JPS6130280B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a method for recognizing the speech of a specific speaker. The type of voice identification device that is currently being commercialized the most is what is called a word voice recognition device for a specific speaker. This is done by having the speaker utter all the words to be recognized one to 10 times in advance, so that the characteristics of the speaker's words are stored in the recognition device, and then recognition is performed. This is because variation due to individual differences in voice patterns (mainly related to frequency structure) is one of the major factors that makes recognition difficult, but this can be avoided. In this specification, words uttered in advance by a speaker are referred to as registered words, individual features of the registered words are referred to as registered feature elements, and a set of registered feature elements related to each registered word is referred to as a registered feature set. The newly uttered word to be recognized is called a recognized word, a recognized feature element, and a recognized feature set, respectively. next,
An example of this specific speaker recognition device is shown in FIG. In Figure 1, 1 is a microphone, 2 is a preamplifier, 3 is a band filter group, 4 is a rectifier/low-pass filter group, 5 is a multiplexer, and 6 is an AD
Converter, 7 is a control unit, 8 is a multiplexer switching signal line, 9 is a control line to the AD converter, 10 and 11 are response signal lines between the control unit 7 and the computer 12, 12 is a microcomputer and its The peripheral portion 13 is configured like an identification result output line. The operation is as follows: First, an audio signal is converted into an electric signal by a microphone 1, amplified by a preamplifier 2, and decomposed into spectra by a group of band filters 3. Generally, the configuration of the band filter group 3 is such that the audio band from about 150 Hz to about 5 kHz is divided into 3 to 15 parts. Bandwidth filter group 3
The output is temporally averaged through a rectifier/low-pass filter group 4 and becomes an input signal to a multiplexer 5. The output signal of the multiplexer 5 selected by the switching signal line 8 is AD at the cycle activated by the control line 9.
The converter 6 converts the analog signal into a digital signal, which is controlled by response signal lines 10 and 11 and transferred to the computer 12. This transfer cycle becomes the sampling cycle of the audio data, and the control unit 7 controls the timing of AD conversion. Generally, the entire output of the band filter group 3 is often sampled at a period of several ms to several tens of ms. Most voice recognition devices use hardware up to the AD converter, and the subsequent processing is executed by a minicomputer or microcomputer. This is because the data input speed is relatively slow, so it is possible to process without using dedicated hardware except for areas unsuitable for computer processing (analog signal processing section), which makes it more compact and cheaper. This is because it is easy to realize. FIG. 2 is a flowchart showing the basic configuration of the computer processing section. 14 is audio data input processing, 1
5 is audio extraction processing, 16 is feature extraction processing, 17
is training data or recognition data, 18
19 is a registration process, 19 is an identification process, and 20 is a result output process, and the operations will be explained below. The audio data input process 14 is a process of importing audio data into the internal memory of the computer.
When an input command for audio data is given in some form, it synchronizes with the control section 7 and stores the output data (hereinafter referred to as sample data) of the AD converter 6 in the memory. Generally, word recognition devices use approximately 1.5 seconds of vocalizations. Since the sample data also includes extra data before and after the utterance, the next step is to detect the utterance section in the voice cutting process 15. As long as the speaker is not in a noisy environment, a simple method for this process is to set and compare a threshold value in which sample data values before and after utterance are smaller than those at the time of utterance. The sample data of the detected voice section has a large amount of data, and a device with a large storage capacity is required. For example, if the utterance time is 1.5 seconds, the sampling period is 10 ms, and AD conversion is 8 bits, a memory of 150 bytes x number of channels is required for each word to store the number of words to be identified in advance.
Sample data is often transformed into some form to reduce the amount of data and then saved. The feature extraction process 16 performs this work. A simple method for feature extraction is a linear compression method of sample data.
