JPH0342480B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field The present invention relates to a pattern comparison method for comparing voice patterns expressed as feature amounts on a two-dimensional plane and determining the degree of similarity between the patterns.

BACKGROUND ART In recent years, various devices have been considered, such as speech recognition devices, that calculate pattern similarity or something similar thereto and select recognition results accordingly. However, when dealing with speech, there are two problems in calculating the similarity of such patterns. One is that the same word phonetic patterns have different durations due to differences in pronunciation speed, making it impossible to directly compare the patterns and calculate the degree of similarity.The other is that the formant frequency changes when the speaker changes, so This means that the difference becomes larger. Currently, the most widely used method for the former is pattern matching using dynamic programming (DP). DP matching will be briefly explained with reference to FIG. This is a case where the similarity is determined by matching the starting and ending ends of patterns f(t) and g(t) and allowing non-linear expansion/contraction in the time axis direction. f(t) and g(t) are discrete quantities sampled at regular intervals, and each sample point is expressed as m ₁ , m ₂ ,
...m _N , n ₁ , n ₂ , ...n _N , the two patterns are f (m ₁ ), f (m ₂ ), ... f (m _N ), g (n ₁ ), g
(n ₂ ),...g(n _N ). Starting points f(m ₁ ) and g(n ₁ ) of the pattern, and ending points f(m _N ) and g(n _N )
are associated with each other, and other points are associated so that the distance between both patterns is minimized. To do this, one point (mi) in f(m) is associated with all points in the vicinity of g(ni), and the point that minimizes the distance is selected from among them and associated. As a result, a slope as shown by A in FIG. 1 is obtained, and the degree of similarity can be calculated by mapping f(t) to g(t) according to this slope. However, this method has the disadvantage that it requires a very large amount of calculations, and although it can absorb variations in the time length of the pattern, it cannot absorb variations in frequency.

In this way, when the audio pattern on the two-dimensional plane formed by the frequency axis and the time axis has fluctuations with respect to both axes, there is no conventional method that can absorb this with a small amount of calculation. Furthermore, Japanese Patent Application Laid-open No. 116185/1983 discloses a method of absorbing and comparing slight distortions in patterns. In this method, the degree of coincidence between a reference pattern and a test pattern is compared for the purpose of pattern checking on printed circuit boards and the like. First, a reference pattern is moved by a certain distance, and a large pattern and a small pattern are created by calculating the logical sum and logical product between the patterns before and after the movement. The moving distance at this time is determined by calculating the AND and OR of the shape of the pattern to be checked and the pattern to determine whether the pattern on the substrate is normal or abnormal. However, although this method is effective for slight distortions such as substrate patterns, it is inappropriate for pattern matching in voice recognition for the following reasons.

(1) In voice recognition, the range of variation is larger than in circuit board patterns. Therefore, the internal shape of the pattern differs depending on the logical sum and product. If the internal shape of the pattern changes using the method described in the above-mentioned publication, not only will there be no effect of absorbing distortion, but it will also become impossible to compare the patterns.

(2) The method described in the above-mentioned publication is suitable for binary graphic processing with or without a pattern such as a board, so it cannot be used for speech recognition patterns unless they are binarized. Therefore, since the amount of information in the binary speech pattern decreases, the number of words to be recognized is limited to a small vocabulary.

(3) Even if a correction is made using the method described in the above publication with a rule that when a multi-value pattern is input, it will be treated as a binary value above and below a predetermined value, the result shown in Figure 13 would be In such a case,
Pattern (a) will have the same degree of similarity in both (b) and (c), making them indistinguishable. In other words, although the described method can check whether a pattern is normal or abnormal, it is not suitable for detailed judgments such as determining which of several standard patterns it most closely resembles.

(4) When a pattern is modified by logical sum or product, not only the time direction of the audio pattern but also the direction perpendicular to it, and in many cases, the number of data on the frequency axis also changes, making it impossible to handle.

As described above, the conventional method described in the above publication cannot be applied to voice pattern matching.

Purpose The present invention was made in view of the above circumstances, and
The present invention aims to provide a speech pattern comparison method that can absorb variations in speech patterns on both axes on a two-dimensional plane formed by the frequency axis and the time axis, and can compare speech patterns with a small amount of calculation. .

Configuration The configuration of the present invention will be described below based on examples.

Figure 2 shows the pattern of the word "size" uttered by a certain speaker. In this figure, the horizontal axis represents frequency, and the vertical axis represents time, and the spectral distribution at the time of occurrence is expressed as "size" in shading; the blacker it appears, the higher the level. The frequency increases from left to right,
250Hz to 6.3kHz is divided into 15 equal logarithmic intervals.
FIG. 3 shows an example in which the same speaker utters the same word on different occasions. As is clear from the figure, both have different lengths in the time axis direction.

Formants are what characterizes the sounds we make. Or the concept of local peaks in the spectrum [Journal of the Acoustical Society of Japan, Vol. 32, No. 1 (1976) No. 12~
[Page 23] may be used, but in any case, in order to produce language, we change the shape of the sound path, and this effect appears as local peaks on the speech spectrum. Therefore, the characteristics of the spoken language appear in the temporal change of such local peaks. Therefore, the time-frequency pattern (hereinafter referred to as time-spectrum) that represents the temporal change of local peaks is
Consider recognizing uttered language by comparing patterns (abbreviated as TSP). Second
Both TSPs shown in Figures and Figure 3 are 10~
15ms is /s/, next 100ms is /a/, etc.
The pattern is /i/ for a little less than 10 ms, followed by /Z/ for several ms, and short at the end, /u/. Incidentally, in addition to changes in time length as shown in the figure, differences in the number of occurrences appear as changes in peak frequency, but neither of these is extreme. Therefore, when comparing two patterns, one pattern is kept wide, taking into account the width of frequency fluctuation and time fluctuation, and the other pattern is created by extracting line features from a wide line figure. Verification is performed after extracting a point or center line near the center of the width using a thinning method, which is one of the methods. At this time, it is desirable to narrow the width in the time axis direction as well. The term "pattern width" in this case refers to a figure in a pattern, or rather a pattern in a pattern, and does not mean changing the size of the entire pattern. By doing this, even if both ends of time and frequency of one pattern fluctuate, the thin line pattern can be matched without protruding from the wider pattern. That is, a standard speech pattern and an input speech pattern are formed on a two-dimensional plane of a time series and an acoustic feature sequence (frequency sequence of audio waves) corresponding to the time series, and the linguistic information generated in the pattern is The linguistic information is formed along a string of linguistic information (a string of local peaks appearing in a spectrum envelope based on short-time spectrum analysis of speech waves) that carries Feature width and/or
Alternatively, the time width formed in the time series direction corresponding to each feature sequence is made different between the standard pattern and the audio pattern, and the size of the overlapping part of the pattern generated by superimposing these two patterns is This method determines the similarity of voice patterns.

FIG. 4 is a hardware configuration diagram of the present invention, and FIG. 5 is a hardware configuration diagram of the present invention.
This is a flowchart for explaining the operation of the configuration diagram shown in FIG.
is generated into a two-dimensional pattern by the two-dimensional pattern generation processing section 11. The two-dimensional pattern stored in the two-dimensional pattern storage unit 13 is a pattern for comparing the degree of similarity with the two-dimensional pattern of data, and will be referred to as a standard pattern. This standard pattern may be prepared and stored in advance, or a two-dimensional pattern of data may be imported and formed into a standard pattern. Prior to comparing the two-dimensional pattern of these data and the standard two-dimensional pattern, pattern width conversion processing is performed as necessary in a pattern width conversion processing section 12. In addition,
The pattern width conversion process is to make the pattern widths of the data two-dimensional pattern and the standard two-dimensional pattern different in each dimensional axis direction corresponding to both the two-dimensional patterns (Step 2). By performing such processing, a difference in width is generated between the pattern width appearing on each dimensional axis in one pattern and the pattern width appearing on each dimensional axis in the other pattern. The matching processing unit 14 overlaps both patterns having different pattern widths formed as described above (step 3), calculates the size of the overlap (step 4),
This is output as a degree of similarity.

Next, an embodiment of the pattern comparison device of the present invention is shown in FIG.

In FIG. 6, an audio signal input from a microphone 1 passes through a filter bank 2 and becomes a frequency-time pattern. A voice section cutout section 3 cuts out a voice portion from the voice section, and a binarization section 4 binarizes it by setting a certain threshold value. This binarization is for reducing the amount of information, and of course binarization is not necessary. This is stored in the dictionary section 6 by the line thinning section 5 as a point or center line that appears to be approximately at the center.
Next, the switch 7 is set to the matching unit 8 side, and after the frequency-time pattern of the input voice is binarized, the dictionary unit 6
When compared with each word stored in , that is, when the two patterns are overlapped, the extent to which the thinning patterns overlap is determined and the degree of similarity is calculated. This comparison is performed for each pattern stored in the dictionary section, and the word with the highest degree of similarity is set as recognition result 9. As an example, FIG. 7 shows a thinned version of the pattern shown in FIG. 3, and FIG. 8 shows a binarized version of the pattern shown in FIG. Various thinning processes can be considered here (for example, see Institute of Electronics and Communication Engineers study group material PRL-75-66, pp. 49-56), but basically the points touching the boundaries of binary figures are connected to each other. Delete the points one by one while preserving the properties, and take out the points near the center or the center line.

Note that in the above explanation, one of the patterns has been made into a thin line, but this is to ensure matching without overlapping when comparing the two patterns, so one may be made into a thick line while retaining the characteristics of the line shape. Alternatively, one may be made into a thin line and the other may be made into a thick line.

