed in phonemes and standard patterns of each phoneme expressed as a distribution of acoustic parameters of each phoneme, input The speech is checked against each dictionary entry in the word dictionary, the input speech is segmented for each phoneme according to the dictionary phoneme sequence that constitutes each dictionary entry, and the segmented speech is segmented using the standard pattern of the phoneme. A word speech recognition method that calculates the probability density that an interval is generated from the phoneme, and uses this probability density value to determine the degree of similarity between each dictionary entry and the input speech to recognize the word.