In this method, the extracted audio section is divided into equal parts (usually 16 to 32 parts), an average value is determined for each channel data for each divided section, and this average value is used as a feature. However, since the loudness of the voice changes even for the same speaker, it is better to normalize the loudness using data such as a filter output added value for each sample data or at the feature level. In the feature registration process 18, the features obtained in this way are obtained for all registered words to be identified and stored in the recognition device (in the computer memory in this example). When the registration is completed, the determination process 17 is executed so that the identification process 19 is executed. In short, the identification process 19 calculates the degree of dissimilarity between the registered feature set of each registered word stored in advance based on the utterances of the same speaker and the recognized feature set of the newly uttered recognized word. It is measured according to the law, and in the result output processing 20, the registered word (code, etc.) with the minimum dissimilarity is output to the outside, and if the minimum dissimilarity does not satisfy a certain condition, a reject output is output. do. Note that although some methods define the identification process by measuring similarity, they are essentially the same. An example of calculation in the identification process 19 will be explained using the following formula. Let the number of saved registered words be N. The registered feature element is f _i (l, m), and the recognized feature element is g
Let it be (l, m). However, the subscript i is a number attached to the registered word, and in this case is a value from 1 to N. l
is a number corresponding to each film output and takes a value from 1 to the number of filters L. m is a number assigned to each time division area and takes a value from 1 to the number of divisions M. Considering points lost as the degree of dissimilarity, the sum of points lost between the registered feature set F _i of a certain registered word stored and the newly uttered recognized word and recognized feature set G is calculated as S.
If _i is

[Formula]. The identification result is a registered word that is MIN (S ₁ , S ₂ , . . . _SN ). However, MIN(S ₁ , S ₂ ,...S _N ) is
This means selecting the one with the smallest total number of points conceded from among S ₁ , S ₂ , ..., _SN . Although this method is simple, it has the disadvantage that if there are similar pronunciations in the target word (for example, ``nakano'' and ``nagano''), the difference in points will be small and it will be difficult to distinguish. Ta. It is extremely difficult for humans to produce exactly the same vocalizations (voice strength, speaking speed, accent, clarity, etc.), so try saying the same word twice and find out the difference (points lost). This is because the amount of points lost is the same as the difference (points lost) when similar words are uttered. The purpose of the present invention is to overcome these drawbacks.
The registered feature elements are weighted so that even words with similar features can be easily distinguished, and will be explained in detail below. In the present invention, all corresponding registered feature element pairs in all registered feature set pairs, regardless of whether they are stored individually for each feature element pair or representatively for each group of feature element pairs. The weighting coefficients of are separately detected and stored. The weighting coefficients are set so that among similar registered word pairs, the weighting coefficient of a registered feature element pair having a large degree of dissimilarity is larger than the weighting coefficient of the remaining registered feature element pairs. In the present invention, when recognition is not possible using the registered feature set, a weighting coefficient is further introduced to perform re-recognition. FIG. 3 is a flowchart of the first embodiment of the present invention. The first stage of computer processing will be explained. 21 is a process for determining whether to process training data, weight calculation, or recognition data; 22 and 27 are audio data input processes; 23 and 28 are audio extraction processes; 2
4, 29 is a feature extraction process, 25 is a feature registration process,
26 is a weight calculation process, 30 is an identification process, and 31 is a result output process, and the operations thereof will be explained below. The judgment process 21 can be carried out in various ways depending on the specifications of the apparatus, but for simplicity, the operator uses a keyboard attached to the apparatus to instruct the process every time or when the judgment changes. Processes 22 to 25 in training data processing (processing to store previously registered characteristics of speakers) and 27 to 25 in identification data processing (processing to perform actual recognition)
Processes 29 and 31 are basically the same processes as those described in the conventional method, so their explanation will be omitted.
When the processing of the training data is completed for all registered words, the decision 21 operates to perform the weight calculation process 26. The weight calculation process 26 will be explained below.
As explained in the conventional method, let any two registered feature elements be f _i (l, m), f _j (l, m), let their registered feature sets be F _i , F _j , and write any registered word The difference D _ij between a pair of registered feature sets and the difference d _ij (l, m) between individual registered feature elements are defined as follows. d _ij (l, m) = | f _i (l, m) − f _j (l, m) | where i≠j ……(1) The weighting coefficient is an arbitrary registered feature set F _i (i=
1, 2, ..., N) and all other registered feature sets F _j (j = 1, 2, ..., N, where i≠j and N is the number _of registered words). do. (1) For D _ij K1. (However, K1 is a predetermined constant) F _i and F _j have sufficient differences in their characteristics, and the difference in points lost during identification is considered to be large, so F _i and F _j
The weighting coefficients _{w ij} (l, m) of the corresponding registered feature request pairs f i (l, m), f _j (l, m) are set to "1". (2) When D _ij <K1. f _i (l, m) and f _j that satisfy d _ij (l, m)K2 (where K2 is a predetermined constant)
For (l, m), let w _ij (l, m)=K3.
