JPH0458638B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Field of Industrial Application> The present invention relates to a speech recognition device that recognizes speech such as Japanese in syllable units and outputs it to an external device.

<Prior art> Conventionally, speech recognition devices determine syllable intervals by classifying input speech into phonemes, cut out the syllables based on the judgment results, and perform recognition in units of the cut out syllables. is going on.

<Problems to be Solved by the Invention> However, in the above-mentioned conventional speech recognition device, since the duration of the speech spectrum of the syllabic sound (“n”) is relatively long and the power is small, As shown in FIG. 3, even though it is a single syllable section, a frame that is determined to be a buzz bar (a frame with the phoneme classification label B indicated by A in FIG. 3) occurs within that syllable section, and is determined to be a buzz bar. There is a problem in that the syllables are classified and extracted as two syllables, one before and one after, with the frame as the boundary.

Here, the busbar is a sound produced by the vibration of the vocal cords before the syllable of a voiced plosive, and is a sound with a very strong low frequency component.
Almost inaudible to humans. In addition, the above frame is
In order to analyze the audio spectrum of the input audio, this is one section in which the audio spectrum is divided into approximately 16 ms.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a speech recognition device that does not erroneously separate and recognize a syllable as two syllables even though the syllable is a single syllable.

<Means for Solving the Problems> In order to achieve the above object, the speech recognition device of the present invention cuts out syllable sections from input speech, and stores feature patterns of the cut out syllables in advance in memory. A speech recognition device that recognizes input speech syllable by syllable by calculating the degree of similarity between the stored characteristic reference patterns of multiple types of syllables.
A syllabic sound discrimination section that determines whether the recognition result of the extracted syllable is a syllabic sound, a feature extraction section that extracts feature information representing the characteristics of the extracted syllable, and a syllabic sound discrimination section that When the extracted syllable is recognized as a syllable, the result of comparing the feature information extracted by the feature extraction section with a predetermined value and the pre-stored syllable immediately before the extracted syllable are compared. Based on the identification result indicating whether the syllable is a syllable or not,
A determination unit that determines whether or not the above-mentioned consonant is erroneously recognized as two syllables even though it is one syllable; The present invention is characterized in that it is provided with a plucked sound rejection section that rejects one of the two syllables.

<Operation> The syllables cut out from the input speech are recognized syllable by syllable, and the percussion discrimination section determines whether or not the recognition result is a percussion.

As a result, if it is determined that the syllable is a syllable, the result of comparing the feature information extracted by the feature extraction unit with a predetermined value and the syllable immediately before the extracted syllable that is stored in advance are Based on the identification result indicating whether or not it is a phrasal sound, the determining unit determines whether the phrasal sound is a phrasal sound that is mistakenly recognized as two syllables even though it is one syllable.

Then, based on the judgment result of the judgment section, 2
One of the two syllables that was mistakenly recognized as a syllable is rejected by the syllable rejection unit, resulting in the 2 syllables being mistakenly recognized as 2 syllables.
without being recognized as a syllable separated into syllables,
It is recognized as exactly one syllable.

<Examples> Hereinafter, the present invention will be explained in detail by referring to illustrated examples.

In FIG. 1, 1 is an analog input section,
It consists of an amplifier (AMP) 11 that amplifies continuously input audio, and an A/D converter 12 that converts the output of the AMP 11 into a digital signal. 2 is a speech analysis section, which performs spectrum analysis by dividing the digital input speech into frames of approximately 16 ms, and then passes through the syllable section extraction section 3 at intervals of approximately 8 ms.
The characteristic pattern of the input voice and information necessary for cutting out syllable sections (for example, power, number of zero crossings, etc.) are transferred to the input voice. The syllable section extraction section 3 extracts syllables from the characteristic pattern transferred from the speech analysis section 2 based on the information necessary for cutting out the syllable section transferred from the speech analysis section 2, and extracts the syllables from the characteristic pattern transferred from the speech analysis section 2. The feature pattern of the extracted syllable (for one syllable) is stored in the feature pattern memory 61 via the CCPU 5, and further includes a feature extraction section, and while exchanging data with the CPU 5, feature information is extracted from the cut out syllable. and stores it in the feature pattern memory 61, and notifies the CPU 5 that syllable extraction of the input voice has been completed. Note that the feature information used includes the length of the cut out syllable, the length of silent time between the cut out syllable and the immediately preceding syllable, and the amount of change in the speech spectrum at the boundary between the cut out syllable and the immediately preceding syllable.

The pattern memory 6 stores a characteristic pattern for one syllable extracted by the syllable section extraction unit 3,
It also has the aforementioned feature pattern memory 61 in which feature information of the syllables is stored, and a standard pattern memory 62 in which feature standard patterns of a plurality of syllables are stored. The warm syllable recognition unit 9
An example of similarity calculation between the feature pattern stored in the feature pattern memory 61 in the pattern memory 6 and the plurality of feature standard patterns stored in the standard pattern memory 62 according to a command from the CPU 5. The input voice is recognized by performing a certain distance calculation. and,
The obtained recognition result is returned to the CPU 5 and further sent to an external device via an interface (I/F) section 10. At the same time, it is stored in the recognition result storage memory 11.

Next, the operation of the speech recognition apparatus will be explained with reference to a flowchart of one syllable recognition process shown in FIG.

In step _S1 , it is determined whether the syllables of the input speech have been extracted by the syllable section extraction section 3. If the syllable extraction unit 3 has completed the syllable extraction, the process proceeds to step _S2 ; if not, the process returns to step _S1 .

In step _S2 , the CPU 5 instructs the monosyllable recognition unit 9 to calculate the degree of similarity. The monosyllable recognition unit 9 uses a feature pattern of one syllable (hereinafter referred to as the current syllable) cut out from the input speech, which is stored in the feature pattern memory 61 in the pattern memory 6, and a standard pattern. A similarity calculation is performed between the feature standard patterns of multiple syllables stored in the memory 62, and a recognition result of the current syllable is obtained.

In step _S3 , it is determined whether the recognition of the current syllable has been completed. As a result, if the process has been completed, the process proceeds to step _S4 ; if the process has not been completed, the process returns to step _S3 .

In step _S4 , it is determined whether or not the speech recognition result of the current syllable recognized in step _S2 is a syllable. As a result, if the sound is not a repellent sound, the process proceeds to step _S10 , and if the sound is a repellent sound, the process proceeds to step _S5 .

In step _S5 , the length of the current syllable (one of the characteristic information of the extracted syllable stored in the characteristic pattern memory 61 in the pattern memory 6) is compared with the set value L1. If it is greater than the set value L1, the current syllable compared is considered to be an independent syllable and the process proceeds to step _S10 , and if it is less than the set value L1, the process proceeds to step _S6 .

In step _S6 , the silent time length (or the time length determined as a buzz bar) between the current syllable and the immediately preceding syllable is compared with a set value L2. As a result, if it is greater than or equal to the set value L2, the current syllable compared is considered to be an independent syllable and the process proceeds to step _S10 , and if it is smaller than the set value, the process proceeds to step _S7 .

In step _S7 , the amount of change in the speech spectrum at the boundary between the current syllable and the immediately preceding syllable is compared with a set value V. The reason why one syllable of a syllable is recognized separately is mainly because some frames are checked as a buzz bar due to power fluctuations in the audio spectrum of the syllable of a syllable, as shown in Figure 3. This method takes advantage of the fact that there is no change in the voice spectrum, and if it is greater than the set value V, the current syllable being compared is considered to be an independent syllable, and the process proceeds to step _S10 , and if it is less than or equal to the set value V. If so, proceed to step _S8 .

In step _S8 , it is determined whether or not the speech recognition result of the syllable immediately before the current syllable compared in step _S7 , which is stored in the recognition result storage memory 11, is a syllable. As a result, the current syllable is considered to be an independent syllable if it is not a plitic sound, and step S ₁₀
If the current syllable is a plitic syllable, it should be recognized as one syllable together with the previous plitic syllable, but it is two syllables.
It is determined that it is one of the recognized syllables, and the process proceeds to step _S9 .

At step _S9 , the current syllable that has been misrecognized as two syllables is rejected, and the one-syllable recognition process is completed.

In step _S10 , the syllable recognition result of the current syllable recognized in step _S2 is stored in the recognition result storage memory 11, and is sent to an external device (not shown) via the I/F section 10. The syllable recognition process ends. Here, the above-mentioned step _S4 constitutes a repellent sound discriminating section, and the above-mentioned steps _S5 to _S8
will constitute the judgment section.

Therefore, according to the above-described procedure, only one recognition result can be output for a plucked sound that is erroneously recognized as being separated even though it is one syllable.

<Effects of the Invention> As is clear from the above, the speech recognition device of the present invention includes a phrasal sound discrimination section that determines whether the recognition result of the current syllable is a phrasal sound or not, and extracts feature information representing the characteristics of the current syllable. When the current syllable is determined by the feature extracting section and the syllabic sound discriminating section to be a syllabic sound, the result of comparing the feature information extracted by the feature extracting section with a predetermined value is stored in advance. a determination unit that uses the identification result of the syllable immediately before the current syllable to determine whether or not it is one of the syllables that are misidentified as two syllables even though the syllable is one syllable;
Based on the judgment result of the above-mentioned judgment section, we have provided a plucked sound rejection section that rejects one of the two syllables misidentified as a plucked sound of two syllables.
There is no misidentification of the syllable as a syllable, and the recognition performance for syllables is improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the speech recognition device of this invention, and FIG. 2 is a block diagram of the speech recognition device of this invention.
FIG. 3 is a flowchart of syllable recognition processing, showing an example of syllables that are erroneously recognized as separated. 1...Analog input section, 2...Speech analysis section, 3
...Syllable section extraction unit, 5...CPU, 6...Pattern memory, 9...Single syllable recognition unit, 11...Recognition result storage memory.

Claims (1)

[Claims] 1. A syllable section is cut out from the input speech, and the degree of similarity is calculated between the feature pattern of the cut out syllable and the standard feature pattern of a plurality of types of syllables stored in advance in a memory. Accordingly, in a speech recognition device that recognizes input speech in units of syllables, there is provided a phrasal discrimination unit that determines whether or not the recognition result of the extracted syllable is a pylase, and a feature that represents the characteristics of the extracted syllable. a feature extraction unit that extracts information; and when the extracted syllable is recognized as a rhyme by the rhyme discrimination unit, the feature information extracted by the feature extraction unit is compared with a predetermined value; Based on the result and the pre-stored identification result indicating whether or not the syllable immediately before the extracted syllable is a syllable, the syllable is mistakenly separated as two syllables even though the syllable is one syllable. The present invention is characterized in that it is provided with a judgment section that judges whether or not it is a phrasal sound that has been recognized as a phrasal sound, and a phrasal sound rejection section that rejects one of the two syllables that has been mistakenly recognized as a phrasal sound based on the judgment result of the judgment section. speech recognition device. 2. The characteristic information of the syllable for judgment by the judgment unit includes the length of the cut out syllable, the length of silence between the cut out syllable and the immediately preceding syllable, and the length of the cut out syllable and the preceding syllable. 2. The speech recognition device according to claim 1, which uses any one or a plurality of combinations of the amount of change in the speech spectrum at the boundary with the immediately preceding syllable.