JPS5936758B2 - Voice recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention provides a speech recognition method, particularly speech recognition that recognizes a word including a sentence and/or a speaker who uttered the word. By creating a time-sequence pattern in which a composite standard phoneme set consisting of at least a first standard phoneme set and a second standard phoneme set paired in the same manner is arranged in a time series, it is possible to create a time-series pattern in which a composite standard phoneme set, which is made up of at least a first standard phoneme set and a second standard phoneme set, is arranged in a time series. This invention relates to a voice recognition method that can correctly recognize slight changes in the road regardless of differences in voice signals.

In the speech recognition method, the present inventors previously invented a patent application filed in 1972-4 in order to perform recognition by specifically grasping the characteristics of the speaker.
In the invention related to No. 1341 "Speech recognition method", a standard phoneme set made up of a plurality of standard phoneme pairs is used, and a standard time for recognizing a time series pattern in which the standard phoneme set is arranged in a time series. We proposed that it be used as a series pattern.

The present invention is a further development of the above-mentioned invention, and is designed to enable accurate recognition within a certain time interval, regardless of differences in audio signals caused by slight changes in the vocal tract caused by the situation when a speaker pronounces. To achieve correct recognition, a composite standard phoneme set is constructed by extracting a first standard phoneme set with the highest degree of similarity and a second standard phoneme set with the next highest degree of similarity. The purpose is

Note that the process of extracting these standard phoneme sets can be performed more accurately and easily by using a data processing device. In order to achieve the above object, the speech recognition method of the present invention recognizes a word including a sentence and/or a speaker who uttered the word. It has a time-series pattern in which a composite standard phoneme set consisting of a first standard phoneme set and a second standard phoneme set, which are also paired, is arranged in a time series, and the unknown input speech is divided at predetermined time intervals. means for calculating the degree of similarity between a phoneme within a time interval and the first standard phoneme set and the second standard phoneme set; The recognition is performed according to which of the time series patterns formed by combining the first standard phoneme set and the second standard phoneme set, Based on a plurality of composite standard phoneme sets, one of the first standard phoneme set and the second standard phoneme set in one composite standard phoneme set and the first standard phoneme set in the other composite standard phoneme set. A time series pattern configured by combining a first standard phoneme set and a second standard phoneme set, such as combining either one of the first standard phoneme set and the second standard phoneme set, A degree of similarity is calculated for each of the time series of phonemes within the divided time intervals of the known input voice and the time series pattern, and based on the calculation result, the time series of the phonemes of the known input voice are divided into a plurality of the time series. If the patterns have the same degree of similarity within a predetermined range, the time series pattern with the greater number of first standard phoneme sets in the time series pattern is selected as the standard time series pattern for unknown input speech recognition. In this case, the plurality of time-series patterns having the same degree of similarity are classified according to the form of their relationship with each other, and the classification information according to form is used for unknown input speech recognition. It is characterized by being one piece of information. This will be explained below with reference to the drawings. Figures 1 to 4 conceptually explain the previously filed invention. Figure 1 is an explanatory diagram explaining the sampling of audio signals, and Figure 2 is a standard diagram for pairing a plurality of standard phonemes. An explanatory diagram illustrating the process of extracting phoneme sets,
Figure 3 is an explanatory diagram explaining the process of determining similarity based on a standard phoneme set, Figure 4 is a configuration diagram of the speech recognition device, and Figure 5
The figure is a processing block diagram of an embodiment according to the present invention, FIG. 6 is a block diagram of an embodiment for determining a standard time series pattern, and FIGS. 7 to 12 are diagrams showing the standard time series pattern and format-based classification information according to the present invention An explanatory diagram illustrating the process of obtaining .

A patent application filed in 1973 prior to the explanation of the present invention
The invention related to No.-41341 will be briefly described.

