JPH02298996A - Word voice recognition device - Google Patents

Abstract

PURPOSE:To improve the recognizing ability by detecting the characteristic pattern of the high-reliability voice section candidate of an input word voice and generating a high-performance standard pattern. CONSTITUTION:A word border temporary generation part 2 sets plural voice section candidates for the feature parameter of the input word voice from an acoustic analysis part 1. Then the voice section candidate found from a combination of head-end and tail-end candidates is detected. Then a similarity arithmetic part 7 extracts and supplies feature patterns having the largest similarity to a standard pattern generation part 8 when their combination is coincident between voice inputs and the generation part 8 generates and registers a standard pattern as to a word to be recognized in a pattern dictionary 5. Then the similarity between the feature pattern of the input word voice and a word to be recognized which is registered in a dictionary 5 is calculated 4. Then a recognition result output part 6 compares the similarity to each other to output the object of recognition having the largest similarity.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a word speech recognition device that can efficiently and highly accurately recognize input speech words.

(Prior Art) Speech recognition technology plays an important role in realizing an excellent man-machine interface. In this speech recognition, speech section detection is an important preprocessing for improving the recognition accuracy, and various research and developments have been carried out in the past.

Conventionally, this speech section detection generally involves obtaining a power temporal sequence of input word speech, and setting the speech power P to a predetermined threshold value Tl.
This is carried out by detecting the point in time when the voice power P becomes smaller than a predetermined threshold T2 as the start point S of the input voice word, and detecting the point in time when the voice power P becomes smaller than a predetermined threshold T2 after detecting this voice start point as the end point E of the input voice word. be exposed.

However, in this type of speech section detection, the speech section is uniquely determined, so if there is, for example, breath leakage or tongue clicking noise before and after the actual speech section, this will also be detected as part of the speech section. There is a problem with it being stored away. On the other hand, in the case of word sounds where the beginning or final syllable of a syllable is easily devoiced, the power P of the devoiced syllable part becomes extremely small, so there is a drawback that this part is likely to be dropped from the detected syllable section. .

Such a detection error in a speech section causes a fatal misrecognition or recognition rejection in speech recognition. In particular, the characteristic patterns of the word sounds uttered by the device user are registered as standard patterns, and the degree of similarity between the characteristic patterns of the word sounds input during speech recognition and the previously registered standard patterns is determined and recognized. In the specific speaker word speech recognition device that obtains the results, if a speech segment detection error occurs when creating a standard pattern, the feature pattern of a speech segment different from the speech segment of the word to be registered is registered as the standard pattern. It will end up being put away. If an incorrect standard pattern is registered in this manner, even if the speech section is correctly detected during speech recognition, an incorrect recognition result will be output.

As a technology to reduce recognition errors caused by speech interval detection errors when creating a standard pattern, the word to be recognized is uttered multiple times when creating a standard pattern, and only when the utterance times are approximately equal, the standard There is a method of creating and registering a pattern and requesting re-voice if there are variations in the utterance time. However, if the same noise is added or the voice section is dropped in all of the plurality of utterances, there is a problem in that the above-mentioned comparison of the utterance sections alone cannot prevent erroneous detection of the voice section.

In addition, as a method to reduce the addition of noise and dropout of speech sections, multiple starting point candidates (S1.
S2. ~SM) and termination candidates (E 1.E 2.~EN
) and calculate the probabilities (f sl, f s2.~f sM) for these candidates according to predetermined rules.
.

(f el, f e2. to f eN) are given, respectively. Then, a plurality of speech interval candidates [SIl, En] obtained as a combination of the above-mentioned start end candidates (Sl, S2.~S4) and end end candidates (El, E2.~EN) (however, m=1°2+~M, n −1+2.~N) is Lmn = fsm
A method has been proposed as the word boundary hypothesis method that calculates the recognition result for the input word speech by calculating it as fen and matching it with a standard pattern for speech interval candidates with higher wave numbers with higher likelihoods. .

However, even if such a method is adopted, there is a problem in that if the speech section of the word speech obtained as a standard pattern is incorrectly detected, the input word speech cannot be correctly recognized.

(Problem to be solved by the invention) In this way, in the past, when creating a standard pattern,
Several problems remain as to how to obtain a highly reliable and highly accurate standard pattern that prevents detection errors in the voice section.