FIG. 9 is a diagram showing another embodiment of the present invention in which one of two patterns is made into a thick line and superimposed on the other; contrary to the case of FIG. 6, no processing is applied to the pattern to be registered in the dictionary. , the pattern to be recognized is subjected to thick line cutting processing by the thick line making section 10.

FIG. 10 is a diagram showing another embodiment of the present invention that performs both thinning and thickening, in which the dictionary registration pattern is thinned and the recognition pattern is thickened, but of course this can be reversed. Also good. Note that although the above description has been made using a voice pattern as an example, it is clear that the present invention is not limited to voice patterns and may be applied to other two-dimensional patterns. .

FIG. 11 is a flowchart for explaining one embodiment of the present invention, and this embodiment involves importing a two-dimensional pattern of data as shown (step 1).
→2→3) This is the case of forming this as a standard two-dimensional pattern (step 4). In general,
The two-dimensional pattern of the data is compared with the formed standard two-dimensional pattern (step 6), and the pattern width of the former in each dimension axis direction is determined to be different from the latter pattern in the corresponding dimension axis direction. Go (step 1 → 2 → 3 → 5). However, if the conversion tendency of the width of the data pattern in each dimensional axis direction is always in one direction, either thinner or thicker than the corresponding width of the standard pattern, then the corresponding width of both patterns There is no need to make a comparative judgment (step 6) as to whether or not they are different.

FIG. 12 is a flowchart for explaining another embodiment of the present invention. This embodiment also captures a two-dimensional pattern of data as shown in the figure.
(Step 1 → 2 → 3) This is the case where this is formed as a standard pattern (Step 6). This point is the same as the flowchart shown in FIG. 11. What is different from the flowchart shown in FIG. When forming a standard two-dimensional pattern, while comparing it with the two-dimensional pattern (self) imported as data (step 5), the pattern width in each dimension axis direction of the former is compared with that of each corresponding dimension axis of the latter. The patterns are determined differently (step 1 → 2 →
3→4→5→6→3) points. However, even in this case, if the conversion method is such that the width of the standard pattern formed in each dimensional axis direction is always thinner or thicker than the corresponding width of the data pattern, then There is no need to compare and judge whether the widths of both patterns are different (step 5).

Effects As described above, according to the present invention, patterns can be compared by absorbing fluctuations in each axis of a two-dimensional pattern, such as fluctuations in the time axis of word length during utterance and frequency fluctuations due to the speaker. is possible.

[Brief explanation of drawings]

Figure 1 is an explanatory diagram of DP matching, Figures 2 and 3 are diagrams showing time-frequency patterns, Figure 4 is a hardware configuration diagram of the present invention, and Figure 5 is a configuration diagram of the configuration diagram in Figure 4. Flowchart to explain the operation,
FIG. 6 is a diagram showing an embodiment of the pattern comparison device according to the present invention, FIG. 7 is a diagram in which the pattern in FIG. 3 is thinned, and FIG. 8 is a diagram in which the pattern in FIG. 2 is binarized. 9 and 10 are diagrams showing other embodiments of the pattern comparison device according to the present invention, and FIGS. 11 and 12 are flowcharts for explaining the embodiment of the present invention, respectively. FIG. 13 is a diagram for explaining a conventional example. 1...Microphone, 2...Filter bank, 3...
...Voice section extraction unit, 4...Binarization unit, 5...
Thinning section, 6... Dictionary section, 7... Switch, 8...
... Collation section, 10... Thick line forming section, 11... Two-dimensional pattern generation processing section, 12... Pattern width conversion processing section, 13... Standard two-dimensional pattern storage section, 14
...Verification processing section.

Claims (1)

[Claims] 1. A standard speech pattern and an input speech pattern are formed on a two-dimensional plane of a time series and an acoustic feature sequence corresponding to the time series, and the linguistic information generated in the pattern is carried. A feature width formed along a column of linguistic information and including the linguistic information in the direction of the feature column corresponding to each time series and/or a feature width of the time series corresponding to each feature column. Regarding the time width formed in the direction, the similarity of the voice patterns is determined from the size of the overlapping part of the pattern generated by making either the standard pattern or the voice pattern thinner or thicker and superimposing it on the other pattern. A voice pattern comparison method characterized by determining. 2. Thinning the characteristic width and/or time width of either the standard pattern or the audio pattern;
2. The voice pattern comparison method according to claim 1, wherein the feature width and/or time width of the other pattern is made thicker and the two patterns are superimposed. 3. The acoustic feature sequence is a frequency sequence of a speech wave, and the linguistic information sequence carrying the linguistic information is a local peak sequence appearing in a spectrum envelope based on short-time spectrum analysis of the speech wave. A voice pattern comparison method according to claim 1 or 2. 4. The voice pattern comparison method according to claim 3, wherein the short-time spectrum analysis is an analysis using a group of band filters.