In the invention, first, sampling processing of the audio signal is performed as shown in FIG. That is, for example, ten partial autocorrelation coefficients are extracted within one frame of the illustrated audio signal 1, for example, 30 ms, and the frames are shifted by, for example, 10 ms. The extracted partial autocorrelation coefficients are arranged in chronological order and are arranged into predetermined standard phonemes, for example 12!, as shown in FIG. All, la2l, . . . The degree of similarity to 1r1 is summarized in a table. Second illustration
In the figure, it is shown that the phoneme k parameter V in a certain frame has a degree of similarity expressed by Sj(c) with respect to the standard phoneme k parameter Kj, and the phoneme parameter k' is the most similar to the standard phoneme parameter Kg. It is shown that it has a high degree of similarity and has the next highest degree of similarity to the standard phoneme parameter K2. In addition, C in the above Sj (C)
represents the degree of similarity to the pronunciation by speaker C. As shown in FIG. 2 above, the standard phoneme jl has the highest degree of similarity and the standard phoneme J.1 has the second degree of similarity. I looked up the standard phoneme set Jl,j. and create a time-series pattern of standard phoneme sets that appear when a certain speaker C pronounces a certain word, for example the number "4", for example {8, 2, 48}
:{ 2, 6, 66}; { 3, 11, 78}; {1,
2,47}. In addition, 48,6 in {} above
Reference numerals 6, 78, and 47 represent characteristic coefficients (pitch) related to the vocal cords, although their explanation will be omitted. The time-series pattern of the standard phoneme set obtained as above (when a certain speaker C pronounces a certain word) is arranged on the horizontal axis of the illustrated table in FIG.
The degree of similarity of the phoneme time series of the unknown input speech signal is investigated.

Note that S, (1., j) in FIG. 3 is the first time period T. The similarity to the standard phoneme K8 indicated by the number "8", Sl(2, j), is the same time period T. In the number “
Standard phoneme K indicated by ``2''. represents the degree of similarity to Then, by a process called a dynamic program, for example, the similarities enclosed by the thick black lines in the figure are added up, and for a certain word pronounced by a certain speaker C in the unknown input audio signal, for example, "4". Determine the degree of similarity. Of course, in recognition, the process shown in FIG. 3 is repeated (number of speakers) x (number of words) times to find the one with the highest degree of similarity. Although the explanation has been omitted, in the previous invention, the degree of dissimilarity is taken into consideration in the time series of pitches, and recognition is performed based on the degree of similarity and the degree of dissimilarity. FIG. 4 shows the configuration for voice recognition in the previous invention, and there is essentially no difference in the voice recognition according to the present invention.

In the figure, 1 is a k-parameter extraction/pitch extraction device, 2, 4
, 6, 8, 10... are standard k parameter storage units when a certain word is pronounced by a certain speaker, 3, 5, 7, 9, 1
1, ... are corresponding pitch storage units, A is a similarity calculation circuit, B is a dissimilarity correction circuit, C is a similarity sum calculation circuit, D is a dissimilarity sum calculation circuit, and E is a dissimilarity calculation circuit. Sum absolute value circuit,
W represents a weighting circuit, F represents an addition circuit, and G represents a maximum detection circuit. Although a detailed explanation will be omitted, as described above, the maximum value is selected from the sum of similarities for the phoneme time series pattern minus the sum of dissimilarities multiplied by the weights for the pitch time series pattern.

As mentioned above, we have outlined the previously filed invention, but depending on the situation when a speaker pronounces, slight changes in the vocal tract can cause the audio signal to change, and such changes can cause erroneous recognition. It is necessary to prevent this from occurring.

In order to solve this problem, the present invention uses a first standard phoneme set that occurs most frequently in a certain time period when a certain speaker pronounces the same word, and a second standard phoneme set that occurs next most frequently. A standard time series pattern is prepared using these. FIG. 5 represents a processing block diagram based on the speech recognition method of the present invention. In the case of the above-mentioned invention filed earlier, the first standard phoneme set and the second standard phoneme set are
using a composite standard phoneme set consisting of the standard phoneme set of
and the use of type classification information. The processing within the chain line frame in FIG. 6 is “
In the case of the present invention, in this process, (i) as mentioned above, based on the known input speech, the standard phoneme set with the highest frequency of occurrence and the next Determine a composite standard phoneme set consisting of the second-rank standard phoneme set, (il) create a combination series of these first-rank standard phoneme sets and the second-rank standard phoneme set, and (110 more A known input speech is input and a nonlinear matching calculation is performed on the previously obtained combination sequence, Qψ, thereby determining a standard time sequence pattern for the word by the speaker. Then, based on this determined standard time series pattern, the degree of similarity with the unknown input voice is calculated as shown in Figure 5.Hereinafter, the process diagram shown in Figure 6 will be used in the present invention. Determination of a standard time series pattern and determination of type classification information will be explained.In the case of the present invention, each speaker utters the same word many times, and A composite standard phoneme set is obtained for a certain frame when a specific word is uttered.

That is, by sampling the audio signal 1 when a specific word is uttered by a specific speaker in the same manner as shown in FIG. 1, as shown in FIG. The first standard phoneme set (S, (1, j), S, (2, j)) that occurs most frequently among those uttered and the second standard phoneme set (Sl7 ( 1・j
), SS(2・J)) to obtain a composite standard phoneme set. This composite standard phoneme set is generally expressed as Al{El, Fl} in frame l as shown in FIG. 8, and the composite standard phoneme set Al in the first frame and the composite standard phoneme set Al in the second frame as shown in FIG. A2,・
..., Af! , ..., An are arranged in chronological order. This situation is illustrated in FIG. In the time series shown in Figure 9, the same composite standard phoneme set may exist between adjacent frames, so the same composite standard phoneme set is aggregated and adjacent ones are different from each other as shown in Figure 10. A time-series pattern is obtained that is a composite standard phoneme set Bl, B2, . . ., Bm.

Figure 11 shows that the composite standard phoneme set obtained as above is Bl.
, B2, B3, B4,
Each composite standard phoneme set is represented as consisting of {E, F}.

The time-series pattern of the composite standard phoneme set represents the time-series of the composite standard phoneme set that would be realized when a certain speaker pronounces a certain word and the speech signal is divided into time zones and examined. In the case of the invention filed earlier, the time series E1→E2→E3→E4 consisting of the standard phoneme set ElE2, E3, and E4 shown in FIG. 11 was taken into consideration. When a certain speaker attempts to pronounce a certain word accurately, it is sufficient to use the above-mentioned time series consisting of E1→E2→E3→E4 as a standard time series pattern. However, even when the same speaker pronounces the same word, El. →E2→E3→E
Time series pattern represented by 4, E1 → E2 → E3 → E
Time series pattern represented by 4,...F1→F
There is a possibility that a total of 16 time series patterns, represented by 2→F3→F4, may appear. Therefore, in the case of the present invention, when the same speaker pronounces the same word under various circumstances, the results obtained in the [combination series of standard phoneme sets] shown in the "nonlinear matching" shown in FIG. We will investigate which of the above 16 time series patterns occurs more often, determine the time series pattern with the highest frequency of occurrence, and use it as the standard time series pattern for future unknown input speech recognition. do. That is, when investigating the frequency of occurrence, it goes without saying that the above chronological pattern E1→E2→E3→E is used for the input voice.
4......, F1→F2→F3→F4, respectively, to find out which of them is similar, and thereby the above-mentioned occurrence frequency distribution is determined. However, in this case, the frequency of occurrence of other time-series patterns may be almost the same as that of the one time-series pattern with the highest frequency of occurrence. In such a case, one of the standard time series patterns is determined as follows. That is, for example, for each of two time-series patterns, the number of first standard phoneme sets E and the number of second standard phoneme sets F that are similar to the input voice phonemes within a predetermined range are checked, and the number of first standard phoneme sets F is determined. The time series pattern with the larger number of E's is selected as the standard time series pattern. Also the first
If the number of standard phoneme sets E is the same, the time series pattern on the side with the first standard phoneme set is selected in the time zone where the accent is present. However, even in such cases, it is desirable to supplementally overcome the drawbacks caused by selecting only a single standard time series pattern. For this reason, type-specific classification information is prepared by classifying the relationship between two time-series patterns having substantially the same frequency of occurrence according to type. Also, as shown in Figure 6,
"Nonlinear matching" is a nonlinear matching process performed by the conventionally known dynamic programming (DP) method. That is, the input audio and each standard sequence are matched in a manner that is not easily affected by the time delay, so to speak, taking into account the time delay. FIG. 12 shows the form and classification information representing the relationship between two time-series patterns.

Figure 12A shows the forms in which the same standard phoneme set is common in any one of the four types, and the classification information G for each form.
ll, Gl2, Gl3, and Gl4. 1st
Figure 2B shows the forms when any two standard phoneme sets are common and the classification information for each form G2l, G22, G23, G
24, G25, G26, and so on. Figure 12C
represents the form and its form-based classification information G3l, G3, , G33, G34 when any three standard phoneme sets are common. Needless to say, the fact that the four standard phoneme sets are common means that the two time series patterns are the same, and there is only one form that has nothing in common in any time period, so the classification information by form is For example, G. O is given. In the case of the present invention, the standard time series pattern obtained by considering the composite standard phoneme set as described above is stored in the standard k parameter storage units 2, 4, 6, 8, 10 shown in FIG. It is stored in...

Then, as in the case of the previous invention, a recognition output is obtained from the maximum detection circuit G. The standard time series pattern obtained by considering the composite standard phoneme set is prepared for each speaker and for each word. However, in the case of the present invention, the storage units 2, 4, 6, 8, 10...
Store it in... In order to obtain the degree of similarity with the unknown voice input, in addition to the comparison with a single standard time series pattern, a comparison is made with a time series pattern whose frequency of occurrence is close to that of the standard time series pattern using the type classification information. Make sure to do this at the same time. At this time,
If the comparison is performed with the speaker fixed, word recognition is possible, if the word is fixed, speaker recognition is possible, and if both are not fixed, it is possible to recognize which speaker uttered which word. becomes. As explained above, according to the present invention, by introducing the concept of composite standard phoneme sets, it is possible to avoid misrecognition caused by differences in the pronunciation situation, and to improve speech recognition when the number of speakers and words increases. It is extremely effective.

Note that in extracting the first standard phoneme set and the second standard phoneme set, which are a prerequisite for obtaining the composite standard phoneme set, it is possible to rely on processing using a data processing device.

In this case, the processing is quick and the extraction criteria are standardized.

[Brief explanation of drawings]

Figures 1 to 4 conceptually explain the previously filed invention. Figure 1 is an explanatory diagram explaining the sampling of audio signals, and Figure 2 is a standard diagram for pairing a plurality of standard phonemes. FIG. 3 is an explanatory diagram explaining the process of extracting a phoneme set, FIG. 3 is an explanatory diagram explaining the process of determining similarity based on a standard phoneme set, FIG. 4 is a block diagram of the speech recognition device, and FIG. FIG. 6 is a processing block diagram of an embodiment according to the present invention, and FIG. 6 is a diagram of an embodiment of determining a standard time series pattern.
FIG. 2 shows an explanatory diagram illustrating the process of obtaining a standard time series pattern and type classification information according to the present invention. In the figure, 1 is a k-parameter extraction/pitch extraction device, 2, 4,
6, 8, 10... are standard k parameter storage units,
3, 5, 7, 9, 11... are pitch storage parts, G
is the maximum detection circuit, 1a11,! A2l... is a standard phoneme, (S1(1,j), S2(2,j)) is the first standard phoneme set, (S((1・j), S:(2・j) ) is the second standard phoneme set, {S1 (1, j), S2 (2, j),
S: (1, j), S: (2, j)} represents a composite standard phoneme set.

Claims (1)

[Claims] 1. A speech recognition method for recognizing either or both of a word including a sentence and a speaker who uttered the word, comprising: at least a first standard phoneme set consisting of a plurality of standard phoneme pairs; It has a time series pattern in which a composite standard phoneme set consisting of a second standard phoneme set which is also paired is arranged in a time series, and the unknown input speech is divided into predetermined time intervals, and the phonemes within the time interval and the above. means for calculating the degree of similarity between the first standard phoneme set and the second standard phoneme set; and the second standard phoneme set, the speech recognition method is characterized in that the recognition is performed depending on which of the time series patterns is most similar. 2. Based on a plurality of composite standard phoneme sets arranged in a time series, one of the first standard phoneme set and the second standard phoneme set in one composite standard phoneme set, and the other. Combining the first standard phoneme set and the second standard phoneme set, such as combining either one of the first standard phoneme set and the second standard phoneme set in the composite standard phoneme set of Provided with the constructed time series pattern, the degree of similarity is calculated for each of the time series of phonemes within the divided time intervals of the known input voice and the above time series pattern, and based on the calculation result, the time of the phoneme of the known input voice is calculated. If a sequence has the same degree of similarity within a predetermined range with respect to a plurality of the above-mentioned time-series patterns, the time-series pattern with a larger number of first standard phoneme sets in the time-series pattern is used for unknown input speech recognition. 2. The speech recognition method according to claim 1, wherein a standard time series pattern is used for the speech recognition method. 3 The first standard phoneme set and the second standard phoneme set in one composite standard phoneme set
A plurality of time-sequence patterns such as combining one of the standard phoneme sets of , and one of the first standard phoneme set and the second standard phoneme set in another composite standard phoneme set. If the time series of phonemes of known input speech have the same degree of similarity within a predetermined range, the relationship between the plurality of time series patterns is classified according to the form, and 2. The speech recognition method according to claim 1, wherein the classification information is used as one piece of information for unknown input speech recognition.