The present invention has been made in consideration of the above circumstances, and its purpose is to obtain a standard pattern by effectively preventing speech interval detection errors when creating a standard pattern, and to prevent speech interval detection errors. It is an object of the present invention to provide a word speech recognition device that can reliably and reliably recognize input speech words that are likely to occur.

[Structure of the Invention] (Means for Solving the Problems) The present invention acoustically analyzes input word speech to obtain its feature parameters, and uses the acoustically analyzed feature parameters to obtain the above-mentioned information in a word boundary hypothesis generation unit. The feature parameters are normalized for each speech section candidate of the input word speech to generate a feature pattern of the input speech, and the degree of similarity between this feature pattern and a standard pattern of the recognition target word speech created by the standard pattern creation section is calculated. The word speech recognition device calculates a word speech recognition device for the input word speech, wherein the standard pattern creation unit calculates a plurality of word speech recognition devices for the word speech inputted multiple times regarding the recognition target word. The likelihoods are mutually calculated between the feature patterns, and when the combination of feature patterns with the highest likelihood matches between the feature patterns in the plurality of utterances, that feature pattern is used for the speech of the target word to be recognized. This is characterized in that it is determined as a standard pattern.

(Function) According to the present invention having such a function, when determining the correct speech interval from among the plurality of speech interval candidates obtained by the word boundary hypothesis generation unit, the same word is uttered multiple times. The likelihood is mutually calculated between the plurality of feature patterns obtained for each input speech, and only when the feature pattern with the highest likelihood matches among the feature patterns of the plurality of utterances, those speech Since the segment candidate is determined to be the correct speech segment, the speech segment can be determined with high accuracy and reliability.

As a result, it is possible to obtain a highly reliable standard pattern and improve its recognition performance.

(Example) Hereinafter, an example of an implementation of the word speech recognition device according to the present invention will be described with reference to the drawings.

FIG. 1 is a schematic diagram of the main parts of the apparatus according to the embodiment, and 1 is an acoustic analysis section that acoustically analyzes input speech to obtain its characteristic parameters. The acoustic analysis unit 1 obtains a temporal sequence of speech power as a feature quantity to be used for speech section detection, and also obtains, for example, the output of a group of bandpass filters subjected to frequency analysis as a feature quantity to be used for comparison with a recognition dictionary.

The word boundary hypothesis generation unit 2 adaptively sets various speech interval detection parameters for the characteristic parameters of the input speech determined by the acoustic analysis unit 1, and sets a plurality of speech interval candidates. Specifically, the word boundary hypothesis generation unit 2 generates a plurality of starting point candidates (S l, S 2. to SM
) and a plurality of terminal candidates (El, E2. to EN), respectively, and the probability (f sl.) for each of these candidates is determined according to a predetermined rule.

fs2. ~fsM), (fel, fe2.~
feN) respectively.

Then, a plurality of voice section candidates [Sm, Enl (however,

s-1, 2, ~M, n-1, 2, ~N) as Lmn - fsm A plurality of high-ranking speech section candidates are detected in descending order of .

Then, the resampling unit 3 resamples the characteristic parameters consisting of the bandpass filter group outputs obtained by frequency analysis in the acoustic analysis unit 1 for the plurality of speech interval candidates [Sm, 'Enl obtained as described above. Then, normalized feature patterns of the input word speech in each speech section candidate are determined. Note that if only one speech segment candidate with a high likelihood is obtained, resampling processing is performed only on that speech segment candidate, and only one feature pattern is obtained.

The similarity calculation unit 4 calculates the similarity between the characteristic pattern of the input word sound obtained in this way and the standard pattern described later for the recognition target word registered in the standard pattern dictionary 5. . The recognition result output unit 6 compares the degree of similarity between the characteristic pattern of the input word sound obtained by the similarity calculation unit 4 and the standard pattern of the recognition target word, and selects a predetermined number of results that have obtained a high degree of similarity. The category name of the recognition target word and its similarity value are determined respectively. Then, these top 1fi category names are output as recognition candidates for the input word voice,
Alternatively, the category name of the recognition target word with the highest similarity value is output as the recognition result for the input word speech.

By the way, the standard pattern for the recognition target word registered in the standard pattern dictionary 5 is generated by the similarity calculation section 7 and the standard pattern creation section 8, which are provided one more.

The similarity calculating section 7 uses a plurality of characteristic patterns respectively obtained by the resampling section 3 described above when the speech of a recognition target word with a known category name is uttered multiple times when creating a standard pattern. Calculate the degree of similarity between multiple spoken input word sounds. This similarity calculation is performed using various techniques such as the subspace method and the DP matching method that have been proposed in the past.