However, K3 is a constant value of a predetermined weighting coefficient, and K3>1. Also, d _ij (l, m)<
For K2, w _ij (l, m)=1. Therefore, the weighting coefficient w _ij (l, m) takes the value of "1" or K3 according to the above-mentioned conditions. The above calculations are performed for all registered features and the weighting coefficients w _ij are stored. The storage state of the weighting coefficients on the memory is shown in FIGS. 4 and 5. In Figure 4, the sets of weighting coefficients in the registered feature sets F _i and F _j are shown in capital letters W _ij to make it easier to understand the relationship between them, and the diagonally shaded squares are portions that do not actually exist as storage memory. . FIG. 5 shows weighting coefficients w _ij ( _l ,
m) (l is the channel number, m is the divided area number) is shown in the actual storage state on the memory. When the operation of the weight calculation process 26 is completed in this manner, the process proceeds to identification data processing. In the identification data processing, only the identification processing 30 will be explained. In the conventional example, the total number of points conceded S _i is

Although defined by [Formula], in the present invention, first, the total number of points conceded S is calculated using the conventional method.
Calculate _i . Then, S _i with the fewest points conceded is S _a , and the one with the next fewest goals conceded is also S _b , and when it satisfies S _b - S _a K4 (where K4 is a predetermined constant), it corresponds to S _a Registered word a is output as the identification result, but if S _b −S _a <K4, registered word a,
Since b means similar, registered word a
The next total of points conceded SW _a and SW _b are calculated again for and b. However, the weighting coefficient wab (l, m) takes the value of "1" or K3 according to the above-mentioned conditions. Again, if |SW _a −SW _b |K4 (where K4 is a predetermined constant) is satisfied, then Min(SW _a ,
The registered word a or b that is SW _b ) is the identification result.
Also, if it is not satisfied, it will be rejected (unidentifiable). In addition, when the minimum loss (minimum dissimilarity) S _a is too large for the purpose of reducing identification errors, or
If Min (SW _a , SW _b ) is too large, it is better to reject. Here, a more detailed explanation will be given using one example. In this example, in order to simplify the discussion, the number of filters L = 1, the number of divisions M = 4, the registered feature sets of identification candidates a and b with respect to the recognized feature set G of the recognized word are Fa, Fb, G, Fa , Fb shall take the following values. G = (8, 16, 5, 28) Fa = (8, 15, 4, 32) Fb = (8, 18, 7, 25) Also, K ₁ = 15, K ₂ = 5, K ₃ = 5, K When ₄ =3, Dab=|(8-8)|+|(15-18)|+|(4-7)|+|(32-25)|=13, and Dab<K ₁ . The set Wab of weighting coefficients in each registered feature set Fa and Fb is the difference dab between the individual registered feature elements.
Weighting coefficient wab (l, m) for each registered feature element that satisfies (l, m) K2 (=5) = K3 (=5)
and the weighting coefficient wab (l, m) for each other registered feature element = 1, then Wab = (1, 1,
1, 5). Here, the sum totals Sa and Sb of the recognized feature set G and each registered feature set Fa and Fb are Sa=|(8-8)|+|(15-16)|+|(4-5) ) | + | (32-28) | = 6, Sb = | (8-8) | + | (18-16) | + | (7-5) | + | (25-28) | = 7. , the total points lost Sa, Sb, the registered word a becomes the first candidate for the identification result. However, Sb−Sa=
Since 1<K ₄ (=3), the weighting coefficient Wab
Calculate the total points conceded SWa and SWb by multiplying by As a result, SWa=(0×1)+(1×1)+(1×1)+(4×5)=22 SWb=(0×1)+(2×1)+(2×1)+ (3×5)=19, and |SWa−SWb|=3K4 (=3), so registered word b becomes the final recognition result. The case where there are two registered word candidates to be output has been described above, but even if there are two or more candidates, it can be dealt with by considering the combinations of two at a time. For example, if S _a , S _b , and S _c show similar points, the registered word pairs (a, b), (a, c),
Calculate the points conceded for (b, c) in the same way as above,
Average points lost SW _a for each registered word a, b, c,
Just take SW _b and SW _c and find Min (SW _a , SW _b , SW _c ). Also, if the weighting coefficients are all words that have the same value, store only one representative weighting coefficient (for example, instead of directly referring to the memory that stores the weighting coefficients, store two weighting coefficients). Once the table is drawn from the numbers attached to the words,
The table records the storage start address of the weighting coefficients or the representative weighting coefficients, and by clearly indicating which one is recorded, the memory capacity for storing the weighting coefficients can be reduced. As explained above, in the first embodiment, the difference between the registered feature sets of each registered word is calculated, and the differences among the registered feature elements of similar registered words are determined. Since the feature elements of the part that are recognized are heavily weighted, it is possible to distinguish between similar recognized words that would be difficult to distinguish using conventional methods in which weighting is uniformly applied. There is an advantage that the recognition rate can be increased by focusing on the points where the features are different.
In the first embodiment, different weighting coefficients can be given to all the registered feature elements, but in order to reduce the memory capacity for storing the weighting coefficients, the weighting coefficients are set for each time-divided area. 1
Even if one is prepared and all channel filter data (features) belonging to the divided area have the same weighting coefficient, sufficient effects can be obtained. In this case, the equations (1) and (2) in the first embodiment correspond to It is. The expression of the weighting coefficient is w _ij (l)=K3, whereas w _ij (l, m)=K3 in the first embodiment. Also, weighting coefficients may be assigned to each frequency domain instead of each divided domain, but in this case the effect is somewhat reduced. The present invention automatically calculates and weights the differences in features between target words when all the features of the target words uttered by the speaker are completed, so even similar words can be accurately distinguished. It can be used in speech recognition devices.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an example of a speech recognition device, FIG. 2 is a flowchart of a computer processing section of a conventional speech recognition device, FIG. 3 is a flowchart of a computer processing section of an embodiment of the present invention, and FIG. 4 1 is a schematic diagram showing the weighting coefficients stored in memory, and FIG. 5 is a detailed diagram showing one block thereof. 1...Microphone, 2...Preamplifier, 3
...Band filter group, 4...Low pass filter group, 5...Multiplexer, 6...AD converter,
7... Control unit, 8... Multiplexer switching signal line, 9... AD converter control line, 10, 11...
Interface signal line between control unit 7 and computer 12, 12...Small computer or microcomputer and its peripheral parts, 13...Identification result output line, 14
...Audio data input processing, 15...Audio extraction processing, 16...Feature extraction processing, 17...Judgment of training or identification, 18...Feature registration processing, 19...Identification processing, 20...Result output processing , 21... Judgment of training, weight calculation, or identification, 22, 27... Audio data input processing, 23, 28... Audio extraction processing,
24, 29...Feature extraction processing, 25...Feature registration processing, 26...Weight calculation processing, 30...Identification processing, 31...Result output processing, D _ij ...Difference between registered feature set pairs, d _ij ( l, m)... Difference between registered feature element pair, F _i , F _j ... Registered feature set, f _i (l,
m), f _j (l, m)...registered feature element, g(l,
m)...Recognized feature element, W _i , _j ... Set of weighting coefficients in the registered feature set pair, w _ij (l, m)
...Weighting coefficients of registered feature element pairs, S _i , SW _a ,
SW _b ...Total points conceded.

Claims (1)

[Claims] 1 The degree of dissimilarity between the registered feature set consisting of the registered feature elements of each registered word based on the utterances of the same speaker and the recognized feature set consisting of the recognized feature elements of the recognized word is measured according to a certain rule. In a speech recognition method that outputs a registered word corresponding to a registered feature set with a minimum dissimilarity, the degree of dissimilarity of each corresponding registered feature element pair in a registered feature set pair is measured according to a certain rule, and the registered feature is The dissimilarity between pairs of registered feature sets is measured according to a certain rule, and the weighting coefficient of each corresponding pair of registered feature elements is stored for all combinations of registered feature sets. The weighting coefficient of the registered feature element pair whose degree of similarity is smaller than the first certain value and whose degree of dissimilarity between the registered feature element pair is larger than the second certain value is set higher than the weighting coefficient of the remaining registered feature element pairs. The present invention includes a weighting coefficient setting storage means for setting a large weighting coefficient, and detects a small number of registered words with small dissimilarity when the minimum dissimilarity between each registered feature set and the recognized feature set does not satisfy a certain condition. step, each registered feature set targeting the detected registered word and using the registered feature element obtained by multiplying the registered feature element by the weighting coefficient as a new registered feature element, and adding the above to the recognized feature element. Step of measuring the degree of dissimilarity with the recognized feature set, using the new recognized feature element obtained by multiplying it by the weighting coefficient, and outputting a registered word corresponding to the degree of dissimilarity that satisfies a certain condition. A speech recognition method comprising: