JPH0458634B2 - - Google Patents

Description

[Detailed description of the invention]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method that is advantageously used in Japanese speech input devices that recognize input speech syllable by syllable. BACKGROUND ART In general, when recognizing speech, speech that is uttered with pauses in each syllable can be recognized with high accuracy. However, in continuous speech, each syllable is influenced by the syllables before and after it (articulatory combination), and there are significant differences in the strength and pitch of the voice depending on the position of the syllable in a word, phrase, or sentence, resulting in a decrease in recognition accuracy. There is a problem that this occurs. this is,
This is because, in continuous speech, the characteristic patterns of syllables extracted from different context positions vary in various ways even for a specific person. In order to deal with these problems, conventional methods have been used to deal with this by having many feature standard patterns at the time of registration, and furthermore, when inputting the feature standard patterns created at the time of registration, feature standard patterns with poor recognition rates are used in various syllable environments. This was dealt with by updating the feature standard pattern by replacing it with a syllable feature pattern. Problems to be Solved by the Invention In the above-mentioned prior art, the registered feature standard pattern strongly depends on the input frequency of syllables, so recognition accuracy is improved on average. The accuracy of recognizing syllables when input was reduced. The purpose of the present invention is to solve the above-mentioned technical problems,
It is an object of the present invention to provide a syllable recognition method capable of efficiently modifying and updating even standard characteristic patterns of syllables with low input frequency and improving syllable recognition accuracy. Means for Solving the Problems The present invention recognizes input speech in units of syllables by calculating similarity with standard patterns of syllable characteristics of multiple types registered in advance, and compares the results with a dictionary. Or, in a speech recognition method in which the final input is obtained by modifying the input using an external command such as a keyboard, for each feature standard pattern of a syllable, at least one of the two before and after the feature pattern is extracted from the speech. The syllable feature pattern obtained by analyzing the voice uttered at the time of input can be added as a standard feature pattern, or the standard feature pattern can be modified using the input feature pattern. A feature of the present invention is that when registration is performed at the time of input, an update operation by additional correction is performed on a standard feature pattern that has the same environment information as the phonological environment on at least one side before and after the input feature pattern was extracted. This is a voice recognition method that uses Effect According to the present invention, information representing the phoneme immediately before or after a syllable is added in association with each feature standard pattern and stored in the feature standard pattern and the feature standard pattern. When adding a feature pattern as a pattern or modifying a feature standard pattern using the feature pattern, the feature standard pattern having the same phonological environment information as the input syllable is modified. By updating the syllable characteristic standard pattern based on the phonological arrangement in this way, it is possible to improve the syllable environment that is frequently input (for example, ``ai'' in ``kai'' and ``sai''). A feature standard pattern is formed from the feature pattern extracted from the phoneme "a" immediately before the syllable "i" as seen in The feature standard pattern is updated while retaining the feature pattern extracted from the feature pattern. Embodiments The present invention utilizes the concept of a syllable environment, and this syllable environment will first be explained. Tables 1, 2, and 3 show the results of an analysis of ``nainowa'' regarding the transformation of the syllable feature pattern depending on the phonetic environment before and after the syllable. Table 4 shows the analysis results for "Nondemo". This is a comparison of the similarity of characteristic patterns of syllables of the same type extracted from various phonological environments in a large number of syllables. Note that Table 1 shows the pattern distances for ``i'' in the input voice ``Nainowa'', and Table 2 shows the pattern distances for ``no'' in the input voice ``Nainowa''. Table 3 shows the pattern distance for ``wa'' in the input voice ``nai no wa''.
Furthermore, Table 4 shows the pattern distance for "no" in the input speech "nondemo". In Tables 1 to 4, the pattern distances are expressed in hexadecimal notation.

【table】

【table】

【table】

[Table] The ``no'' in ``nainowa'' comes from the ``no'' in ``garasuno,'' as in ``nioino,'' and ``kanjino.'' ”
is very similar to In other words, the vowel immediately before is "i"
It is similar to ``no''. In this way, it can be seen that the characteristic pattern of a syllable is strongly influenced by the phoneme before and after the syllable is extracted, and especially by the phoneme immediately before it. These transformations are correlated with the syllable environment, and the present invention utilizes this point. However, Japanese syllables are basically monosyllables consisting of a pair of consonants and vowels. According to the symbols in Table 5, for example, “Shiyakai no” is “sja-ka-i-no”,
“C1C2V−C1V−V−C1V”, “CV−CV−V−
It can be seen that it is broken down into "C" and is made up of four monosyllables. There are also semi-vowels between consonants and vowels, and continuous vowels, long sounds, consonants, etc. after vowels.

