JPS592040B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention is an improved speech recognition device, in particular, that recognizes the original word by segmenting word speech into units such as phonemes and syllables and then recognizing each segment by DP matching. The present invention relates to a voice recognition device.

Conventionally, the following two methods have been used as general methods for recognizing word sounds.
There is one.

That is, (a) a method using words as a unit of recognition;

In this method, the word itself is stored as a standard pattern, and recognition is performed by matching the input pattern and the standard pattern. (b) A method that uses basic units that make up speech, such as phonemes and syllables (hereinafter referred to as speech units), as recognition units.

In this method, input speech is segmented into speech units, and each segment is then recognized to recognize the original word. Furthermore, this method requires a word dictionary in addition to a standard pattern dictionary of phonetic units. Method (a) above requires as many standard patterns as there are words to be recognized, but method (b) requires as many standard patterns as there are phonetic units, and all you have to do is change the contents of the word dictionary. All words can be recognized.

Therefore, method b) is advantageous when there are many changes in the words to be recognized or when the number of words is large. The present invention relates to a speech recognition device based on the above method. In method (b), it is easier to recognize segments by using stable phonetic units that are less susceptible to articulatory combinations, rather than using phonemes as phonetic units.
It is advantageous to use longer units as vocal units.

A method for performing word recognition using this method is to use VCV syllables consisting of one vowel, one consonant, and one vowel as units of speech, as described in ``Nakatsu Koda's Word Speech Recognition Using VCV Syllables 2 Communication Society Pattern Recognition and Learning study group material PRL7
363 (1973-09).
In this method, in order to recognize each segmented VCV syllable, a standard VCV syllable pattern stored as a time-series pattern and an input pattern are matched by nonlinear matching (hereinafter referred to as DP matching) using dynamic programming. This is done using . As shown in Figure 1, this matches the standard pattern and the input pattern by matching their starting and ending points and allowing non-linear expansion and contraction, which corrects the non-linear expansion and contraction of the audio pattern in the time axis direction. This has the advantage that correct matching can be performed. The above matching method is one of the effective methods for speech recognition. However, in speech patterns, the expansion and contraction of the constant vowel part is large, but the expansion and contraction of the consonant part is not so large, and the conventional matching method described above has the disadvantage that it is insufficient to correct such deformation. It was hot. For example, standard patterns of syllables are usually created from singly uttered syllables and therefore contain fairly long vowel stations. In contrast, syllables in continuous speech often contain almost no vowel stationary parts. In this way, the conventional method of matching patterns with different properties using DP matching with matching endpoints has the drawback that errors tend to occur in segment recognition, making it difficult to recognize words with sufficient accuracy. It was hot. In order to eliminate these drawbacks, the present invention improves the segment recognition method and, when associating an input pattern with a standard pattern, does not necessarily match the end points, but instead matches the input pattern with the standard pattern as shown in Figure 2. The purpose is to obtain a word recognition device that can obtain a high recognition rate by using a DP matching method that associates parts of words.

The drawings will be explained in detail below. FIG. 3 shows an embodiment of the invention.

In the following explanation, the characteristic parameters of input speech are [Itakura, Saito0 Estimation of Speech Spectral Density and Formant Frequency Using Statistical Methods 7 Journal of the Institute of Electronics and Communication Engineers (
A) The autocorrelation coefficient of the speech waveform is extracted based on the maximum likelihood spectral analysis method proposed in ``53-A, 1, P35 (Showa 45-01)''. It is also assumed that the standard pattern is registered in the form of maximum likelihood spectral parameters. Furthermore, as a measure of the similarity between the input and the standard pattern, we use the likelihood, which is also defined in the above literature and is expressed as the logarithm of the sum of products of the autocorrelation coefficient and the maximum likelihood spectral parameter with a negative sign. . Furthermore, a VCV syllable consisting of one vowel, one consonant, and one vowel is assumed as a vocal unit. Third
In the figure, 1 is a feature extraction unit that performs maximum likelihood spectrum analysis to extract features of input speech. 2 is a segmentation unit that segments input speech into VCV syllable units.

The segmentation method used here may be one that has already been published. For example, a method for segmenting VCV syllables is "Nakatsu, Koda's Segmentation of Continuous Speech and Conversion to Phonological Sequences 2 Acoustical Society of Japan. The one proposed in "Speech Study Group Material S74-24 (1974-12)" may be used. Therefore, details are omitted here. 3 is a control section which controls the operation of the apparatus by issuing various control signals at appropriate times.

a indicates the word to be recognized). number (or its word), and sequentially indicates the numbered words to be recognized.
b is a signal indicating the number of the segment of the word indicated by signal a. Assuming that a word a is made up of b(a) segments, control signals (a, b) are sent out from the control unit in the following order.

