JPH0337199B2 - - Google Patents

Description

[Detailed Description of the Invention] [Summary] In a word speech recognition device that matches an unknown input word speech pattern with each word standard pattern, route restrictions are changed in accordance with the pattern length ratio of the input word speech pattern and the word standard pattern. Go there and check. This makes it possible to eliminate inconvenient time correspondences and improve the recognition rate without narrowing the range of expansion/contraction ratios.

[Industrial application field]

The present invention relates to a word speech recognition device that recognizes an input word speech by comparing an unknown input word speech pattern with each pre-registered word standard pattern.
The present invention relates to a word speech recognition device improved so as to improve the recognition rate by eliminating unnatural time correspondence during verification.

[Conventional technology]

When recognizing unknown input word sounds, the recognition rate is good, so the input word sound patterns created from the unknown input word sounds are compared with pre-registered standard patterns for each word to recognize the input word sounds. Word speech recognition methods that perform this are often used.

FIG. 5 is a block diagram showing the basic structure of a conventional word speech recognition method that matches word standard patterns.

In the word speech recognition method shown in FIG. 5, the speech analysis section 210 analyzes the input word speech, extracts parameters representing the characteristics of the input word speech, and detects intervals to create an input word speech pattern, and the word recognition section 2
Enter 20.

On the other hand, in the word standard pattern section 230, standard patterns created by analyzing each word of a word group to be recognized are registered in advance.

The word recognition unit 220 performs word recognition by comparing the input word audio pattern with the standard pattern of each word in the word standard pattern unit 230.

This word recognition is performed using the well-known DP method (Dynamic Pro
-gramming matching), and the word of the word standard pattern with the smallest distance from the input word speech pattern is set as the recognized word.

FIG. 6 shows the process performed in the word recognition unit 220.
This shows the DP matching method.

In FIG. 6, the horizontal axis is the input word speech pattern, and the numbers are the numbers of each vector forming the input word speech pattern. The vertical axis is the word standard pattern, and the numbers are the numbers of each vector forming the word standard pattern. The solid line is the time warping function (DP path) when the input word speech pattern is matched with the word standard pattern.

By the way, the duration of the word sound actually uttered is
The length of time varies depending on the speaker, and even with the same speaker, the length of time varies each time the speaker speaks. Since the input word speech pattern is uttered slowly or rapidly, the input word speech pattern expands and contracts with respect to the time length of the standard word pattern. Therefore, at the time of matching, the expansion and contraction in the time direction of the input word sound pattern is normalized (time normalization), and then the input word sound pattern is matched with the word standard pattern.

The range allowed as the expansion/contraction ratio of the input word speech pattern during this time normalization process is set to various values depending on the DP method, but for example,
It is often set in a range of 1/2 to 2 times, 1/3 to 3 times, etc. When the expansion/contraction ratio is set to 1/2 to 2 times, word sounds uttered at 1/2 to 2 times the normal time length can be recognized using the same word standard pattern. In this way, by increasing the expansion/contraction ratio, it becomes possible to recognize input word sounds having various time lengths using the same word standard pattern, and thus the recognition range can be expanded.

Increasing the width of the expansion/contraction ratio to 1/2 to 2 or 1/3 to 3 has the advantage of widening the recognition range as described above, but on the other hand, the following disadvantages occur. That is, when uttering one word, uttering the first half of the word quickly and uttering the second half slowly cannot normally occur in natural pronunciation. However, in the conventional DP method, pattern matching is allowed by stretching the first half by 2 to 3 times and shrinking the latter half to 1/2 to 1/3, so if matching is performed under such conditions, This will lead to misunderstandings.

Figure 6 shows that the input word speech is first uttered quickly (frames 1-2), then very slowly (frames 3-4), and then (frames 5-4).
10) This shows a case where misrecognition occurs as a result of matching performed under an unnatural condition in which utterances are uttered quickly on average.

[Problem that the invention seeks to solve]

Conventional word speech recognition methods have a wide range of expansion/contraction ratios for input word speech patterns during matching, making it possible to recognize input word speech with various time lengths (pattern lengths). Although this method has the advantage of being able to obtain a wide recognition range, it has the problem that unnatural time correspondences are allowed, resulting in erroneous recognition and lower recognition rates.

An object of the present invention is to provide a word speech recognition device that improves the recognition rate by preventing matching based on unnatural time correspondence without narrowing the range of expansion/contraction ratios allowed for the word as a whole. With the goal.

[Means for solving problems]

When uttering one word, the utterance time of some of the syllables forming the word changes, such as becoming longer or shorter, which normally does not occur in natural pronunciation. That is, when a word is uttered naturally, each syllable in the word is uttered at a normal rate, and when a word is uttered quickly, each syllable in the word is uttered at a normal rate. Each word is uttered quickly, and when a word is uttered slowly, each syllable is also uttered slowly.

Therefore, it is better to limit the DP path during matching when the word is uttered slowly and slowly in a natural state, to one whose slope is small on average, and when the word is uttered slowly and naturally. In this case, it is better to limit the DP paths during matching to those whose slopes are large on average.

Based on this knowledge, the present invention is capable of collation using unnatural temporal correspondence without narrowing the range of expansion/contraction ratios allowed for the word as a whole by appropriately changing the route restriction in accordance with the natural speaking speed of word speech. This is designed to improve the recognition rate by preventing this from occurring.

Hereinafter, the means taken by the present invention to solve the above-mentioned problems in the conventional word speech recognition system will be explained with reference to FIG.

FIG. 1 is a block diagram showing the basic configuration of the present invention.

In FIG. 1, reference numeral 110 denotes a pattern length ratio calculation means that calculates a pattern length ratio representing the pattern length of the input word speech pattern created from the input word speech and each word standard pattern. Here, the pattern length ratio is a quantity that determines the pattern length ratio of the input word speech pattern and the pattern length of the word standard pattern, and is expressed, for example, as the difference between the two pattern lengths or the ratio of the two pattern lengths. Further, the pattern length of each pattern is expressed by the number of frames, time length, etc. of each pattern.

Reference numeral 120 denotes a restricted route distance calculation means, which changes the route restriction at the time of matching in accordance with the magnitude of the pattern length ratio, and calculates the distance between the input word speech pattern and each word standard pattern.

[Effect]

When the input word speech pattern created from the input word speech is input, the pattern length ratio calculation means 11
0 calculates the pattern length ratio between the input word speech pattern and the word standard pattern. This calculation is performed for each word standard pattern.

The restricted route distance calculation means 120 changes the route restriction during matching in accordance with the magnitude of the pattern length ratio input from the pattern length ratio calculation means 110, and calculates the distance between the input word audio pattern and the word standard pattern. Calculate.

For example, when the pattern length ratio expressed as a ratio, that is, the value obtained by dividing the input word speech pattern length by the word standard pattern length, is larger than the vicinity of 1 (when it is uttered slowly), the path restriction during matching can be changed. When the pattern length ratio is smaller than the neighborhood of 1 (when uttered rapidly), the route restriction is changed to exclude paths with a slope of less than 1. will be changed to

As described above, by changing the route restriction in accordance with the natural speaking rate of the input word and performing matching, it is possible to perform matching using unnatural temporal correspondence without narrowing the range of expansion/contraction ratios allowed for the word as a whole. This makes it possible to prevent this from occurring, thereby improving the recognition rate.

〔Example〕

Embodiments of the present invention will be described with reference to FIGS. 2 to 4.

FIG. 2 is an explanatory diagram of the configuration block of one embodiment of the present invention, FIG. 3 is an explanatory diagram of route restriction in the same embodiment, and FIG.
The figure is an explanatory diagram of the DP matching method of the same embodiment.

(A) Configuration of the embodiment In FIG. 2, the pattern length ratio calculation means 11
0 and the restricted route distance calculation means 120 are as described in FIG.

130 is a microphone, and a speaker (not shown)
The word utterance uttered by is input.

140 is a speech analysis unit that analyzes the input word speech, extracts parameters representing the characteristics of the input word speech, detects sections, and creates input word speech patterns or word standard patterns.

150 is a switching circuit that performs switching corresponding to the word standard pattern registration operation and the input word speech pattern recognition operation.

160 is a word standard pattern dictionary in which each word standard pattern created by the speech analysis section is registered.

In the restricted route distance calculation means 120, 12
1 is an inter-vector distance calculation unit, which performs speech analysis 140
From this, the distance between each vector of the input word speech pattern and each vector of each word standard pattern in the word standard pattern dictionary 160 is calculated.

122 is a first DP calculation unit that calculates the distance between the input word speech pattern and each word standard pattern based on the first route restriction.

A second DP calculation unit 123 calculates the distance between the input word speech pattern and each word standard pattern based on the second route restriction.

A third DP calculation unit 124 calculates the distance between the input word speech pattern and each word standard pattern based on the third route restriction.

The specific contents of each of the first to third route restrictions will be explained in the next section "(B) Operation of the embodiment".

Reference numeral 125 denotes a route selection section, which selects each of the first to third DP calculation sections 122 to 124 based on the pattern length ratio inputted from the pattern length ratio calculation means 110.
Select one of them and output the calculation result.

170 is a word recognition unit that detects the minimum value among the distances on the restricted routes between the input word audio pattern input from the route selection unit 125 and each word standard pattern, and detects the minimum value among the distances for the restricted routes between the input word voice pattern input from the route selection unit 125 and the word standard pattern. Let the word be the recognized word.

(B) Operation of the embodiment The operation of the embodiment is explained with reference to FIGS. 2 to 4.
Each operation during input word speech recognition will be explained separately.

(B-1) Registration operation Before recognition processing is performed on the word sounds uttered by the speaker, a standard word pattern for each word to be recognized is registered.

When registering a word standard pattern in the word standard pattern dictionary 160, the switching circuit 150
is connected to the word standard pattern dictionary 160 side,
Word speech for registration is input to the speech analysis section 140 through the microphone 130.

The speech analysis unit 140 analyzes the input word speech, extracts parameters representing the characteristics of the word speech for registration, detects sections, creates a word standard pattern, and stores the word standard pattern dictionary 1.
Register at 60. Each word standard pattern is expressed as a time series of vectors representing the characteristics of each word standard pattern.

(B-2) Input word speech pattern creation operation When recognizing input word speech,
The switching circuit 150 is connected to the vector distance calculation unit 1
Connect to the 21 side.

When an unknown word speech to be recognized is input from the microphone 130, the speech analysis unit 140 extracts parameters representing the characteristics of the input word speech and extracts the interval The detection is performed to create an input word speech pattern and input to the inter-vector distance calculation unit 121.

Each created input word speech pattern is expressed as a time series of vectors representing the characteristics of the input word speech pattern, similarly to the word standard pattern.

(B-3) Pattern length ratio calculation operation The pattern length ratio calculation means 110 calculates the pattern length ratio between the input word speech pattern and the word standard pattern.

The pattern length ratio is expressed, for example, by the difference between the length of the input word audio pattern and the standard word pattern length, or by the ratio of the two. In this example, the ratio of the two, that is, the input word audio pattern length divided by the word standard pattern length, This value will be used as the pattern length ratio.

(B-4) Restricted matching distance calculation operation The inter-vector distance calculation unit 121 calculates each vector of the input word speech pattern input from the speech analysis unit 140 and the word standard pattern dictionary 16.
The inter-vector distance between each word standard pattern in 0 and each vector is calculated.

First to third DP calculation units 122 to 124
refers to the inter-vector distances calculated by the inter-vector distance calculation unit 121, and calculates the distance between the input word speech pattern and each word standard pattern based on the first to third route restrictions shown in FIG. Calculated using the known DP method.

In FIG. 3, DP ₁ to _{DP 4} indicate each conventional route restriction, i to (i-2) are each vector number of the input word speech pattern, and j to (j-3) are each vector number of the input word speech pattern. These are the vector numbers of the word standard pattern, and p ₁ to _{p 5} are each pass when performing DP calculation.

DPa ₁ to DPa ₄ are the first path restrictions, and as shown in the figure, by excluding paths whose slope is greater than 1 from the corresponding conventional path restrictions,
The configuration is such that it does not include paths with slopes greater than 1.

DPb ₁ to DPb ₄ are second route restrictions, and as shown in the figure, except for DPb ₄ , they have the same configuration as the corresponding conventional route restrictions. DPb ₄ has a configuration in which paths with slopes less than 1 and paths with slopes greater than 2 are removed from the corresponding conventional path restriction DP ₄ .

DPc ₁ to DPc ₄ indicate the third route restrictions, and as shown in the figure, by excluding paths with slopes smaller than 1 from the corresponding conventional route restrictions, the paths with slopes smaller than 1 are obtained. The configuration does not include any paths. DPc ₄ is
Furthermore, paths with slope 1 are also excluded.

The first path restriction (DPa ₁ -DPa ₄ ) is selected when the pattern length ratio is greater than the neighborhood of 1, for example when the pattern length ratio is >1.2.
This is the case when words are uttered naturally and more slowly than usual.

The second path restriction (DPb ₁ to DPb ₄ ) is, for example, when the pattern length ratio is close to 1.
Selected when 0.8≦pattern length ratio≦1.2. This is the case when words are uttered naturally and at a normal rate.

The third path restriction (DPc ₁ to DPc ₄ ) is selected when the pattern length ratio is smaller than the neighborhood of 1, for example, when the pattern length ratio is <0.8.
This is the case when words are uttered more quickly than normal in a natural way. These pattern length ratios are determined experimentally to obtain the best recognition results.

Note that the first to third route restrictions are selected from the same group (DPak to DPck).
For example, if DPa ₁ is selected as the first path restriction, DPb ₁ is selected as the second path restriction, and DPc ₁ is selected as the third path restriction.

First to third DP calculation units 122 to 124
The respective DP calculations in are performed in parallel, and when the distance between input word patterns based on the first to third route restrictions for one word standard pattern is calculated, the route selection unit 125 uses pattern length ratio calculation means. Based on the pattern length ratio regarding the word standard pattern input from 110, each of the first to third DP calculation units 122 to 1
Select one of 24. For example, if the pattern length ratio is smaller than 0.8, the distance calculated by the third DP calculation unit that performed the DP calculation according to the third route restriction is selected and input to the word recognition unit 170.

The above-described DP calculation process based on the route restriction is performed for each word standard pattern in the word standard pattern dictionary, and each word is input to the word recognition unit 170.

(B-5) Word recognition operation The word recognition unit 170 detects the minimum value of each distance on the restricted route between the input word audio pattern input from the route selection unit 125 and each word standard pattern, The word corresponding to the standard word pattern is defined as a recognized word.

Figure 4 shows the DP matching method of the present invention. Figure a is a reproduction of Figure 6 as is, and Figure b shows the conventional case of figure a compared to the DP matching method of the present invention. This figure shows an example of matching according to the method.

Since the pattern length of the input word audio pattern is 10 frames, and the pattern length of the word standard pattern during matching is 13 frames,
The frame length is 0.77 (=10/13). Therefore, the calculation result of the third DP calculation unit 124 that performs DP calculation is selected according to the third route restriction.

As a result, matching is performed using the DP path with the natural time correspondence shown in FIG. 4b instead of the one shown in FIG. 4a, and a correct recognition result can be obtained.

Although one embodiment of the present invention has been described above, each structure of the present invention is not limited to each structure of this embodiment.

For example, the mode of route restriction is not limited to the mode shown in FIG. 3, and various other route restrictions can be used, and the number of DP calculation units can be changed accordingly.

Also, the DP calculation units 122 to 124 are arranged in one DP instead of being arranged in parallel as shown in FIG.
The calculation unit may perform the calculation in a serial format. However, as shown in FIG. 2, by performing DP calculations based on each route restriction in parallel using a plurality of DP calculation units, it is possible to speed up the DP calculation process.

〔Effect of the invention〕

As explained above, according to the present invention, the path restriction during matching is changed in accordance with the natural speaking speed of the input word, without narrowing the range of expansion/contraction ratios allowed for the word as a whole. , it is possible to prevent matching based on unnatural time correspondences, and it is possible to improve the recognition rate.

[Brief explanation of drawings]

Fig. 1...Explanatory diagram of the basic configuration of the present invention, Fig. 2...Explanatory diagram of the configuration of an embodiment of the present invention, Fig. 3...
...An explanatory diagram of the route restriction in the same embodiment, Fig. 4...An explanatory diagram of the DP matching method in the same embodiment, Fig. 5...
FIG. 6 is an explanatory diagram of the conventional word recognition method. An explanatory diagram of the DP matching method of the conventional word speech recognition method. In FIGS. 1 and 2, 110...Pattern length ratio calculation means, 120...Limited route distance calculation means, 130...Microphone, 140...Speech analysis unit, 150...Switching circuit, 160...Word standard Pattern dictionary, 170...word recognition means.

Claims (1)

[Scope of Claims] 1. In a word speech recognition device that recognizes an input word speech by comparing an unknown input word speech pattern with each word standard pattern, (a) an input word speech pattern created from the input word speech and each (b) pattern length ratio calculation means 110 for calculating a pattern length ratio representing the magnitude of each pattern length of the word standard pattern; A word speech recognition device comprising: restricted path distance calculation means 120 for calculating the distance between a word speech pattern and each word standard pattern. 2. If the pattern length of the input word speech pattern is larger than the pattern length of the standard word pattern by a predetermined amount, the restricted route distance calculation means 120 changes the route restriction to one that does not include a route with a slope greater than 1, and Claim 1: If the pattern length of the voice pattern is smaller than the pattern length of the word standard pattern by a predetermined amount, the route restriction is changed to one that does not include a route with a slope smaller than 1. Word speech recognition device described in Section 1.