【table】

【table】

[Table] Therefore, the phonemes are encoded in hexadecimal as shown in Table 6, and these are used to create the Japanese syllable structure.
It is expressed in bits and encodes the surrounding syllables before and after. FIG. 1 is a block diagram showing the configuration of a Japanese speech input device 1 according to an embodiment of the present invention, and FIG. 2 is a flowchart showing the procedure of speech recognition processing in the Japanese speech input device 1. As shown in FIG. This Japanese speech input device 1 has a function of recognizing continuously uttered speech in units of syllables, correcting the recognition results using a dictionary, and then transmitting the results in units of words etc. to an external device. There is. First, the process moves from step n1 to step n2, where an audio signal is input. That is, the uttered and input voice is input to the analog input section 3 via the microphone 2, and after being amplified by the amplifier 4 in this analog input section 3, it is converted into a digital signal by the analog/digital converter 5. The digital signal is input to the speech analysis section 6 and the syllable segmentation section 7. Next, the process moves from step n2 to step n3, where acoustic syllable extraction processing is performed. That is, the speech analysis section 6 divides the input speech into frames of about 16 ms and performs spectrum analysis, and transfers the characteristic parameters necessary for syllable segmentation to the syllable segmentation section 7 at intervals of about 8 ms.
In the syllable segmentation section 7, the speech analysis section 6
While temporarily storing various feature parameters from the ring-shaped feature pattern buffer 8, syllables are cut out and the features of each syllable are patterned.
It is stored in the feature pattern memory 9 at n4. The feature pattern buffer 8 is configured to be able to store a plurality of syllables. The processing of the syllable segmentation unit 7 is carried out by the central processing unit (hereinafter referred to as CPU).
It is configured such that start and stop are controlled by commands from 10 (10). Next, in step n5, syllable recognition processing is performed,
In step n6, recognition result candidates are selected. That is, the syllable recognition unit 11 calculates the pattern distance between the feature pattern of each syllable and all the feature standard pattern memory 12 stored as standards in advance, and presents candidates in descending order of similarity frequency. Then, according to the labels stored in the feature standard pattern memory 16, candidates having the same type of syllable name are integrated and stored in the recognition result memory 13 as a syllable recognition result. Next, in step n7, errors in the syllable recognition results are corrected, and in step n8, the finalized syllable recognition results are derived. That is, the correction processing section 11a in the syllable recognition section 11 automatically corrects errors in the syllable recognition results using the dictionary stored in the language processing dictionary memory 14. Or the operator himself can use the keyboard 1
5, a correct candidate may be selected from the recognition candidates for the input speech, or an error location may be directly corrected. The correct result determined in this way is converted into Kanji and output as a character string. For example, when the input voice is "kaiwa", the pattern matching shown in Table 7 is performed, and the pattern matching shown in Table 8 is performed.
The same syllables shown in the table are combined, the correction processing shown in Table 9 is performed, and the final character string "Kaiwa" is
is output.

【table】

【table】

[Table] Next, learning processing is performed at step n8, and the processing ends at step n9. In addition, except for the voice analysis section 6, all CPUs are 10.
controlled by. FIG. 3 is a flowchart showing a more detailed procedure of the standard pattern learning process in step n8. After the syllable is determined, the process moves from step n8 to step m1 and step m2, where the syllable recognition unit 1
The learning control unit 11b in 1 learns the feature pattern of each input syllable using information such as recent correct/incorrect trends (this means updating the feature standard pattern using the feature pattern). Determine whether or not. For example, when comparing the automatically corrected syllable string ``kaiwa'' with the syllable recognition results, the syllable ``i'' is incorrectly recognized as ``pi.'' Next, a case will be shown in which it is determined that the characteristic pattern of the input syllable "i" is to be learned. So step
In m3, the update pattern limiting unit 11c in the syllable recognition unit 11 uses the same syllable label as the input syllable and the same environment information as the phonological environment from which the input syllable was extracted (stored in the feature standard pattern memory 18).
The update target is limited to feature standard patterns with . For example, the phoneme immediately before the syllable is shown in Table 108.
The class number (using the lower three bits of the monosyllable 8-bit code) is stored as syllable environment information.