To recognize input speech, first send a reset signal rl to the word likelihood addition register T1, which will be described later, and then input signals a and b.
The word dictionary storage unit 5b (described later) and the reading circuit 4 (described later)
Send to 2. Next, a likelihood calculation start signal s is sent to a likelihood calculation control section 60 (described later), and upon receiving an end signal e, a signal a is sent to a segment likelihood maximum value selection circuit 70. Next, another set of a and b is sent to the word dictionary storage section 50 and the reading circuit 42. The same operation is repeated thereafter, and when all pairs of a and b have been sent, an end signal z is sent to the maximum value selection circuit 8. With the above operations, recognition of one word is completed. The operation of each part will be described in detail below. Block 4 surrounded by a broken line is a buffer of the input pattern. The input pattern correlation coefficient buffer 40 has input audio as a time series of correlation coefficients, /pi=(ρoi
．． /pl i,・・・・・・・・・/ppi),p
It is stored in the form of ol=1.

Further, 41 is a register in which data of segment boundaries extracted by the segmentation unit is stored. Reference numeral 42 denotes a readout circuit, which, when a signal b indicating a segment number is sent from the control unit 3, refers to data at segment boundaries and outputs a time series of correlation coefficients of the corresponding segment.

Further, when all the contents of the buffer register 40 have been read out, a read end signal f is sent to the control section 3. The block indicated by the dashed line 5 is a storage section for a word dictionary and a VCV syllable standard pattern.

50 is a word dictionary storage unit in which words to be recognized are stored as a series of VCV syllables.

For example, the word / asahi / (morning sun) is / a
A message is stored consisting of two VCV syllable sequences: sa / and / ahi /. The word dictionary storage section instructs the corresponding VCV syllable to a vCV syllable standard pattern storage section 51, which will be described later, by signals a and b sent from the control section. 51 is a VCV syllable standard pattern storage unit, which stores all VCs appearing in the word to be recognized.
A standard pattern of V syllables is stored as a time series of maximum likelihood spectral parameters.

When a VCV syllable is specified from the word dictionary storage section, a time series of the corresponding standard pattern is output. A block 6 surrounded by a broken line 6 is a segment likelihood calculation unit, which calculates the likelihood between the input speech segment pattern and the VCv syllable standard pattern.

60 is a likelihood calculation control unit which, upon receiving the calculation start signal s from the control unit 3, sends control signals to each unit to control the likelihood calculation, and when finished, sends an end signal e to the control unit.

61 is a buffer of input VCV syllables, and readout circuit 4
Store the time series of correlation coefficients sent from 2.

62 is a buffer for VCV syllable standard patterns, in which the vCv syllable standard patterns sent from the storage section 51 are stored.

The time series of the inputs stored in buffer 61 is now /pi=(/poi,.pli)...
・, /p pi), /pol=1, and the time series of the standard pattern stored in 62 is, however, Aj=(Ao
j, Alj, ゜゜゜゜゜゜゜゜゜゜゜, Apj).

63 is a likelihood calculation circuit, which calculates the input ``i'' and the standard pattern Aj by the control circuits i, j sent from the
It is possible to store the likelihood* of (1, .j) in separate registers and use the output.

The register 64 first receives a set signal R from the control section 60.
After the internal state is set by 2, it is sent from the calculation circuit 63. * As shown in Figure 4, this is l(1
．． This corresponds to finding the path with the maximum likelihood sum from the first column to the nth column on the likelihood matrix created by j) (this point will be described later with reference to FIG. 5).

Previously, this type of method was used instead of the (support) formula condition. This means, as shown in FIG. 1, that the input pattern and the standard pattern are matched with each other by matching their end points. On the other hand, in the device of the present invention, as shown in FIG. 2, the input pattern is associated with a part of the standard pattern. Correct correspondence will be made even if the pattern does not include a vowel stationary part. Equation (7) can be expressed as a recurrence expression as follows.

, Y Therefore, the dynamic programming calculation circuit 65 receives the control signals 1, j
According to (1-
The operation of reading the values of 1, j), (1-1, j-1), (1-1, j-2) and adding the maximum value to register (1.j) is based on the time series of the input pattern. are P1, P2, P3
, /P4, and the time series of one standard pattern is given by A1, A2C...A7, and will be explained with reference to FIGS. 5A to 5G. be done.

Likelihood 1 (1.j) according to equation (5) above is 4X7
Calculated for each piece.

Next, the likelihood sum of the paths from the first column to the third column, that is, the fifth
The sum of the likelihoods for each of the 27 paths as shown in Figure A, the sum of the likelihoods for each of the 26 paths as shown in Figure 5B, etc. The sum of likelihoods along one path as shown in FIG. 5G is calculated to find a path with the maximum likelihood sum, for example, a path as shown by the bold line in FIG. 5B. In this way, the path with the maximum likelihood sum for one standard pattern can be found, or this likelihood sum is similarly compared with the likelihood sum of the path with the maximum likelihood sum found for each standard pattern. go.

Block 7 surrounded by a broken line is a word likelihood calculation section.
70 is a segment likelihood maximum value selection circuit, and the control unit 3
When signal a is received from , Equation G3) is calculated, and the result is added to a corresponding register of word likelihood addition register 71, which will be described later.

71 is a word likelihood addition register, which has registers equal to the number of words to be recognized.

Then, the value sent from the control unit 3 at the start of the recognition process or set by the set signal r1 and then sent from the maximum likelihood selection circuit 70 to the segment is added to the word likelihood addition register 71. . That is, the register 7
The value of each register 1 is the likelihood between the input speech and each word to be recognized, defined as the sum of the likelihoods for each segment. When calculating the likelihood between the input speech and word a in the dictionary,
If the numbers of segments do not match, the input speech cannot be word a. Therefore, at the time when the control section 3 sends out the signal (A.b(a)) instructing the last segment of word a, it checks that the reading end signal f is sent. Otherwise, i.e. the signal (A.
b(a)) but does not receive the signal f. If the signal f is received when the signal (A. Clear. When the likelihood of the input speech and all words in the word dictionary is calculated and the result is stored in the register 71, a signal Z is sent from the control unit 3, and in response to this, the maximum value selection circuit 8 adds the word likelihood. The maximum value of the register 71 is determined, and the word name and maximum likelihood are output as recognition results.

As explained above, the present invention provides a speech recognition device that segments word speech into speech units and recognizes each segment to recognize the original word. When performing matching, performing DP matching without fixing the endpoints has the advantage that segment recognition can be performed with high accuracy and, as a result, a high word recognition rate can be obtained.

In the above embodiment, the autocorrelation coefficient was used as the voice feature parameter, but even if this is replaced with another parameter, if another measure of similarity is adopted, the device can make partial changes. It is clear that it can be used by Furthermore, in the above embodiments, the VC syllable was explained as an example of a vocal unit, but it is clear that other units such as a VC syllable consisting of a vowel and one consonant, a CV syllable consisting of a consonant and one vowel, etc., may also be used. These are also included within the scope of the present invention.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing how input patterns and standard patterns are matched in conventional DP matching, FIG. 2 is an explanatory diagram showing how input patterns and standard patterns are matched in the present invention, and FIG. 3 is a diagram showing how input patterns and standard patterns are matched in the present invention. A block diagram showing the configuration of an embodiment of the present invention, FIG. 4 is an explanatory diagram showing how DP matching is performed between an input pattern and a standard pattern, and FIGS. 5 A to G are for finding the path with the maximum sum of likelihoods. An explanatory diagram illustrating the path of is shown. In the figure, 1 is a feature extraction unit, 2 is a segmentation unit, 3 is a control unit, 4 is an input pattern buffer, 40 is an input pattern correlation coefficient buffer, 41 is a segment boundary register, 42 is a VCV syllable reading circuit, 5 50 is a dictionary storage unit, 50 is a word dictionary storage unit, 51 is a VCV syllable standard pattern storage unit, 6 is a segment likelihood calculation unit, 60 is a likelihood calculation control unit, 61 is an input VCV syllable buffer, and 62 is a VC syllable standard pattern. 63 is a likelihood calculation circuit, 64 is a likelihood addition register, 65 is a dynamic programming calculation circuit, r is a word likelihood calculation unit, 70
71 represents a segment likelihood maximum value selection circuit, 71 represents a word likelihood addition register, and 8 represents a maximum value selection circuit.

Claims (1)

[Claims] 1. An input pattern buffer that stores feature parameters of input speech and values of segment boundaries, a word dictionary storage unit that stores recognition targets as symbol strings of segments, and a time series pattern of all segments. When matching a segment of an input voice with a segment of a standard pattern with a dictionary storage unit consisting of a standard pattern storage unit, the end points of the two do not necessarily match and non-linear expansion/contraction is allowed. Word similarity that stores the sum of the similarities between each segment candidate word of the input speech and the corresponding segment as the similarity between the input speech and the candidate word. What is claimed is: 1. A speech recognition device, comprising: a degree calculating section, and outputs a candidate word giving the maximum value of the degree of similarity as a recognition result. 2 The segment similarity calculation unit calculates the element similarity l(
i, j) (i=1, 2,...n; j=1, 2,...
...... m), the sum of n element similarities starting from element similarity l(1, 1), and n elements starting from element similarity l(1, 2). 2. The speech recognition device according to claim 1, further comprising a processing circuit section that calculates the sum of similarities. 3 In calculating the above element similarity sum, V(1,
j)=l(1,j)(j=1,2,...m)V(
i, j)=1(i, j)+max{V(i-1),V(
i-1, j-1), V(i-1, j-2)}, and obtain the sum of element similarities. The speech recognition device according to scope 2.