JPH0333280B2 - - Google Patents

Description

[Detailed description of the invention]

FIELD OF INDUSTRIAL APPLICATION The present invention relates to a speech recognition device for automatically recognizing speech signals uttered by a human 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 instructions from humans to computers and various machines. The pattern 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 calculate the degree of matching (hereinafter referred to as similarity). The word is calculated and determined to be the same word as the standard pattern that yields the maximum match.
In this pattern matching method, standard patterns must be prepared for all words to be recognized, so if the speaker changes, a new standard pattern must be input and stored. Therefore, in cases where hundreds of types of words or more 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. Ru. Furthermore, there is a drawback that the time required for pattern matching between the input pattern and the standard pattern increases as the number of words increases. On the other hand, input speech is divided into phoneme units and recognized as combinations of phonemes (hereinafter referred to as phoneme recognition).
The method of determining similarity with a word dictionary written in phoneme units has the advantage that the memory capacity required for the word dictionary is significantly smaller, the time required for pattern matching is shorter, and the contents of the dictionary can be easily changed. ing. For example, the utterance of "red" is /
Since the three phonemes a/, /k/, and /i/ can be combined and expressed in an extremely simple format called AKAI, it is easy for non-specific speakers to deal with the sounds of many words. FIG. 1 shows a block diagram of a speech recognition system characterized by phoneme recognition. Audio input through a microphone or the like is analyzed by the acoustic analysis section 1.
As an analysis method, a band filter group uses linear predictive analysis to obtain spectrum information every frame period (about 10 ms). The phoneme discrimination section 2 uses the spectrum information obtained by the acoustic analysis section 1 and performs phoneme discrimination for each frame based on the data in the standard pattern storage section 3. The standard patterns stored in the standard pattern storage section 3 are obtained in advance for each phoneme from the voices of multiple speakers. In segmentation section 4,
Based on the analysis output of the acoustic analysis unit 1, voice sections are detected and boundaries for each phoneme are determined (hereinafter referred to as segmentation). The phoneme recognition unit 5 performs a task of determining what phoneme is in one phoneme interval based on the results of the segmentation unit 4 and the phoneme discrimination unit 2. As a result, a series of phonemes is completed. The word recognition unit 6 compares this phoneme sequence with a word dictionary 7 that is similarly expressed in phoneme sequences, and outputs the word with the highest degree of similarity as a recognition result. In the conventional segmentation unit 4, consonant segmentation was performed as follows.
Figure 2a shows the magnitude of the change in power over time, and Figure 2b shows the magnitude of the change in the rate of change of power over time. When has a concave shape (this is called a dip), the frame where the power shows the minimum value is n ₁ , and the rate of change of power over time in the frames before and after n ₁ (this is called the power difference value) Let n ₂ and n ₃ be frames in which 9 indicates negative and positive broadside values. Also, if the difference value at a certain frame n is WD(n), when the condition of WD(n ₃ )−WD(n ₂ )θ〓 (1) is satisfied, the interval from n ₂ to _{n 3} is a consonant interval. It was. Here, θ〓 is a threshold value for preventing 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 the dip detection using the power of the band filter obtained through acoustic analysis, and determines the consonant interval. In the section 32, the period between n ₂ and _{n 3} in FIG. 2 is determined as a consonant interval. Consonant determination unit 33 based on the spectrum shape for this section.
Consonant judgment is performed. On the other hand, the phoneme discrimination unit 2 includes a vowel candidate extraction unit 35 and a vowel interval determination unit 36, and uses the LPC cepstral coefficients obtained by acoustic analysis to
Similarity calculation for the standard pattern storage unit 3 is performed by the vowel candidate extraction unit 35, and the phoneme with the highest similarity is extracted as a vowel candidate. In this case, the standard pattern 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 those determined by the consonant interval determination unit 32, and the vowel interval and vowel type are determined by the vowel interval determination unit 36. By combining this result with the result of the consonant determination section 33, the phoneme recognition section 5 performs phoneme recognition and sends it to the word recognition section 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 the power becomes unstable in a vowel, it is detected as a dip, so a consonant is added, and since a vowel is inevitably added according to Japanese rules, the result is a single consonant. The problem is that the addition of 2 phonemes results in the addition of two phonemes. Another problem is that the dip section does not necessarily represent the correct boundary, so the correct boundary between vowels and consonants is no longer guaranteed. 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 the label a indicates the position of each phoneme.
The dip detecting section 31 in FIG. 3 detects the dip c, 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 35 is shown in e,
The vowel segment determination unit 36 performs vowel recognition by combining the result of the consonant segment determination unit 32 and the vowel candidate e. The results are shown in f. The vowel recognition result f and the consonant recognition result d are transferred to the phoneme recognition unit 5 to obtain the recognition result b. In the section of consonant recognition d, the area between n ₂ and _{n 3} in Figure 2 is determined as a consonant interval according to the position of dip c, and the type of phoneme is determined according to the similarity of the spectrum to the standard pattern. show. In the section for vowel candidate e, phonemes with the highest spectral similarity are shown for vowels and nasals. By mechanically combining vowel candidates e using the boundaries of consonant recognition d as correct boundaries, a phoneme sequence as shown in the section of recognition result b is created. Comparing label a and recognition result b, /h/
and /u/ are added. Also, /N/ is replaced with /n/, and the /〓/ 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 varies from person to person, and errors occur in speakers whose vocalization method is unstable or speakers whose frequency characteristics deviate greatly from the band filter used to detect dips. As a result, additions, omissions, and substitutions of phonemes occur frequently, resulting in a drawback that the performance of word recognition deteriorates. OBJECTS OF THE INVENTION An object of the present invention is to eliminate the above drawbacks and provide a high-performance speech recognition method by improving the accuracy of phoneme segmentation and phoneme discrimination. Composition of the Invention The present invention achieves the above object by determining the similarity of phonemes to a standard pattern, determining the position of a consonant candidate based on a change in power, and determining the similarity of each vowel candidate and consonant candidate to the standard pattern. By providing a phoneme boundary determination unit that corrects and determines phoneme intervals by comparing the continuity and strength of the This makes it possible to perform sound immersion recognition. DESCRIPTION OF EMBODIMENTS Examples 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. Although dips are caused by power fluctuations, they do not necessarily correspond to spectral fluctuations. In other words, even if a dip exists, if there is no spectrum variation, it can be considered that no consonant exists there. Furthermore, if the spectrum is stable at the start or end position 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. 5 shows a block diagram of the main parts of a speech recognition device which is an embodiment of the present invention. The standard patterns stored in the standard pattern storage unit 44 are created using p-order LPC cepstral coefficients of n frames near the phoneme center for vowels and nasal sounds. That is, it is composed of a two-dimensional pattern on the time-frequency axis. The p-th LPC cepstral coefficient in the n-th frame of phoneme i is expressed as C _iop , and a vector 〓 _i is created. 〓 _i = (C _i11 , C _i12 ,…, C _i1p , C _i21 ,…, C _i31 ,…
C _io1 , ..., C _iop ) Total 〓 _i from many voices and calculate the average value of 〓 _i .
m _ij (j represents the order of parameters, the maximum is k=
n×p). The covariance matrix is the same regardless of the type of phoneme, and is expressed as 〓. The inverse matrix of 〓 ^-1
and the (j, j′) element is σ ^jj ′, then j of phoneme i
The weighting coefficient a _ij for the th pattern can be expressed as a _ij =2 _K 〓 ^j=1 σ ^ij ′m _ij ′ (2). Mahalanobis distance D _i ² for the distribution of phoneme i of parameters x (x ₁ , x ₂ , ..., x _j , ..., x _k ) obtained from voice data of multiple speakers is D _i ² = x ^t 〓 ^-1 It can be expressed as x− _K 〓 ^j=1 a _ij x _j +m _i ⁱ 〓 ^-1 m _i (3). 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 L _i is simply expressed as L _i = _K 〓 ^j=1 a _ij x _j −m _i ^t W ^-1 m _i (4 ) can be found. Therefore, a _ij of equation (4) is stored in the standard pattern storage section 44.
and the constant m _i ^t W ^-1 m _i . Next, the parameters x (x ₁ ,
x ₂ , ..., x _j , x _k ) is calculated by the vowel candidate extraction unit 45 using equation (4), and the vowel candidate extraction unit 45 calculates the similarity _L Candidates are extracted and the results are transferred to the vowel interval storage section 46. On the other hand, after performing acoustic analysis, the dip detection unit 4
Dip detection of the power of the band filter is performed at 0. The consonant section detecting section 41 sets the period between n ₂ and _{n 3} shown in FIG. 2 as a temporary consonant section and transfers the result to the consonant section storage section 42 . The dip detection section 40 and the consonant interval determination section 41 constitute a consonant candidate extraction section 49. 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 pattern 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 candidates with large errors in the consonant interval are corrected, and the results are stored in the consonant interval storage unit 42 for consonants.
Furthermore, vowels can be returned to the vowel section storage section 46. Next, for the consonant section determined by the phoneme boundary determining section 47 and stored in the consonant section storage section 42, the consonant determining section 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 in the vowel section storage section 46, the phoneme recognition section 48 performs phoneme recognition and transfers the result to the word recognition section. FIG. 6 shows an example of recognition performed by this embodiment. In the figure, a indicates a label determined by inspection. 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. Also e
5 is a consonant candidate modified by the phoneme boundary determining unit 47, and 5 is a consonant recognition result determined by the consonant determining unit 43 for the consonant candidate shown in e. Further, g indicates a vowel candidate extracted by the vowel candidate extracting section 45, and h indicates a vowel recognition result modified by the phoneme boundary determining section 47. b indicates a recognition result recognized by the phoneme recognition unit 48 from the consonant recognition result f and the vowel recognition result h. 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 section determination unit 41 selects the consonant candidates /b/, / shown in FIG. 6d.
n/, /m/, /h/ are extracted and transferred to the consonant section storage section 42. On the other hand, for vowel recognition, the vowel extraction unit 45 selects the phoneme with the highest degree of similarity for each frame using a standard pattern composed of time-frequency patterns stored in the standard pattern storage unit 44.
The vowel candidates shown in FIG. 6g are extracted and transferred to the vowel section storage section 46. In the phoneme boundary determination unit 47, the consonant interval storage unit 42
A highly accurate phoneme boundary is finally determined by referring to the results in the vowel interval storage unit 46. As mentioned above, extracting vowel candidates using a time-frequency pattern as a standard pattern has the following properties. It is possible to absorb small disturbances in the spectrum in vowel intervals and extract vowels stably. Since the spectrum in the crossing part is not stable over time, it is possible to prevent unnecessary vowel candidates from being extracted. Consonant candidates added by dips 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, the consonant / shown in Figure 6 d.
In the addition of h/, / of the vowel candidate shown in g is added.
Since o/ is extracted stably over a long interval, it can be removed using the property of .
In addition, in the part /N/ shown in label a, by using the property that when looking at vowel candidates g, no stable vowel candidates other than /N/ are extracted, c
The interval up to the next dip shown in can be determined as /N/ as shown in h. Also, looking at the vowel candidates for g in the /m/ section shown in d, /
A part of the interval of m/ overlaps with /o/, and the interval of /m/ is corrected by utilizing the property that /o/ is stable over a long interval. be able to. 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 Fig. 6h. The consonant determination unit 43 reviews the phoneme /m/ whose phoneme boundary has been corrected among the consonant candidates shown in FIG. High element/
〓/ is corrected 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 realized by using the detection of consonant candidates by dips in combination with the stability of the spectrum. Using this method, 2120 utterances by 10 adult males were used.
The table shows the results of phoneme recognition and evaluation of words.

[Table] As is clear from the table, the average recognition rate for all phonemes
A good value of 82.6% can be obtained. Furthermore, it is possible to create a highly accurate phoneme sequence with very few addition and omission errors, such as a phoneme addition rate of 4.8% and a phoneme dropout rate of 3.9%. In the above embodiment, the spectrum information is
Although the case where LPC cepstral coefficients are used has been described, other information such as filter bank output may be used. Effects of the Invention In summary, the present invention calculates the similarity of phonemes to a standard pattern, and also calculates consonant candidates based on changes in power, and compares the continuity and strength of the similarity of vowel candidates and consonant candidates with respect to the standard pattern. By correcting the phoneme interval through comparison, 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 speech recognition can be realized.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device, Fig. 2 is a diagram showing changes in power and the rate of change of power over time, and Fig. 3 is a block diagram of the main parts of a conventional speech recognition device. Fig. 4 is a diagram showing an example of recognition performed by the same device, Fig. 5 is a block diagram of the main parts of the 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. 40...Dip detection section, 41...Consonant section determination section, 42...Consonant section storage section, 43...Consonant determination section, 44...Standard pattern storage section, 45...Vowel candidate extraction section, 46...Vowel Section storage section, 47...
... Phoneme boundary determination unit, 48... Phoneme recognition unit, 49...
...Consonant candidate extraction section.

Claims (1)

[Scope of Claims] 1. A standard pattern storage unit that stores in advance a standard pattern configured based on a statistical distance measure using spectrum information obtained from voices of multiple speakers;
a vowel candidate extraction unit that uses spectral information to extract the degree of similarity of phonemes to the standard pattern for each analysis interval based on a statistical distance measure; and a consonant candidate extraction unit that calculates consonant candidates using dips based on temporal changes in power. , a phoneme boundary determination unit that corrects and determines a phoneme interval using continuity and strength of similarity for the vowel candidates and consonant candidates extracted by the vowel candidate extraction unit and the consonant candidate extraction unit; A speech recognition device comprising at least a phoneme recognition unit that determines the type of phoneme in the determined phoneme interval. 2. A patent claim characterized in that 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.