JPH041916B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device capable of recognizing speech.

In general, it is known that not only does the utterance time differ for each utterance even by the same speaker, but there is also some variation in the accent, and furthermore, if the speaker is different, individual differences will be added to the above-mentioned variation. There is. Therefore, it has been proposed to compensate for these fluctuating components of speech by using statistical methods, and in recent years, speech recognition that aims to improve the recognition rate by introducing such statistical methods has been proposed. Devices are being developed.

The configuration of a conventional speech recognition device of this type is shown in FIG. In the figure, reference numeral 1 denotes a microphone that converts audio into electrical audio signals. 2 is a parameter extraction circuit that extracts voice characteristic parameters from the voice signal obtained by the microphone 1, and extracts 8 frequencies obtained by dividing the voice band (100Hz to 4KHz) into 8 for the voice signal of one recognized voice. Characteristic parameters [xij] of matrix configuration in which spectral values are arranged in 16 samples, (i=1, ..., 8; j=1,
..., 16) is output. 3 is a function memory, based on feature parameters _(o) [wij] extracted in advance from multiple speakers for a specific number n of recognized speech,
Distribution relationship of feature parameter values regarding speaker differences derived assuming that these have a normal distribution _(o) [fij]
is stored for each recognized voice. If this recognized voice consists of five vowels (a, i, u, e, o), the first to fifth function memories 3 of the function memory 3
(1) [fij], -, ( _{5) [fij] are stored in the order of A, I, C, E, and O in 1, . . . , and 35} , respectively. 4 is a probability calculation circuit, which calculates characteristic parameters of the input speech obtained from the parameter extraction circuit 2 based on each distribution function _(o) [fij], n=1, 2, 3, 4, 5 of the function memory 3; _(o) Derive the existence probabilities [ _(o) fij (xij)] for the five vowels of [xij], n=1, 2, 3, 4, 5, and output these. 5 is a probability memory that stores these existence probabilities [ _(o) fij(xij)], and the first to fifth probability memories 51, to, 55 each store [ _(o) fij(xij)], n =1, 2, 3, 4, 5 are stored. Reference numeral 6 denotes a recognition speech determination circuit, which calculates the sum of the components of each existence probability [ _(o) fij (xij)] in the probability memory 5, that is, d(n)= ₈ 〓 ⁱ⁼¹ ₁₆ 〓 ^j=1 _(o) fij(xij) n = 1, 2, 3, 4, 5, and by detecting n when d(n) is maximum, the input speech at this time is the nth recognized speech. It is determined that That is, if n=4, it is determined that (d).

FIG. 2 shows the distribution function _{(o) f IJ} showing the existence probability of the frequency spectrum value of a specific sample (j=J) at a frequency of 1KHz (i= _I ) for the five vowels stored in the function memory 3. ( _xIJ ) is illustrated. Based on the same figure, if we pay attention to the frequency spectrum value of the characteristic parameter of the input voice x〓 _{IJ , we can see that} this
₎ x _IJ , _-(4) x _IJ , _-(5) x _IJ have the same error, but
Depending on the function memory 3 and the probability calculation circuit 4 described above, the respective probabilities ₍₂₎ f _IJ (x〓 _IJ ), ₍₄₎ f _IJ (x〓 _IJ ), ₍₅₎ f _IJ (x〓 _IJ ) As is clear from the figure, ₍₄₎ f _IJ (x〓 _IJ ) is the largest and has the highest probability of belonging to the vowel (E). In this way, in the conventional device shown in Fig. 1, the recognition rate can be improved to some extent by performing recognition processing that takes into account the degree of distribution of frequency spectrum values, which are characteristic parameters of speech. ing.

However, in Figure 2, if we pay attention to the frequency spectrum value x _{¨ IJ} , this x ¨ _IJ belongs to the vowel (a), and its probability ₍₁₎ f _IJ (x〓 _IJ ) is The probability that x〓 _IJ belongs to the vowel (D) ₍₄₎ f _IJ (x〓 _IJ ) is equal to the probability that x〓 _IJ belongs to the vowel (E), but in reality,
),
Compared to the fact that there is almost no probability of belonging to (d) or (e),
x¨ It can be seen that there is a sufficient probability that _IJ belongs to vowels (i) and (o) other than vowel (e). In this way, the probability that x¨ _IJ , which is important for identifying speech, belongs to the vowel (A) ₍₁₎ f _IJ (x¨ _IJ ) and x _〓, which is important for identifying speech, belongs to the vowel (E). The probability of belonging ₍₄₎ f _IJ (x〓 _IJ ) is treated equally, that is,
There is a disadvantage in directly treating the probability depending on the distribution function of each recognized voice as a similarity score.
Due to this inconvenience, conventional speech recognition devices
It was not possible to expect a significant improvement in the recognition rate.

The present invention has been made for the purpose of solving the above-mentioned disadvantages, and will be described in detail below.

FIG. 3 shows an embodiment of the speech recognition device of the present invention. In the figure, 1 and 2 indicate a microphone and a parameter extraction circuit as in Figure 1, and the frequency spectrum value [xij] of the characteristic parameter of the voice input to the microphone 1 is extracted as a parameter. obtained by circuit 2.

The difference between the device according to the present invention and the conventional device shown in FIG. 1 is that the distribution function _(o) [fij] (hereinafter referred to as n
= 1, 2, 3, 4, 5), instead of the function memory 3 that stores the existence probability calculated based on a specific distribution function _(o) [fij] and other distribution functions _(o) A weighted distribution function _{[(o) gij] in which the relative ratio (o)} [gij] with the existence probability calculated based on [fij] is weighted to the specific distribution function _{( o) [} fij]. _(o) fij], n=1,
2, 3, 4, and 5 are provided in the first and fifth weighting function memories 71 to 75, respectively. The point is that the weighting function memory 7 is provided. The weighting distribution function stored in the weighting function memory 7 [ _(o) gij・_(o) fij]n=1,
Specifically, 2, 3, 4, and 5 are calculated as follows. That is, based on the distribution function _(o) [fij] of the frequency spectrum value of each vowel (a, i, u, e, o),
The probability at each average value of the frequency spectrum value that maximizes each of these distribution functions _-(1) xij, ~, _-(o) xij [ ₍₁₎ fij ( _-(o) xij)], ~, [ ₍₅₎ fij ( _-(o) xij)], n=1, 2, 3, 4,
5, and further calculate the relative ratio between these existence probabilities. By calculating n = 1, 2, 3, 4, 5 and multiplying these ratios _(o) gij by the weighting coefficient and each of the above distribution functions _(o) fij, the weighting distribution function _(o) gij・_(o) fij is obtained. Reference numeral 8 denotes a similarity calculation circuit, which calculates the characteristic parameters [xij] of the input speech obtained from the parameter extraction circuit 2 based on each weighting distribution function [ _(o) gij・_(o) fij] in the weighting function memory 7. 5 vowels (a, i, u, e,
The weighted existence probability [ _(o) gij・_(o) fij( _-(1) xij)] as the similarity for e) is derived and output. 9 is a similarity memory that stores these similarity values [ _(o) gij・_(o) fij(xij)], and each value is stored in each of the first to fifth similarity memories 91, -, 95. [ _(o) gij・_(o) fij(xij)],
n=1, 2, 3, 4, 5 are stored. 6' is a recognized voice determining circuit which operates in the same manner as the conventional device shown in _FIG .
j・_(o) fij(xij)], that is, d(n)= ₈ 〓 ⁱ⁼¹ ₁₆ 〓 ^j=1 _(o) gij・_(o) fij(xij) n=1, 2, 3 , 4, 5 and detecting n when d'(n) is maximum, the input speech at this time is determined to be the n-th recognized speech. That is, if n=2, it is determined that (a).

In the speech recognition device having such a configuration, the similarity calculation circuit 8 calculates the input speech based on the weighting distribution function [ _(o) gij・_(o) fij] stored in advance in the weighting function memory 7. The probability of existence of each recognized voice is weighted with respect to the feature parameters of .
For example, if we look at the frequency spectrum value x〓 _IJ in Figure 2, we can see that this x〓 _IJ belongs to the vowel (E) with probability ₍₄₎ f _IJ
For (x〓 _IJ ), the average value of the frequency spectrum value of this vowel (E) - ₍₄₎ x Each vowel in _IJ (A, I, U,
E, E) existence probability _(o) f _IJ ( _-(4) x _IJ ), n=1, 2, 3,
Since the ratio of the values of 4 and 5 is approximately 0:0:0:7:3, the ratio of the existence probability for this vowel (E) is
₍₄₎ g _IJ will be approximately 70%. On the other hand, in the case of the frequency spectrum value x¨ _IJ , as is clear from the figure, the vowel
Ratio of existence probability in (a) ₍₁₎ g _IJ is approximately 100%.
Therefore, the similarity calculation circuit 8 calculates the similarity of x〓 _IJ to the vowel (D) {0.7 ₍₄₎ f _IJ (x〓 _IJ )} and x ¨ The similarity of _IJ to the vowel (A) { 1.0 ₍₁₎ f _IJ (x¨ _IJ )} is obtained, and a lower weight is given to the existence probability where the existence probabilities of the feature parameters of each recognized voice overlap and the recognition of the recognized voice becomes ambiguous. Based on the similarity obtained through effective weighting in this manner, the recognized speech determining circuit 6' determines the recognized speech that has the highest degree of similarity to the input speech.

In the above description, the frequency spectrum value was exemplified as the voice characteristic parameter, but the autocorrelation coefficient of the voice waveform or the formant frequency may also be used. Although five vowels were treated as the recognized speech to simplify the explanation, it is also possible to perform recognition processing on a specific number of recognized speech, for example, 50 sounds.

As is clear from the above description, the speech recognition device of the present invention has a plurality of feature parameters x of the same recognized speech.
Find the distribution function of this specific distribution function f
and a weighting function memory storing a weighted distribution function g·f obtained by weighting the relative ratio g between the existence probability calculated based on the distribution function f and the existence probability calculated based on another distribution function f. Equipped with
Based on the distribution functions g and f of the memory, the similarity calculation circuit obtains a weighted similarity for each recognized voice with respect to the characteristic parameters of the input voice. Among the probabilities, those with low importance for identifying recognized speech can be given less weight with respect to their existence probabilities. Therefore, it is possible to recognize input speech using the existence probability of recognized speech, which is highly important for identifying recognized speech, and it is possible to recognize the fluctuating components of the speech not only for specific speakers but also for unspecified speakers. It is possible to sufficiently compensate for this and improve the recognition rate.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing the configuration of a conventional speech recognition device, Fig. 2 is a distribution function curve diagram showing the existence probability of frequency spectrum values of speech, and Fig. 3 shows an embodiment of the speech recognition device of the present invention. A block diagram is shown. 2... Parameter extraction circuit, 6'... Recognized speech determining circuit, 7... Weighting function memory, 8... Similarity calculation circuit, 9... Similarity memory.

Claims (1)

[Claims] 1. A parameter extraction circuit that extracts the characteristic parameter x of the voice from the input voice, and a distribution function f of the characteristic parameter It is obtained by weighting the distribution function f of a specific recognized voice and calculating the relative ratio g between the existence probability calculated based on the distribution function f and the existence probability calculated based on the distribution function f of other recognized voices. Based on a weighting function memory that stores weighting distribution functions g and f in advance, and weighting distribution functions g and f of each voice in the function memory, each of the characteristic parameters x of the input voice obtained from the parameter extraction circuit is Weighted existence probability g・f(x) for recognized speech
What is claimed is: 1. A speech recognition device comprising a similarity calculation circuit that calculates a similarity as a similarity, and recognizes the recognized speech with the maximum similarity obtained from the similarity calculation circuit as the input speech at this time.