[Table] "*": Silence (word-initial or consonant "tsu") "N": Decal sound "n" The vowel immediately before the input syllable "i" is "a",
The syllable environment information is 1 from Table 10. Here, the structure and contents of the M feature standard patterns are shown in Table 11.

【table】

[Table] From Table 11, there are three feature standard patterns to be updated: P7, P11, and P13. In step m4, the updating unit 11b in the syllable recognition unit 11 replaces the worst pattern among the selected standard feature patterns with the feature pattern of the input syllable according to the first equation, or uses the feature pattern to replace the worst pattern with the feature pattern of the input syllable. The averaging operation is performed using the formula 2, and the update of the feature standard pattern is completed. P'11=Pin...(1) P'11=(P11+Pin)/2...(2) Here, Pin indicates the input feature pattern, and P'11
indicates a new feature standard pattern when the feature standard pattern P11 is updated. For example, the pattern P11 with the smallest value of the recognition counter (stored in the feature standard pattern memory 16) indicating the degree of recognition contribution (that is, the smallest contribution to recognition) is updated as shown in Table 12. . At the same time, the value of the recognition counter is reset.

[Table] When the updating of one syllable is completed in this way, it is determined in step m5 whether or not there is a next syllable, and if there is a next syllable, the process returns to step m1. In this way, the operations from step m1 to step m5 are repeated to update the input syllable. If there is no next syllable in step m5, the process moves to step n10 and the process is completely completed. Note that everything except the voice analysis section 6 is controlled by the CPU 10. In the present invention, if the number of feature standard patterns for a type of syllable is very large, such as several dozen or more, it is thought that the effect will be even greater if the syllable environment class is set in detail. In the above embodiment, the immediately preceding phoneme is used, but the syllable environment may be set using the preceding and following phonemes. FIG. 4 conceptually shows the difference in the distribution of feature standard patterns after updating a certain syllable. In the present invention, patterns are arranged for each environment and distributed sparsely in the phonological distance space, as shown in FIG. It can be said that it has a vine distribution. In FIG. 4, numbers 1 to 6 indicate standard pattern numbers, and symbols a to e indicate syllable environments. Effects As described above, the present invention provides the following effects. (1) Feature standard patterns are well-organized using information defined as the environment before and after a syllable by associating the physiological actions of humans who perform similar articulation with phonemes in speech. is efficiently corrected and updated, ultimately resulting in a high recognition rate (high-accuracy recognition). (2) The final recognition accuracy can be quickly reached without being strongly constrained by the input frequency of syllables (high convergence).

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a Japanese speech input device 1 according to an embodiment of the present invention, FIG. 2 is a flowchart showing the procedure of speech recognition processing in the Japanese speech input device 1, and FIG. is the step
FIG. 4 is a flowchart showing a more detailed process procedure of the n8 standard pattern learning process, and is a diagram conceptually showing the difference in the distribution of feature standard patterns after updating a certain syllable. 1...Japanese speech input device, 2...Microphone, 6...Speech analysis section, 7...Syllable segmentation section, 8...Feature pattern buffer, 9...
Feature pattern memory, 10... CPU, 11...
Syllable recognition unit, 12, 16, 18... Feature standard pattern memory, 13... Recognition result memory, 14...
Language dictionary memory, 15...keyboard.

Claims (1)

[Scope of Claims] 1. Recognizes the input voice in syllable units by calculating the similarity with a plurality of types of syllable feature standard patterns registered in advance, and compares the results with a dictionary or with an external device such as a keyboard. In a speech recognition method that obtains the final input by modifying it through instruction operations, for each feature standard pattern of a syllable, the phonological environment of at least one side before and after the feature pattern was extracted from the speech is calculated. Add the syllable feature pattern obtained by analyzing the voice uttered at the time of input as a standard feature pattern, or modify the standard feature pattern using that input feature pattern and register it at the time of input. A speech recognition method characterized in that, when an input feature pattern is extracted, an update operation is performed by additional correction on a standard feature pattern that has the same environment information as the phonological environment on at least one side before and after the input feature pattern is extracted. .