JPH0246957B2 - - Google Patents

Description

[Detailed description of the invention]

[Industrial Field of Application] The present invention relates to a monosyllabic speech recognition method, and is particularly concerned with a monosyllabic speech recognition method that can improve the recognition rate even though it has a particularly simple configuration. [Prior Art] In recent years, voice recognition devices have come into practical use as input devices for computers, various control devices, and the like. Speech recognition devices that can recognize human speech as they are have various advantages, such as requiring no special training for use and no restrictions on line of sight or limbs. However, many of the speech recognition devices currently in use are word speech recognition devices that perform recognition on a word-by-word basis, and while they have the above-mentioned advantages, they have the drawback of being limited in the number of words. For these reasons, monosyllabic speech recognition systems that recognize speech in units of syllables have recently been attracting attention. As is well known, the Japanese language system is composed of phonetic characters, meaning that each syllable corresponds almost one-to-one to each kana, so monosyllabic speech recognition systems can be used as various input devices. . In particular, with the spread of Japanese word processors and workstations, many attempts have been made to use this monosyllabic speech recognition system as an input means for these devices. By the way, in monosyllabic speech recognition systems,
The consonant parts of the syllable characteristic parameter time series (hereinafter simply referred to as patterns) are extracted from all the syllable parts, and the consonant parts are matched. Syllable patterns have fewer characteristics between patterns than word patterns, and vowel parts are generally much longer in duration than consonant parts, and minute differences in similarity between consonant parts can cause fluctuations in the pattern of the following vowel part. This is because it is masked by Yoshida et al.'s paper "Japanese monosyllabic speech recognition experiment"
(Proceedings of the Acoustical Society of Japan, 3-2-16, 1979)
shows an example of such extraction of consonant parts. In this example, the consonant part is cut out using an algorithm that sets the consonant/vowel boundary at a predetermined position among the points where the feature vector of the syllable changes significantly.
This consonant part is subjected to a dynamic programming matching method (hereinafter referred to as DP matching) with free endpoints using a standard registered consonant pattern to obtain consonant information. In addition, Mr. Furui's paper ``Single syllable recognition and its application to speech recognition of large word sounds'' (Proceedings of the Institute of Electronics and Communication Engineers A, 65-A; 2, pp175-182, 1982) is about cutting out consonant parts. shows other techniques. In this example, the beginning of a word is cut out at a predetermined ratio to the total length of the syllable, and this is compared with similar registered beginnings of words to obtain consonant information. For example, performing linear matching. Furthermore, the paper by Mr. Nakagawa et al., “Large vocabulary word speech recognition using monosyllable unit input from unspecified speakers” (IEICE Transactions D, 65-D, 12, pp1558-1565,
(1982) also disclose other extraction methods. In this example, since the subject is an unspecified speaker, the registered patterns can be fixed, and for this reason, the boundary point j between consonants and vowels is determined in advance for each registered pattern by visual inspection. This decision point is estimated by taking into consideration, for example, the waveform and formant of the audio signal. Regarding the unknown input pattern, the registered pattern for the entire syllable and this unknown input pattern
DP matching is performed, and the point i where the optimal path passes through the above boundary point j is the consonant of the unknown input pattern,
It is used as a vowel boundary point. Various proposals have been made to obtain consonant information, including endpoint-free DP matching. This example also discloses consonant extraction using the same algorithm as in Furui's paper mentioned above. Both of the above-mentioned papers use a simpler method than DP matching for identifying vowel information. Then, this vowel information and the above-mentioned consonant information are combined to identify the syllable. However, in the conventional configuration as described above, it was necessary to employ a complicated algorithm to separate the consonant part and the vowel part. In other cases, fixed registered patterns must be visually checked in advance, resulting in complicated work and requiring deep experience. Furthermore, errors in identifying separation points between consonants and vowels may increase the number of errors in recognizing syllables, especially their consonant information. In Mr. Furui's example, the consonant part is cut out in a fixed manner based on the ratio to the total syllable length, so there is no problem like the one mentioned above, but since the variation in the consonant/vowel boundary point for each syllable is ignored, it naturally cuts out the consonant part for each syllable. It is thought that the recognition rate is different. The final stage of recognizing syllables is as follows:
There is a method of obtaining consonant information and vowel information separately and determining a syllable from a combination of them. For example, the syllable ``ki'' is identified from the vowel information of ``i'' and the consonant information of ``k''. Another method is
It is known that vowel information is obtained in advance, syllables with that vowel as the subsequent vowel are recognized candidates, and DP matching is performed on these candidate syllables. In the latter case, the articulatory coupling elements between consonants and vowels can be fully considered in the final matching, so that good discrimination can be achieved. Such two-stage evaluation is described in JP-A-54-145409, JP-A-58-52694, and JP-A-58-59498. However, in this case,
It is complicated to perform two-stage evaluation. As described below, one embodiment of the present invention can solve this problem. [Problems to be Solved by the Invention] This invention has been made in consideration of the above circumstances, and provides a monosyllabic speech recognition method that can easily and reliably obtain consonant information without distinguishing consonants from vowels. is intended to provide. [Means for Solving the Problems] In order to achieve the above object, the present invention uses DP for unknown input patterns and registered standard patterns.
A distance calculation is performed during matching, and the consonant information of the unknown input pattern is identified based on the minimum cumulative distance at a predetermined midpoint between the beginning and end of these unknown input patterns or registered standard patterns. . In one aspect of the present invention, the minimum cumulative distance is also determined for a second intermediate point that is closer to the end of the word than the intermediate point from which consonant information is obtained, that is, contains more vowel information sources,
Based on this, vowel information may be identified, and then consonant information may be identified for candidate standard patterns in which the identified vowel is the subsequent vowel. That is, two-stage matching may be performed. In this case, the minimum cumulative distance of the midpoint for consonants can be obtained as a by-product of the distance calculation when discriminating vowel information, and the distance calculation is
This process can be completed in one go. In another aspect of the present invention, when obtaining vowel information, the ending of a word may be known and the DP matching calculation may be performed up to a predetermined midpoint. In this case, 5
Since the vowels of the unknown input pattern can be reliably identified using a standard pattern of one vowel, the amount of calculation is significantly reduced compared to the case where all syllables are referred to. [Embodiment] Hereinafter, an embodiment of the present invention applied to a speech recognition device for a specific speaker will be described with reference to the drawings. FIG. 1 shows this embodiment as a whole. In FIG. 1, a speaker's voice is supplied to a microphone 1, and this voice is converted into an audio signal and supplied to an A/D converter 2. be done. This A/D converter has a sampling frequency of 20KHz, for example.
The data bit length is 12 bits. A/
The data from the D converter 2 is supplied to the feature parameter extraction section 3, where a feature parameter time series (pattern) is formed based on the above-mentioned data. In this example, a logarithmized spectrum is used as this characteristic parameter, as will be described in detail later. This example targets a specific speaker, so
Training is performed prior to speech recognition. That is, a speaker utters a predetermined number of syllables to be identified, for example, 68, into the microphone 1, which are sequentially calculated by the A/D converter 2 and the feature parameter extractor 3, and then sent to the recognizer 4 to identify each syllable. We will supply standard patterns. In this case, the switching circuit 5 of the recognition section 4 is switched to "a" so that the information is registered in the storage section 6 of the recognition section 4. After this preliminary step, the speaker divides the syllables into, say, 100 words.
As speech is input at a rate of syllables/minute, each syllable is derived as an unknown input pattern via the feature parameter extraction section 3 and stored in the storage section 6 of the recognition section 4. At this time, the switching circuit 5 is switched to side b. The unknown input pattern is sequentially referenced to 68 standard patterns, and the most optimal one among the reference results is outputted via the output circuit 7 to an output device 8 such as a printer or a monitor. Of course, in addition to the most optimal candidate, second and third candidates may also be output. The feature parameter extraction unit 3 in FIG. 1 forms a time series of logarithmized spectra as shown in FIGS. 2 to 6. That is, the digital data from the A/D converter 2 is pre-emphasized, and based on this pre-emphasized data, the energy Ei for each time frame i, 10 msec in this example, is determined. However, Ei = 10log ₁₀ (average of squared amplitudes). After this, the normalized energy Ein is calculated from the maximum energy Emax and the minimum energy Emin, and Ei
For example, normalize to a value between 0 and 32. However, Ein=32(Ei−Emin)/Emax−Emin. And the normalized energy Ein obtained in this way
The boundaries between syllables are determined by setting an appropriate threshold based on the time distribution of . On the other hand, short-term spectral analysis is also performed on the pre-emphasized digital data. In other words, while moving the digital data every 10 msec per frame, Winograd's fast Fourier transform using the time function of the Hamming window is performed over a range of 20 msec (400 points). The power spectrum thus obtained is logarithmized, and the spectrum from 10Hz to 7900Hz is further divided into 19 frequency bands. Specifically, 100Hz and 200Hz form one band, similarly 300Hz and 400Hz,...
700Hz to 7900Hz each form a band. Thereafter, the average value of the power spectrum of the unvoiced part is subtracted from the power spectrum of the voiced part (each syllable) as background noise. The power spectrum from which the background noise has been subtracted, that is, the characteristic pattern, is normalized in the time direction and is further subjected to nonlinear transformation of frequency components. That is, consider a characteristic pattern of times 1 to n _F and frequency bands 1 to m as shown in FIG. Here, _nF is the number of frames for each syllable, m is the number of bands, and in this example, m=19. The logarithmized power spectrum at each time and each band is simply indicated by circles.
To perform normalization in the time direction, both the standard pattern (word length _TR ) and the unknown input pattern (word length T _I ) are set to a predetermined pattern length (Tn) as shown in Figure 4.
Convert to by linear interpolation. The pattern length Tn is determined separately by experiment. For example, it may be selected to be 1.2 times the average standard pattern syllable length. Nonlinear transformation of frequency components is performed as shown in FIGS. 5 and 6. That is, as shown in Fig. 5, the characteristic pattern after time normalization is expressed as Vij,
The maximum energy is expressed as Vi _MAX = MAX ^j (Vij). Then, as shown in FIG. 6, Vij is converted into a value Vij from 0 to 255 according to the formula below. In the case of VijVi _MAX −Vb, Vij=0 In the case of Vij>Vi _MAX −Vb, Vij=255/Vb (Vij−Vi _MAX +Vb) Note that Vb is determined separately by experiment. Vb=30,
When Vb was changed to 40 and 50, the optimum value was Vb=40. The optimal value of Vb is considered to be related to the ratio of audio peak signal to noise power. This nonlinear conversion can alleviate the adverse effects of noise power. Next, the recognition unit 4 shown in FIG. 1 will be explained with reference to FIGS. 7 and 8. The recognition unit 4 includes a storage unit 6, a cumulative distance calculation unit 9, and the like. The storage unit 6 stores an input pattern P _I and a standard pattern P _R i, and a cumulative distance calculation unit 9 executes a DP matching operation between these input patterns P _I and the standard pattern P _R i. Ru. However, this cumulative distance calculation section 9 does not require calculation for matching over the entire pattern. It is sufficient that the calculation is performed at least up to time t _I =t _V (FIG. 7) on the time axis of the input pattern. Here, the time point tc in FIG. 7 is an intermediate point for obtaining consonant information, and the time point _tV is an intermediate point for obtaining vowel information. This is detailed below. First, 68 standard patterns P _R i are matched with respect to a predetermined unknown input pattern P _I. In other words, as shown in Fig. 8, the unknown input pattern P _I and the first
A DP matching operation with the standard pattern P _R i is executed, and at this time, the minimum cumulative distance Dc at t _I =tc is determined. Starting point (Figure 7, beginning of word)
There are various paths from to the grid on t _I = tc (limited to the matching window, so the grid is within the range indicated by Wc), but we find the minimum cumulative distance of each of these paths. . Thereafter, the DP matching calculation is continued to obtain the same minimum cumulative distance Dv at t _I =t _V. These minimum cumulative distances Dc and Dv are stored in the storage section 6. Similarly, the unknown input pattern and the remaining 67
DP between standard patterns P _R i (i = 2 to 68)
A matching operation is performed to obtain the respective minimum cumulative distances Dc and Dv. Once the minimum cumulative distances Dc and Dv are determined for all standard patterns P _R i , vowels are determined from the 68 minimum cumulative distances Dv at t _I =t _V.
That is, the syllable that minimizes the minimum cumulative distance Dv is found, and the vowel of this syllable is used as detected vowel information. After this, a syllable in which the vowel of the detection result is the subsequent vowel (for example, if the vowel is "a", it is "a", "ka", "sa", etc.)
...) at t _I = tc is selected as candidate syllable data, and among them, the one with the smallest minimum cumulative distance Dc is output as a syllable detection result. The above operation is performed by the vowel detection section 1 in Fig. 1.
0, is executed by the selection circuit 11 and the syllable detection section 12. The intermediate points t _I =tc and t _V mentioned above are determined by experiment.
For example, good results were obtained with approximately tcTn×1/3 and tv=Tn×0.7. Table 1 shows the actual results of applying this example to 68 syllables. The number of standard patterns per syllable is one (single template method). In this table, the second candidate is the syllable with the second smallest minimum cumulative distance in the syllable detection unit 12 of FIG.

