JPH01285996A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To exactly extract the syllable sections of speech input words at the time of recognition of the speech input words by comparing the acoustical characteristics of the speech input words and the acoustical characteristics of the candidate for recognized words and executing correspondence of the time base. CONSTITUTION:There are two modes; a registration mode and a recognition mode. The syllable sections of the speech input words from a user are extracted from the acoustical characteristics of said speech input words and the syllable standard patterns are formed for each of the syllable sections and are registered in a standard pattern memory part 5, in the registration mode. The syllable sections of the speeches pronounced by the user are extracted and the distance calculation from the previously registered syllable standard pattern is executed for each of the syllable sections, then the speech input words are recognized in accordance with the calculated distances and the results of the recognition are outputted, in the registration mode. The time series of the acoustical characteristics of the candidates for the recognized words stored in a recognized word memory part 6 is corresponded to the time series of the acoustical characteristics of the speech input words in such a manner. The syllable sections of the speech input words are thereby correctly extracted.

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to a speech recognition device that recognizes input speech by dividing it into syllables.

<Prior art> As a method for recognizing words for a specific speaker, features representing the entire word for one word are stored in a storage unit, and speech input words are recognized based on the features of the entire word. This method is common. However, in the case of this method, it is necessary to register the feature amounts of all words, and it is not suitable for recognizing large words (several hundred to several thousand). Furthermore, it is not easy to change recognized words that have been registered. On the other hand, in a method that stores syllable feature quantities in a storage unit and performs recognition by dividing the voice input word into syllables based on the syllable feature quantities, the number of syllables for which special @ quantities are stored is small. Even when inputting all of Japanese, it is possible to recognize the amount of human speech by simply registering the features of about 100 types of syllables. Furthermore, by registering the feature amount of the -.variant syllable, changes in the recognized word can be handled simply by changing the notation of the kana character string of the word dictionary HY.

Conventional speech recognition devices that recognize speech input words by dividing them into syllables divide the speech input words into syllables, and compare the characteristic patterns of the divided syllables with syllable standard patterns registered in advance by the user. It calculates the distance and then uses the distance to recognize the voice input word. Then, when dividing the voice input word into syllables, it asks the user to say j-. The syllable section is extracted only based on the acoustic features of the input word 1.

<Problems to be Solved by the Invention> A voice recognition device that decomposes a voice input word into syllables and recognizes the voice input word based on the distance between the decomposed syllables and a syllable standard pattern, The accuracy of syllable segment extraction when dividing speech input words into syllables does not have a large effect on the accuracy of word recognition. this is,
This is because if a syllable is recognized based on an incorrect syllable interval, the recognized syllable candidates will not include correct syllables, and therefore the recognized word candidates will not include correct candidates.

By the way, in the conventional speech recognition device described above, when extracting the syllable section of the speech input word, the syllable section is extracted using only the acoustic features of the speech input word. There is a problem in that the frequency of incorrect extraction is high, and therefore the recognition accuracy of speech input words is low.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech recognition device that can correctly extract syllable sections of speech input words.

<Means for Solving the Problems> In order to achieve the above object, the present invention registers a syllable standard pattern in advance by a speaker, divides a voice input word input by voice into syllables, and divides the divided syllables into syllables. a feature extractor for extracting acoustic features of the speech input word; a recognition word storage unit that stores recognition words used when recognizing voice input words;
a feature generation unit that generates acoustic features regarding recognition word candidates selected from recognition words stored in an i-word storage unit; an acoustic feature of the speech input word extracted by the feature extraction unit; "generated by the feature generator"
A syllable segment extractor that compares the acoustic features of bipedal recognition word candidates to create a temporal axis correspondence, and extracts a syllable segment of the voice input word based on the association, and a syllable segment extraction unit that extracts the syllable segment. The syllable interval of the voice input word and "−
The present invention is characterized by comprising a distance calculation unit that calculates the distance from the two-syllable standard pattern and calculates the average value of the calculated distance for the recognition word candidates.

<Effect> 14 = When recognizing a voice input word input by voice,
Acoustic features of the speech input word are extracted by a feature extractor. Furthermore, select the voice input word from among the recognized words stored in the recognized word storage unit. Accordingly, one of the preselected recognition word candidates is selected, and acoustic features related to the word are generated by the feature generation section. Then, the syllable interval extraction unit compares the acoustic features of the speech input word extracted by the feature extraction unit and the acoustic features of the recognition word candidate generated by the feature generation unit, and A correspondence is made, and a syllable section of the audio input word is extracted based on the correspondence. Then, the distance calculation section calculates the distance between the syllable section of the voice input word extracted by the syllable section extraction section and the syllable standard pattern written at -,
An average value of the obtained distances for the recognized word candidates is calculated.

Therefore, the syllable section of the voice input word can be extracted more accurately than when the syllable section is extracted only by the acoustic features extracted by the feature extraction section.

<Example> Hereinafter, the present invention will be explained in detail with reference to illustrated embodiments.

FIG. 1 is a block diagram of the speech recognition device of the present invention, in which 1 is a microphone for inputting voice, 2 is an amplifier for amplifying only the voice band of the voice input from microphone 1, and 3 is from amplifier 2. Special @ used for distance calculation when extracting syllable sections from input speech waveforms and recognizing speech input words.
4 is the feature extraction unit that calculates the village, and 4 is the entire speech recognition device.
1 a CPU (central control unit) to control; 5 a standard pattern storage unit that stores syllable standard patterns; 6 a recognition word storage unit that stores recognition words used when recognizing voice input words;
7 generates an acoustic feature related to the recognized word from the character string of the recognized word in the recognized word Hotaru stored in the recognized word storage unit 6, and generates a syllable section of the input voice so as to match the generated acoustic feature at most. A top-down segmentation section for extracting; 8 a distance calculation section for calculating the distance between the extracted syllable section of the audio input word and the syllable standard pattern stored in the standard pattern storage section 5; 9 a syllable standard pattern registration operation and a syllable working memory used during recognition operations;
(2) 0 is an I10 interface for exchanging data with external devices (not shown) (display section, keyboard section, host CPU, etc.).

The speech recognition device of the present invention has two main modes: a registration mode and a recognition mode. The registration mode is a mode in which a syllable section of a voice input word is extracted from the acoustic characteristics of the voice input word from the user, a syllable standard pattern is created for each syllable section, and the created standard pattern is registered in the standard pattern storage unit 5. . In addition, as will be detailed later, the recognition mode extracts syllable sections from the voice uttered by the user, and calculates the distance between each extracted syllable section and the syllable standard pattern registered in advance using the registration mode. This mode recognizes the voice input word based on the calculated distance and outputs the recognition result. In this way, performing voice recognition using the registration mode and the recognition mode is a common voice recognition method.

17 = FIG. 2 is a diagram of the recognition mode described above.

Hereinafter, the operation of the recognition mode of the speech recognition apparatus having the above configuration will be explained using a specific example according to the diagram of FIG.

In step Sl, the voice input from the micropon 1 is amplified by the amplifier 2 and input to the feature extraction section 3. Then, the feature extraction unit 3 extracts the phoneme sequence 5 for each frame of the input speech.
Power series and other features used for distance calculations (e.g.
A time series of acoustic features such as 16-channel filter bank output) is calculated.

In step S2, in addition to the above-described processing for the input speech, several recognition words are selected according to the speech input word by preliminary selection, for example, by DP matching of phoneme strings, from among the recognition words stored in the recognition word storage unit 6. A candidate is selected.

Then, from the character string of one of the selected recognition word candidates, the top-down segmentation unit 7 generates an expected acoustic feature for this recognition word candidate.

In step S3, it is determined whether or not processing has been completed for all of the recognition word candidates selected in step S2. As a result, if it has been completed, the process advances to step S8,
Otherwise, the process proceeds to step S4.

In step S4, syllable sections of the input voice are extracted and divided into syllables.

This syllable division is performed as follows. That is, the time series of the acoustic features of the input speech calculated in step Sl above is compared with the time series of the expected acoustic features of the recognition word candidates generated in step S2, and the generated expected acoustic features are The acoustic features of the input speech are associated in a manner that best matches the time series of the input speech. Then, a time-series change point of the acoustic features of the input speech that corresponds to a time-series change point of the predicted acoustic features is determined as the optimal syllable interval of the input speech. As a method for finding consistency between the time series of acoustic features and features of the input speech described above and the time series of expected acoustic features of recognition word candidates, there is an algorithm that takes nonlinear temporal correspondence without causing a reversal of the time axis. There are methods such as DP matching.

In step S5, the distance calculation between the syllable section of the input voice extracted in step S4 and the syllable standard pattern stored in the standard pattern storage unit 5 in the two registration motes is performed to calculate the total number of currently recognized word candidates. This is performed by the distance calculation unit 8 for all syllables.

At this time, only the degree of similarity between the current syllable of the voice input word and the currently recognized word candidate is determined, so it is only necessary to determine the distance between the syllable patterns of the syllables that compare the currently recognized word candidate.

That is, for example, suppose that the current recognized word candidate is "red". Voice input jlj.

For the first syllable section of a word, it is only necessary to find the distance from the syllable standard pattern /a/.

In step S6, the sum of the distances for all syllables of the currently recognized word candidate obtained in step S5 is divided by the number of syllables of the currently recognized word candidate? The average syllable distance of the input speech to the currently recognized word candidate is calculated.

In step S7, the average syllable distance of the currently recognized word candidate calculated in step S6 is stored in, for example, the working memory 9, and then the process returns to step S2.

In step S8, the average syllable distances of all recognized word candidates calculated and stored as described above are sorted in descending order of average syllable distance.

In step S9, the recognized word candidate with the smallest average syllable distance is output from the I10 interface 9 as a recognition result.

FIGS. 3 and 4 show extracted syllable sections and syllable distances when two words "Kashima" and "1 and Kushima" are selected as recognition word candidates.

FIG. 3 shows the power of the input speech and its extracted syllable sections.

The top row is a time series of the power obtained when the user utters /Tokushima/. The first column in the lower row is the optimal syllable interval obtained when the input word is "Kashima", that is, when "Kajima J" is used as a recognition word candidate. This is the optimal syllable interval obtained when "Tokushima" is used as a candidate. In addition, the lower column (■) is a syllable section obtained only from the acoustic features of the input speech without using recognition word candidates for the speech input word "Tokushima". In the case of these 11 columns, since the syllable /</ of the input voice "Tokushima" is short, the next syllable / /
It is recognized as being merged with the word ``,'' and it is incorrectly divided into three syllables.

Figure 4 shows the syllables and standard patterns of the input speech word divided as shown in columns 1 and 2 of Figure 3 when "Kashima" and "Tokushima" are selected as recognition word candidates. Memory section 5? The distance M from this stored syllable standard pattern (distance determined in step S5 of the flowchart in FIG. 2) is shown. At the same time, the average syllable distance for each recognized word candidate based on this distance (the distance determined in step S6 of the flowchart in Figure 2) is also shown.

That is, in the case of this embodiment, the recognition word candidate "Kikushima j" with a small average syllable distance is output as a recognition word.

In this way, several recognition word candidates (for example, "Kashima" and "Tokushima" in FIG.
), and by associating the time series of the acoustic features of this recognition word candidate with the time series of the acoustic features of the speech input word, the syllable interval of the speech input word is extracted. As shown in the figure, the input audio word "Tokushima" is not erroneously divided into three syllable sections as shown in the ■ column. That is, a three-syllable word obtained based on three syllable sections that are erroneously divided, such as a double-syllable, is not output as a recognized word.

Therefore, according to the present invention, input speech can be divided into accurate syllable sections and speech input words can be correctly recognized.

In the above embodiment, when calculating the distance between the syllable interval of the voice input word and the syllable standard pattern, the syllable interval is extracted for each recognition word candidate, and the distance of the syllable standard pattern 7 is calculated for all the syllable intervals. However, the present invention is not limited to this. In other words, when extracting the syllable interval, the syllables /ka/ and syllables /ma/to of the recognition word candidate "Kashima" in Figure 3, and the recognition word candidate [
Tokushima] corresponds to the syllable /to/ and the syllable /ma/,
The same sections A and 13 of the audio input word are extracted as syllables. Therefore, the distances between the syllable standard patterns /ka/ and / and / are calculated at once to the syllable intervals in the voice input, and the distances between the syllable interval I3 and the syllable standard patterns /ma/ and /ma/ are calculated at once. Times may be omitted.

In addition, in this implementation (1+jl), when generating the P-like acoustic features of the recognition word candidates, they may be generated according to a predetermined rule that is stored in advance, or the character string and its acoustic characteristics may be generated for each recognition word. The time series of the acoustic features may be stored together in the recognized word storage unit 6, and the acoustic features may be generated in accordance with the stored time series of the acoustic features.

<Effects of the Invention> As is clear from the above, the speech recognition device of the present invention has the following effects:
When recognizing a voice input word, the acoustic features of the voice input word are compared with the acoustic features of the recognition word candidates, and the syllable intervals of the voice input word are extracted by comparing the acoustic features of the recognition word candidates and The voice input word is recognized based on the average value of the distance between the identified syllable section and the syllable standard pattern for the recognized word candidates.

Therefore, according to the present invention, it is possible to accurately extract the syllable section of a speech input word during speech input word recognition, and highly accurate speech recognition can be performed in accordance with this correct syllable section.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the speech recognition device of the present invention, FIG. 2 is a flowchart of syllable recognition operation in recognition mode,
Fig. 3 shows an example of the power of an input speech word and its extracted syllable section, and Fig. 4 shows an example of the distance between the syllables of the divided speech input word and the syllable standard pattern and the average syllable distance of recognition word candidates. FIG. 1. Microbon, 2. Amplifier, 3. Feature extraction section, 4. CPU, 5. Standard pattern storage section, 6. Recognition word storage section, =15- 7. Top-down segmentation section, 8. Distance. Calculation unit, 9... working memory, IO... I10 ink file.

Claims (1)

[Claims] (1) A syllable standard pattern is registered in advance by the speaker,
Divide the voice input word input by voice into syllables,
In the speech recognition device, which calculates the distance from the syllable standard pattern for each divided syllable and recognizes the speech input word based on the calculation result, a feature extraction unit extracts acoustic features of the speech input word. a recognition word storage unit that stores recognition words used when recognizing the voice input word; and generating acoustic features regarding recognition word candidates selected from the recognition words stored in the recognition word storage unit. a feature generation unit, which compares the acoustic features of the voice input word extracted by the feature extraction unit and the acoustic features of the recognition word candidates generated by the feature generation unit, and correlates the time axis; a syllable interval extraction unit that extracts syllable intervals of the audio input word based on the correspondence; and a syllable interval extraction unit that calculates the distance between each syllable interval of the audio input word extracted by the syllable interval extraction unit and the syllable standard pattern. A speech recognition device comprising a distance calculation section.