JPS61145600A - Word recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a word recognition device used in a speech recognition device.

Structure of the conventional example and its problems In the conventional example of the word speech recognition method, the input speech is first recognized in units of phonemes, and the recognized phoneme sequence and the recognition target word stored in the word dictionary are used. A method is known in which the degree of similarity between a phoneme sequence and a phoneme sequence is determined, and the word corresponding to the phoneme sequence in a word dictionary with the highest degree of similarity is selected as a recognized word. This is shown in "Word speech recognition system using the outline form of the speech spectrum and its dynamic characteristics" by Miwa et al., Journal of the Acoustical Society of Japan 34 (1978).

A conventional word recognition method will be described below with reference to the drawings. FIG. 1 shows the configuration of a conventional word recognition device. In the figure, 1 is a similarity calculation unit, 2(r
a word dictionary storing phoneme sequences of i-recognition target words;
3 is an inter-phoneme likelihood storage unit in which the likelihood between phonemes is stored;
4 is a word determination unit that selects the word with the highest degree of similarity. The recognition method of the word recognition device configured as described above will be explained below.

The phoneme recognition device (
The input speech signal is converted into a phoneme sequence (not shown) and sent to the similarity calculation unit 1 shown in FIG. The degree of word similarity between the phoneme sequences of the 166 words stored in the 166 words to be recognized is calculated using the likelihood between phonemes stored in the likelihood between phonemes storage unit 3. Then, the obtained 166 word similarities are sent to the word determination section 4, the one with the maximum word similarity is selected, and the word corresponding to the phoneme sequence of the word dictionary for which the maximum similarity has been obtained is outputted as a recognized word.

Next, a method for calculating word similarity in the word determination section 4 will be explained. The phoneme series of dictionary entries in word dictionary 2 is D
(D,,D2...D□), input phoneme sequence w
(w, , W2...-WJ), fc dashi,
Let J be the dictionary entry and the number of phonemes in the input phoneme sequence W. The word similarity S (D, W') is determined by the recurrence formula shown in Equation 1.

Here, L=qCi-1*1-1) La=q(i-1,j-2)+1a(j-1)Laa=
g(i-1,j-3)+1aa(j-2)+1aa(
j-1) Lo = g(i-2,1-1)+4o(1-1)Lo
o ==g (i -3, 1-1) + Joo (i-2)
10o(i-1)g(o,o)=J(1+1,J
+1)=Oq(t,o)=q(o+j)=-ω (however, i≠'+]≠0) qCi, j+1>=q(1+1*i)=-■
(However, i≠I+1, i≠J+1) In equation 1, j
i (i, j) is the i-th phoneme D in the phoneme series of the dictionary entry
The likelihood between i and the input phoneme sequence WOj-th phoneme Wj is shown. Ha (j) is the likelihood that the wi phoneme is added to the same root, !
10(i) indicates the likelihood that the phoneme of Di is dropped. Also,
AaaN), 1oo(i)l: Shows the likelihood that the phonemes of tWi are added consecutively and the likelihood that the phonemes of Di are consecutively dropped. Equation 1 is: ■ Word boundaries always correspond. ■ Addition or omission of phonemes is within two consecutive additions or omissions. ■Addition and omission do not occur consecutively. It represents the degree of similarity obtained as a result of the optimal correspondence when the phoneme sequence of the dictionary entry is made to correspond to each phoneme Di, Wj of the input phoneme sequence W with the following restriction. Also, β(i9mouth) in Formula 1.

ga(i), gaa(i), 6o(i), 1oo(i)
Each likelihood value is a Confusion Matri that represents the probability of an error in phoneme recognition, including the addition or omission of phonemes, obtained from the results of performing phoneme recognition R on a large number of voices in advance.
It is calculated as the logarithm value of each component of x and stored in the inter-phoneme likelihood storage unit 3. This Confusion M
In atrix, the sum of the appearance probabilities of all recognized phonemes (including omissions) for one input phoneme is 1.

In other words, the conventional word recognition method calculates the word similarity obtained by equation 1 for the input phoneme sequence for all n dictionary items included in the word dictionary, and selects the word corresponding to the one with the highest word similarity. This is a recognized word.

However, the conventional method is effective when noises such as breathing sounds or sighs generated from the mouth, or meaningless sounds such as "um" are added to the beginning or end of the input speech. It often doesn't work. Figure 2 shows an example of the phoneme sequence resulting from phoneme recognition when noise due to breathing sounds is added before the voice ``Ichikawa''. Since the silent part up to this point was recognized as 101 (consonant), the three phonemes 1ha01 are added consecutively to the beginning of the word. When the noise emitted from the mouth is added to the beginning or end of a word in this way, the noise part is easily recognized as two or more phonemes. Since a silent part is easily recognized as a consonant 101, three or more phonemes are often added consecutively. The recurrence formula in Equation 1 is a recurrence formula with the restriction that the addition of phonemes is limited to two consecutive times, so if three or more phonemes are added consecutively, as shown in Figure 3, This method has the disadvantage that it is not possible to obtain an optimal correspondence between the phoneme Di of the dictionary entry and the input phoneme Wi, and the word similarity of the correct word with the dictionary entry is a small value, making it difficult to obtain correct recognition results. .

OBJECTS OF THE INVENTION The present invention eliminates the above-mentioned drawbacks of the prior art, and provides a word recognition device that exhibits little performance deterioration even when noise is added to the beginning or end of a speech word.

Structure of the Invention The basic structure according to the present invention includes a word dictionary storage unit storing phoneme sequences of words to be recognized, a phoneme-to-phoneme likelihood storage unit storing phoneme-to-phoneme likelihoods, and a word dictionary storage unit storing phoneme sequences of words to be recognized; A phoneme sequence testing unit that determines whether there is a possibility that noise is added to the end of a word.

If there is a possibility that noise has been added, a word boundary from which the noise part has been removed is determined, and a word boundary re-determining unit generates all modified input phoneme sequences, and the input phoneme sequence or the input phoneme sequence and the modified input phoneme sequence and the word. The word similarity calculation unit calculates the word similarity with the phoneme sequence of the dictionary storage unit, and the word similarity calculation unit selects the maximum word similarity and calculates the maximum word similarity corresponding to the phoneme sequence of the word dictionary. It is equipped with a word judgment unit that outputs words as recognized words, and when there is a possibility that noise is added to the beginning or end of the speech, there are two types of input phoneme sequences: an input phoneme sequence and a modified input phoneme sequence with word boundaries corrected. DESCRIPTION OF AN EMBODIMENT The degree of word similarity between an input phoneme sequence and a phoneme sequence in a word dictionary is determined.An embodiment of the present invention will be described below with reference to the drawings. FIG. 4 shows a block diagram of a word recognition device incorporated in a speech recognition device according to an embodiment of the present invention. In FIG. 4, reference numeral 5 denotes a word similarity calculation unit that calculates the phoneme sequence of the input voice recognized by the previous stage phoneme recognition device (not shown) and the recognition target stored in the word dictionary storage unit 6. The word similarity with the phoneme sequence of the word is calculated.

Further, 7 is a storage unit for storing likelihood between phonemes, and a word similarity calculation unit 5
The inter-phoneme likelihood used in calculating word similarity is stored. This inter-phoneme likelihood can be obtained by performing phoneme recognition in advance on a large number of voices, and then calculating the logarithm value of each component of ConfusionMatxix, which represents the total probability of phoneme recognition errors, including phoneme additions and omissions. I'm using something like this. Further, reference numeral 8 denotes a phoneme sequence verification section which verifies whether the phoneme sequence sent from the previous stage phoneme recognition device includes the consonant 1Q1. If the input phoneme sequence includes the consonant 1o1, the word boundary is redetermined in the word boundary redetermination unit 9, and the redetermined phoneme sequence is sent to the word similarity calculation unit 6 again. It will be done. 1o is a word determination unit that determines the maximum word similarity among the word similarities calculated by the word similarity calculation unit 6.

The word corresponding to the phoneme sequence of the dictionary item stored in the word dictionary storage unit 6 for which the maximum word similarity has been obtained is output as a recognized word.

The operation of the word recognition device configured as described above will be explained. Speech input from a microphone or the like is converted into a series of phonemes by a phoneme recognition device and sent to the word recognition device shown in FIG. The manually generated phoneme sequence is sent to the word similarity calculation section 5 and the phoneme sequence verification section 8 shown in FIG. In the word similarity calculation unit 6, the word similarity between the input phoneme sequence and the phoneme sequence of the recognition target word stored in the word dictionary storage unit 6 is stored in the total inter-phoneme likelihood storage storage unit 7. Calculated using inter-phoneme likelihood. This inter-word similarity is calculated using the equation shown in equation 1. Then, this word similarity is calculated for all phoneme sequences stored in the word dictionary storage section 6.

On the other hand, the phoneme sequence input to the phoneme sequence testing section 8 is tested to see if the phoneme sequence includes a consonant 101. If the input phoneme series does not include a consonant, the following processing is not performed, but if the input phoneme series does include a consonant, noise may have been added to the beginning or end of the word. The input phoneme sequence is sent to the word boundary re-determining unit 11 and the word boundaries are corrected. The word boundary re-determining unit 11 determines that if the position of a consonant included in the input phoneme sequence is closer to the beginning of a word than to the end of the word, there is a possibility that noise has been added to the beginning of the word, and sets the word boundary to the beginning of all phonemes following the consonant. Make corrections. On the other hand, if the consonant is closer to the end of the word than the beginning, there is a possibility that noise has been added to the end of the word, so the word boundary is corrected so that the phoneme before the consonant becomes the end of the word. The phoneme sequence whose word boundaries have been corrected in the word boundary re-determination unit 9 is sent to the word similarity calculation unit 5, where the word similarity with the phoneme sequence in the word dictionary storage unit 6 is calculated. In other words, if the input phoneme sequence contains a consonant, in addition to the word similarity between the input phoneme sequence and the phoneme sequence in the word dictionary, the word similarity between the input phoneme sequence with word boundaries corrected and the phoneme sequence in the word dictionary is determined. Degrees are also calculated. The calculated word similarity is then sent to the word determination unit 1o, and the word that matches the phoneme sequence of the word dictionary with the maximum word similarity is output as a recognized word.

According to this embodiment, if n is at the beginning or end of a voice.

Even if noises such as breathing sounds, coughing, or meaningless voices such as "um" are added, the word boundary re-determination unit 9 removes the noises, making it possible to obtain correct recognition results. This is based on the fact that there is often a silent section between noise and speech as described above, and the phoneme recognition result of this silent section is a consonant 101.

As a result of conducting evaluation experiments using 274 words uttered by a total of 40 male and female speakers using the speech recognition device incorporating the word recognition device according to this embodiment, we found that it works effectively even when noise is added as described above. The average word recognition rate was 95.
I was able to write the whole thing with a good result of 6%.

Effects of the Invention As described above, the present invention includes a word dictionary storage unit storing phoneme sequences of words to be recognized, an inter-phoneme likelihood storage unit storing likelihoods between phonemes, and an input A phoneme sequence verification unit that determines whether noise may be added to the phoneme sequence; a word boundary re-determination unit that corrects word boundaries and generates a modified input phoneme sequence; A word similarity calculation unit that calculates the word similarity between the corrected input phoneme sequence and the phoneme sequence stored in the word dictionary storage unit, and a word similarity calculation unit that selects the maximum word similarity among the calculated word similarities. The present invention is a word recognition device that includes a word judgment unit that outputs a word corresponding to the phoneme sequence of the word dictionary storage unit used in the word dictionary storage unit as a recognition result. When noise such as meaningless speech such as "um" is added, there is often a silent interval between the noise and the speech. If there is a possibility that noise has been added, the word boundaries are corrected in the word boundary re-determining unit, and the input phoneme sequence and the phoneme sequence with the corrected word boundaries are combined. This method calculates word similarity using two types of input phoneme sequences, and has the advantage that correct recognition results can be obtained even for speech with added noise.

[Brief explanation of drawings]

Figure 1 is a block diagram showing the functional configuration of a conventional word recognition device, Figure 2 is a diagram showing an example of a phoneme sequence when noise is added to the beginning of a word, and Figure 3 is a diagram showing the relationship between input phonemes and phonemes in a word dictionary. A diagram showing an example of taking an incorrect response. FIG. 4 is a functional block diagram of a word recognition device in one embodiment of the present invention. 6... Word similarity calculation unit, 6... Word dictionary storage unit, 7... Inter-phoneme likelihood storage storage unit, 8
. . . Phoneme sequence testing section, 9 . . . Word boundary re-determination section, 1o . . . Word judgment section.

Claims (1)

[Claims] A word dictionary storage unit stores the phoneme sequence of the word to be recognized, an inter-phoneme likelihood storage unit stores the likelihood between phonemes, and a word dictionary storage unit stores the phoneme sequence of the word to be recognized, and an inter-phoneme likelihood storage unit stores the likelihood between phonemes. a word boundary re-determining unit that corrects word boundaries and generates a corrected input phoneme sequence when noise is added; and an input speech sequence or input phoneme sequence and corrected input phoneme sequence. a word similarity calculation unit that calculates a word similarity between a phoneme sequence and a phoneme sequence stored in the word dictionary storage unit using an inter-phoneme likelihood stored in the inter-phoneme likelihood storage unit; , a word determination unit that selects the maximum word similarity among the word similarities calculated by the word similarity calculation unit and outputs the word corresponding to the phoneme sequence in the word dictionary storage unit used for calculating the word similarity as a recognition result; A word recognition device comprising: