JPS60147797A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition device for automatically recognizing a speech signal uttered by Iruma's voice.

Conventional configurations and their problems A speech recognition device that automatically recognizes speech is considered to be very effective as a means for providing data and commands from humans to computers and various machines.

The slam matching method is often adopted as the operating principle of speech recognition devices that have been researched or published in the past.

This method memorizes standard patterns for all types of words that need to be recognized in advance, and compares them with unknown input patterns to determine the degree of matching (hereinafter referred to as similarity). The word is calculated and determined to be the same word as the standard batan that yields the maximum match. In this batan matching method, a standard batan must be prepared for every word to be recognized, so if the speaker changes, it is necessary to input and memorize a new standard batan. Therefore, if more than several hundred types of words are to be recognized, it is expected that it will take time and effort to pronounce and register all types of words, and the memory capacity required for registration will be enormous. . Furthermore, there is a drawback that the time required for matching the input button and the standard button increases as the number of words increases.

On the other hand, the method of dividing the input speech into phoneme units and recognizing them as combinations of phonemes (hereinafter referred to as phoneme recognition) and calculating the similarity with a word dictionary expressed in phoneme units requires less memory capacity for the word dictionary. It has the advantage that it requires much less time, the time required for slam matching is short, and it is easy to change the contents of the dictionary. For example, the utterance of “obey” is /
a/.

Since the three phonemes /ni/ and /i/ can be combined and expressed in an extremely simple format called AKAI, it is easy for unspecified speakers to deal with speech of many words.

FIG. 1 shows a professional diagram of a speech recognition method characterized by phoneme recognition. Audio input through a microphone or the like is analyzed by an acoustic analysis section 1.

As an analysis method, a bandpass filter group or a winning side analysis is used, and spectral information is obtained every frame period (about 1'oms). The phoneme discrimination unit 2 uses the spectrum information obtained by the acoustic analysis unit 1 to discriminate phonemes for each frame based on the data in the standard pattern storage unit 3. The standard patterns stored in the standard pattern storage section 3 are prepared in advance for each phoneme from the voices of many speakers. The segmentation unit 4 detects speech intervals and determines boundaries for each phoneme (hereinafter referred to as segmentation) based on the analysis output of the acoustic analysis unit 1. The phoneme recognition unit 6 performs segmentation. As a result, a series of phonemes is completed.

The word recognition unit 6 compares this phoneme sequence with a word dictionary 7 similarly written in phoneme sequences, and outputs the results of recognition of all words with the highest degree of similarity.

The conventional segmentation unit 4 performs consonant segmentation as follows.

Figure 2a shows the magnitude of the change in power over time.
The figure shows the magnitude of the change in the rate of change of power over time. When the shape of the temporal change in power using a bandpass filter 8 is concave (this is called a dip), Let n be a frame line in which the power shows a local minimum value, and n,
The frame in which the rate of change of power over time (this is called the power difference value) 9 shows extremely large negative and positive values in the frames before and after n2. Let it be n. Also, if the difference value in a certain frame n is WD(n), when the condition of VD (n,) -WD (n2)<θa) (1) is satisfied, the interval from n2 to n is considered as a consonant interval. there was. Here, 0 is a threshold value for ω to prevent the addition of consonants, and is determined in advance based on statistical distribution.

Details of the segmentation unit 4 and phoneme discrimination unit 2 are shown in FIG. The segmentation unit 4 includes a dip detection unit 31, a consonant interval determination unit 32, and a consonant determination unit 33. The dip detection unit 31 performs dip detection using the power of a bandpass filter determined by acoustic analysis, and the consonant interval determination unit At step 32, the area between n2 and n3 in FIG. 2 is determined as a consonant section. The consonant determination section 33 performs consonant determination for this section based on the spectral shape. On the other hand, the phoneme discrimination unit 2 includes a vowel candidate extraction unit 36 and a vowel interval determination unit 36, and the vowel candidate extraction unit 36 calculates the similarity with respect to the standard slam storage unit 3 using the LPC cepstral coefficients obtained by acoustic analysis. , the phoneme with the highest degree of similarity is extracted as a vowel candidate.

In this case, the standard slam storage section 3 is created using LPC cepstral coefficients for each frame, targeting the five vowels and nasal sounds. This result is applied to consonant intervals other than the consonant interval determined by the consonant interval determination unit 32, and the vowel interval and vowel type are determined by the vowel interval determination unit 36''.
The phoneme recognition unit 6 performs phoneme recognition by combining the results with the results of , and sends the result to the word recognition unit 6 shown in FIG.

According to this method phoneme segmentation.

Although good discrimination can be made, there are two drawbacks because phoneme boundaries are uniquely determined by the presence of dips. One of them is that even when no-no-wah becomes unstable in a vowel, it is detected as a dip, so a consonant is added. The problem is that the addition of one phoneme results in the addition of two phonemes. Another reason is that the dip section does not necessarily represent the correct boundary.
The problem is that the correct boundaries between vowels and consonants are not established. This causes errors in discrimination between vowels and consonants, errors in discrimination between simple vowels and long vowels, and the like.

An example is shown in FIG. This is an example of saying "number" and indicating the position of each phoneme using the Otsu label.

Dip C is detected by the dip detection sound 631 in FIG.
The result is transferred to the consonant section determining section 32, and the result determined by the consonant determining section 33 is shown in d.

On the other hand, the extraction result of the vowel candidate extraction unit 36 is shown in e, and the vowel segment determination unit 36 performs vowel recognition by combining the result of the consonant interval determination unit 32 and the vowel candidate e. The results are shown in f. The vowel recognition result f and the consonant=S result d are transferred to the phoneme recognition section 6 to obtain a recognition result. The consonant recognition section d shows the results of determining the consonant section between n2 and n3 in FIG. 2 according to the position of dip C, and determining the type of phoneme according to the degree of spectral similarity to the standard pattern. vowel candidate e
The section shows the phonemes with the highest spectral similarity for vowels and nasals. By mechanically combining the vowel candidates e using the boundaries of the consonant recognition d as correct boundaries, a phoneme sequence as shown in the section ``Recognition Results'' is created.

Comparing label a and recognition result, /h/ and /u
/ is added. Also, /N/ is replaced with /n/, and the /ri/ section is incorrect.

This is just one example, and is caused by the fact that the dip sections shown in FIG. 2 do not necessarily represent consonant boundaries.

The frequency with which such errors occur differs from person to person, and errors occur more frequently for speakers whose vocalization method is unstable or for speakers whose frequency characteristics deviate greatly from the bandpass filter for detecting dips.

As a result, additions, omissions, and substitutions of phonemes occur frequently, which has the disadvantage of deteriorating word recognition performance.

OBJECTS OF THE INVENTION An object of the present invention is to eliminate the above-mentioned drawbacks and provide a high-performance speech recognition method by improving the accuracy of phoneme segmentation and phoneme discrimination.

SUMMARY OF THE INVENTION The present invention achieves the above-mentioned objects by determining the similarity of phonemes to a standard pattern, determining the position of consonant candidates based on changes in power, and determining the continuity of similarity and intensity of the spectrum with respect to vowels. By comparing vowel candidates extracted based on stability with consonant candidates, the position of the boundary between phonemes and the type of phoneme between the boundaries can be determined with high accuracy, making it possible to perform high-performance speech recognition. It is something to do.

Description of examples Embodiments of the present invention will be described below with reference to the drawings.

The error shown in FIG. 4 occurs because the dip section does not necessarily represent the consonant boundary. Dips are caused by power fluctuations, but do not necessarily correspond to spectral fluctuations. In other words, even if a dip exists, if there is no change in the spectrum, it can be considered that no consonant exists there. Furthermore, if the spectrum is stable at the position of the start or end of the dip, and the spectrum changes significantly at other positions, it can be considered that the true phoneme boundary is at that position. This embodiment makes it possible to accurately determine the boundary between a consonant and a vowel by actively utilizing this property.

FIG. 6 shows a block diagram of the main parts of a speech recognition device that is an embodiment of the present invention.

The standard pattern stored in the standard baton storage unit 44 is a p-order LP of n frames near the phoneme center for vowels and nasals.
It is created using the G kepnutrum coefficient. In other words, it is composed of two-dimensional bumps on the time-frequency axis. p-order LPG key nutrum coefficient woC1n in the n-th frame of phoneme i
Denote it as p and create a vector y.

yl=(cl, 1.C engineering, □,..., c14.c, □
1. ..., Cmu1. ...Ra1. ...+C1np)
1 from many voices and calculate the average value of y#i in mg
(3 represents the order of parameters, maximum = nXp)
shall be. The covariance matrix is the same regardless of the type of phoneme, and is expressed as IW. Let the inverse matrix of IW be IW-1, (]I
I') If the element is σj, then the weighting coefficient a for the j@th parameter of phoneme i 1. It can be expressed as .

Parameter x(?+
+ −”2 + ”’ + xj + ”’ + !
k) Ma Lanobis distance operator for the distribution of phoneme operators'
It can be expressed as . t represents a transposed matrix.

The first term of equation (3) is omitted because it does not depend on the type of phoneme, and the similarity Li can be simply determined as L1=j4 aijxj miw ml (4).

Therefore, in the standard button storage section 44, a:Ij of equation (4)
It is sufficient to include the constant m and Wm.

Next, parameters X(X1+x2.
..., 乃, ...xk)
is calculated by the vowel candidate extracting unit 45 using equation (4), and vowel candidates are extracted based on the continuity of the similarity to the vowel and the stability of the spectrum by intensity, and the result is transferred to the vowel interval storage unit 46. .

On the other hand, after performing the acoustic analysis, the dip detection section 40 performs dip detection of the power of the bandpass filter. The consonant section detection section 41 sets the period between n2 and n3 shown in FIG. 2 as a temporary consonant section and transfers the result to the consonant section storage section 42. Consonant candidate extraction unit 49 with dip detection unit 40 and consonant interval determination unit 41
Configure.

A phoneme boundary determination unit 47 collates the consonant interval storage unit 42 and the vowel interval storage unit 46 to determine the phoneme boundary. In this case, since the standard slam storage section 44 is statistically configured with a plurality of frames near the center of the phoneme, slight fluctuations in the spectrum in the vowel can be absorbed as mere disturbances in the spectrum in the vowel. Furthermore, in the ambiguous region at the boundary with the consonant, the spectrum is not stable over time, so a large degree of similarity does not appear. By utilizing this property, vowel intervals can be extracted with high accuracy.

Therefore, consonant candidates for which there is no possibility of a phoneme boundary are removed, consonant intervals with large errors are corrected, and the results are stored in the consonant interval storage unit 42 for consonants and in the vowel interval memory for vowels. It can be returned to section 46.

Next, the phoneme boundary determination unit 47 determines the consonant interval storage unit 42.
Based on the results, the consonant determination unit 43 calculates the degree of spectral similarity to the standard pattern in the new section and performs consonant determination. By combining this result with the result of the vowel interval storage section 46, the phoneme recognition section 48 performs phoneme recognition, and the result is transferred to the word-identification section.

FIG. 6 shows an example of recognition performed by this embodiment.

In the figure, a indicates the label determined by the festival.

C indicates a dip region detected by the dip detection section 40 in FIG. 5, and d indicates a consonant candidate determined by the consonant section determination section 41. Further, e is a consonant candidate modified by the phoneme boundary determining unit 4, and 6 is a consonant recognition result obtained by determining the consonant candidate shown in e by the consonant determining unit 43. Further, q indicates a vowel candidate extracted by the vowel candidate extraction section 46, and h indicates a vowel recognition result modified by the phoneme boundary determination section 47. b indicates the recognition result recognized by the phoneme recognition unit 48 from the consonant recognition result f and the vowel recognition result.

In this embodiment, for consonant recognition, the dip detection section 40 first detects the dip position shown in FIG. 6C. For this dip position, the consonant interval determination unit 41 extracts the consonant candidates /bA/n/; /IIl/ and /h/ shown in FIG. 6d, and transfers them to the consonant interval storage unit 42.

On the other hand, for vowel recognition, the vowel extraction unit 46 selects the phoneme with the highest degree of similarity for each frame using the standard baton composed of time-frequency bangs stored in the standard baton storage unit 44. 6 Extract the vowel candidates shown in Figure q,
It is transferred to the vowel interval storage unit 46.

The phoneme boundary determination unit 47 refers to the results of the consonant interval storage unit 42 and the vowel interval storage unit 46 to make a final determination of highly accurate phoneme boundaries.

As mentioned above, the following properties are obtained by extracting vowel candidates using the time-frequency button as the standard button.

■ It is possible to absorb small disturbances in the spectrum during vowel intervals and extract vowels stably.

■ Since the spectrum is not stable over time in the crossing part, it is possible to prevent the extraction of extra vowel candidates.

■ Consonant candidates added by dip due to unstable power among vowels can be removed based on the stability of the vowel candidates.

This embodiment actively utilizes this property and performs the following processing. First, in the part where the consonant /h/ is added, shown in FIG.

■It can be removed using the properties of. In addition, in the /N/ part shown in label a, by using the fact of ■ that no stable vowel candidates other than /N/ are extracted when looking at the vowel candidate q, the section up to the next tip shown in C can be calculated. h
/N/ can be determined as shown in . Also, looking at the vowel candidates for q in the /m/ interval shown in d, 10/ overlaps with a part of the /m/ interval, and this 10/ is stable over a long interval. By utilizing the property of ``Gara'', the interval of /m/ can be modified. The results are shown in Figure 6e.

As a result of the above processing, the consonant section is transferred to the consonant section storage section 42 as FIG. 6e, and the vowel section is transferred to the phoneme recognition section 48 via the vowel section storage section 46 as shown in FIG.

The consonant determination unit 43 reviews the phoneme /m/ whose phoneme boundaries have been corrected among the consonant candidates shown in FIG. It is corrected to a high element /S/ and transferred to the phoneme recognition unit 48 as a consonant recognition result f.

In this way, in this method, more precise phoneme segmentation and phoneme discrimination can be achieved by combining detection of consonant candidates by dissipation and spectral stability.

Using this method, phoneme recognition was performed on 2120 words uttered by 10 adult males, and the results of the evaluation are shown in the table.

As is clear from the table, the average recognition rate for all phonemes is 82.6.
A good value of % can be obtained. Also, the sound addition rate is 4.
It is possible to create a highly accurate phoneme sequence with extremely low addition and omission errors of 8% and phoneme omission rate of 3.9%.

In the above embodiment, a case was described in which hpc cepstral coefficients were used as spectrum information, but other information such as a filter bank output may be used.

Effects of the Invention In short, the present invention measures the similarity of phonemes to standard batan, selects consonant candidates based on changes in power, and selects consonant candidates based on continuity of similarity to vowels and stability of spectrum according to intensity. By comparing candidates with consonant candidates, the position of the boundary between phonemes and the type of phoneme between the boundaries can be determined with high accuracy, and there is an advantage that highly reliable phoneme recognition can be realized.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a conventional speech recognition device, FIG. 2 is a diagram showing changes in power and power change rate over time, and FIG. 3 is a block diagram of main parts of a conventional speech recognition device. FIG. 4 is a diagram showing an example of recognition performed by the same device, FIG. 6 is a block diagram of main parts of a speech recognition device in an embodiment of the present invention, and FIG. 6 is an example of recognition results by the same device. It is a diagram. 4o...Dip detection section, 41...
Consonant interval determining unit, 42...Consonant interval storage unit, 4
3... Consonant determination section, 44... Standard pattern storage section, 46... Vowel candidate extraction section, 46.
...Vowel interval memory section, 47 degrees. ... Phoneme boundary determination section, 48 ... Phoneme recognition section,
49...Consonant candidate extraction unit. Name of agent: Patent attorney Toshio Nakao and 1 other person 2nd
Figure R No. 31!1 C number) bαNri00

Claims (2)

[Claims] (1) A standard pattern storage unit that stores in advance a standard pattern constructed based on a statistical distance measure using spectral information obtained from the voices of multiple speakers, and a statistical distance for each analysis interval using spectral information. a vowel candidate extraction unit that calculates the similarity of phonemes to the standard pattern based on a measure;
A consonant candidate extractor extracts consonant candidates using a dip number based on a temporal change in power, and a vowel candidate extractor extracts consonant candidates based on the temporal continuity of the similarity to the phoneme or the stability of the spectrum depending on the strength of the similarity. The vowel candidates obtained are compared with the consonant candidates in the consonant candidate extraction section, and the positions of the boundaries between phonemes and the number of phonemes between the boundaries are determined. A speech recognition device comprising at least a phoneme boundary determination unit that determines S*. (2) The standard pattern is constructed based on a statistical distance measure by a two-dimensional time-frequency pattern using spectral information of a plurality of analysis interval lengths near the center of a phoneme. The speech recognition device according to scope 1.