Therefore, when the combination of feature patterns with the maximum similarity matches between the plurality of uttered input word sounds, the similarity calculation unit 7 extracts the feature pattern and provides it to the standard pattern creation unit 8. .

Note that if the combination of feature patterns with the maximum similarity does not match among the plurality of input word sounds, re-input of the word sound is prompted.

The standard pattern creation unit 8 sets the characteristic pattern of the input word sound obtained in this way as a standard pattern for the recognition target word, and registers it in the standard pattern dictionary 5 with its category name attached.

To explain in more detail how to create this standard pattern, when creating the standard pattern, known word sounds of the category name are uttered and input multiple times.

These input word sounds are acoustically analyzed by the aforementioned acoustic analysis section 1 to obtain their characteristic parameters, and the word boundary hypothesis generation section 2 obtains probable speech section candidates. Then, the resampling unit 3 obtains characteristic patterns for each of these voice section candidates.

Specifically, as shown in FIG. 2, the speech section candidates Lll and L12 are obtained at the first utterance, and these speech section candidates L11. Characteristic pattern pH at L 12.

Find each PI3. Similarly, when speaking for the second time, the voice section candidate L 21. The characteristic pattern P21 for L22. P 22 is calculated respectively, and the voice section candidate L 31. is also calculated for the third utterance.
Regarding L 32, its characteristic pattern P 31.
Find P32 respectively.

Here, two speech section candidates are obtained for each input word speech and their characteristic patterns are obtained, but characteristic patterns may be obtained for three or more speech section candidates. Furthermore, if a sufficient result cannot be obtained with three utterances, it is of course possible to urge the user to utter four or more times.

The similarity calculation unit 7 mutually calculates the degree of similarity (likelihood) between the plurality of characteristic patterns of each utterance input voice obtained in this way, and selects the combination of characteristic patterns that has the maximum degree of similarity as described above. Check whether or not there is a match between multiple spoken input word sounds. For example, the degree of similarity Q between the feature patterns of the first and second utterance input voices described above is calculated as follows.

Qll-[pH-P21], Q12 bound [pHΦP2
2] Q13-[PI3・P21], Q14-[P
12φP22] and a combination of feature patterns that takes the maximum value among these similarities is determined.

After that, the similarity Q between the feature patterns between the first and third utterance input voices is calculated as Q21- [pH
・P31], Q22- [pH-P32]Q23-
[PI3・P31], Q24- [PI3・P32
], and find the combination of feature patterns that takes the maximum value among these similarities.

Similarly, the degree of similarity Q between the feature patterns between the second and third utterance input voices is determined by Q31- [P21-
P31], Q32- IP21- P32] Q3
3-[P22・P31], Q34- [P22・P
32], and find the combination of feature patterns that takes the maximum value among these similarities.

Through such processing, when a feature pattern combination with the maximum similarity between each feature pattern matches among multiple vocal input voices, the corresponding feature pattern for each of these vocal input voices is Extract.

For example, Qll-[pH-P21] is the highest among the similarities between the first and second utterance input voices, and
Q22- [P
ll-P32] When the maximum similarity is reached, it is required that the combination of feature patterns that have obtained the maximum similarity matches the feature pattern pH of the first utterance. 2 more
Q32 in the similarity between the utterance input voice for the first time and the third time
- When it is detected that [P21-P32] has reached the maximum, the combination of these feature patterns becomes the feature pattern P21 of each voice input voice described above. P 3
2 is required.

When it is confirmed that there is no contradiction between the combination of feature patterns with maximum similarity and the feature pattern of each utterance, that is, the combination [pH-P2
1] [Pll・P 32] [P 21・P32
] is the characteristic pattern of each utterance P 11゜P21
．． Since they match P31, these are determined to be characteristic patterns in the correctly detected voice section.

When such a determination result is obtained, these characteristic patterns pH, P21. P31 are each extracted and provided to the standard pattern creation section 8. The standard pattern creation unit 8 creates these characteristic patterns P11. P 21゜P3
1 is averaged, etc., to create a standard pattern for the word to be recognized, and this is registered in the standard pattern dictionary 5.

Note that [PI3・
P 21] Similarity Q1 of the combination of feature patterns
3 takes the maximum value, and the similarity Q22 of the combination of feature patterns [Pll-P32] takes the maximum value between the first utterance and the third utterance, then the feature pattern that takes these maximum values is Since the combination does not match the feature pattern of the first utterance (different feature patterns are required for P11 and PI3), it was determined that it is inappropriate as a feature pattern for creating a standard pattern. , prompts you to input the word audio again.

Also, for each of the first and second utterances, for example, Qll-
[pH-P21] is the maximum, Q22-[pH・P32] is the maximum for the first and third utterances, and the combination of feature patterns that obtains the maximum similarity is 1
Even if it is required to match the characteristic pattern Pit of the first utterance, for the second and third utterances,
For example, when Q32-[P22·P32] is the maximum, a contradiction occurs due to the mismatch between the characteristic patterns P21 and P22 of the second utterance, and therefore, in this case as well, re-input of the word voice is prompted.

In this way, this device calculates the degree of similarity (likelihood) between feature patterns obtained for each input word voice of the same category multiple times (three or more times), and determines the combination of feature patterns that takes the maximum value. When there is no inconsistency between the feature patterns of each utterance, this is extracted as a feature pattern that has been correctly detected in a voice section. Then, a standard pattern is created based on the characteristic pattern when the speech section is correctly detected, and is registered in the standard pattern dictionary 5 to be used in word speech recognition processing.

Therefore, according to the present device, the standard pattern itself can be used as a characteristic pattern when the speech section is correctly detected, so that the recognition dictionary performance can be made sufficiently high. As a result, the recognition performance can be made sufficiently high.

In addition, as mentioned above, the speech sections of the input speech are evaluated as combinations of feature patterns based on the similarity between the feature patterns, and only the feature patterns in the correct speech sections are extracted to create a standard pattern, so the process The procedure is very simple and processing efficiency is high. Moreover, effects such as being able to effectively create the standard pattern with a small number of utterances can be achieved.

Note that the present invention is not limited to the embodiments described above. For example, the similarity (likelihood) may be calculated between the feature patterns of four or more uttered input voices to extract the correct feature pattern of the voice section. Furthermore, if the similarity (likelihood) between feature patterns does not reach a predetermined threshold, a standard pattern may be created by prompting re-voice input even if there is no contradiction in the combination. .

In this way, it becomes possible to increase the reliability of standard pattern creation and to construct standard patterns (recognition dictionaries) with higher performance. In addition, the present invention can be implemented with various modifications without departing from the gist thereof.

[Effects of the Invention] As explained above, according to the present invention, the speech sections of input speech are correctly detected by eliminating the effects of noise addition and omission of speech sections, and the characteristic patterns thereof are extracted. It is possible to efficiently create a high-performance standard pattern using only the feature patterns of voice section candidates with high performance, and it has great practical effects such as being able to effectively improve its recognition performance. It will be done.

[Brief explanation of drawings]

FIG. 1 is a schematic configuration diagram of a word speech recognition device according to an embodiment of the present invention, and FIG. 2 is a schematic diagram of speech interval candidates and their characteristic patterns for multiple input speeches when creating a standard pattern in the embodiment device. % expression% 3... Resample unit, 4... Similarity calculation unit (similarity calculation between feature pattern and standard pattern), 5... Standard pattern dictionary, 6... Recognition result output unit, 7 ... Similarity calculation unit (likelihood calculation processing between feature patterns), 8...
・Standard pattern creation department.

Claims (1)

[Scope of Claims] An acoustic analysis section that acoustically analyzes input word speech to obtain its feature parameters; a word boundary hypothesis generation section that obtains speech section candidates of the input word speech from the acoustically analyzed feature parameters; a resampling unit that normalizes the feature parameters for each speech section determined by the word boundary hypothesis generation unit to generate a feature pattern of the input speech; A standard pattern creation unit that creates a pattern, and a similarity calculation that calculates the degree of similarity between the standard pattern of the recognition target word audio created by the standard pattern creation unit and the characteristic pattern of the input word audio obtained by the resample unit. Department and
and a recognition result output section that obtains a word speech recognition result for the input word speech according to the similarity calculation result, and the standard pattern creation section is configured to perform a recognition result output section for each word speech that is uttered multiple times for the recognition target word. A means for mutually calculating the likelihood between a plurality of desired feature patterns, and a means for calculating the feature pattern when the combination of the feature pattern with the highest likelihood matches among the feature patterns in the plurality of utterances. A word speech recognition device comprising: means for obtaining a standard pattern of the speech of the word to be recognized.