[Table] As is clear from Table 2, this example provides superior recognition results compared to the previous example described above.

[Table] As described above, in this embodiment, segmentation of consonants and vowels is not required, so single syllables can be recognized with an extremely simple configuration. Furthermore, consonant information can be obtained by using a byproduct of the DP matching operation when obtaining vowel information, and the amount of calculations can be reduced. Furthermore, the configuration is extremely suitable for hardware implementation. This embodiment can also improve the recognition rate. This can be considered as follows. As mentioned above, the syllable feature vector includes consonant information at the beginning of the word, and vowel information over a wide range from the middle of the word to the end. This is schematically shown in FIG. 9. Then, by visual inspection or experience, the boundary between the two is arranged as indicated by a broken line. However, consonants influence the following vowels, and the transition region in particular is thought to contain important elements for determining consonants. Therefore, detecting the boundaries between consonants and vowels to extract consonants makes the consonant information uncertain. In this example, the intermediate point tc for obtaining consonant information is determined by experiment without being concerned about the consonant/vowel boundary, and is determined as shown in FIG. 9, for example, so that more preferable recognition can be performed. Also, the end points of paths that can be realized in terms of consonant determination are shown in Figure 7.
Since it is a lattice group indicated by Wc, the corresponding end point of the standard pattern can be selected within this range, and delicate matching in the consonant region between the input pattern and the standard pattern can be performed with a higher degree of freedom. It is thought that this will improve. Note that this invention is not limited to the above-described embodiments, and various changes can be made without departing from the spirit thereof. For example, the present invention may be applied to 101 monosyllables including syllables, and multiple templates may be used. With the same configuration as the above embodiment
Table 3 shows the 101 monosyllable recognition results, and Table 4 shows the comparison results with the previous model using multiple templates.

【table】

[Table] Furthermore, in the above-described embodiment, DP matching was performed from the beginning of the word when obtaining vowel information, but DP matching may be performed when the ending of the word is known when obtaining vowel information. In this case, vowels can be determined sufficiently by setting 20 to 30% of the word length as the midpoint. There are only five standard vowel patterns: "A", "I", "U", "E", and "O". This is because recognition errors are less likely to occur due to the influence of consonants, and unlike the above-described embodiments, there is no need to obtain 68 or 101 consonant information as a subsidiary from the calculation for vowel information. Therefore, the amount of calculation can be extremely reduced. Of course, corresponding candidate syllables are narrowed down based on the recognition results regarding vowels. Separately for consonant information identification
68 or 101 standard patterns are prepared, and the candidate syllables are referred to for consonant information according to the present invention, and the input syllables are finally identified. Furthermore, in the DP matching operation when obtaining consonant information, the starting point and midpoint tc may be configured to be variable, or syllable recognition is performed for each of multiple sets of starting points and midpoints, and the final decision is made by majority vote. It is also possible to have the syllable determined by the syllable. Of course, it is also possible to select an intermediate point on the time axis of the registered standard pattern and obtain consonant information based on the minimum cumulative distance at this intermediate point. [Effects of the Invention] As explained above, according to the present invention, distance calculation is performed by DP matching for unknown input patterns and registered standard patterns, and distance calculation between unknown input patterns or registered standard patterns is performed between the beginning and the end of the word of these unknown input patterns or registered standard patterns. The consonant information of the unknown input pattern is obtained based on the minimum cumulative distance at a predetermined midpoint. Therefore, there is no need to discriminate consonant/vowel boundary points, and the configuration can be simplified, and errors associated with determining boundary points can be eliminated. Moreover, since the matching for obtaining consonant information can be given a degree of freedom, the recognition rate can be improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention, FIGS. 2, 3, 4, 5, and 6.
The figure is a diagram for explaining the feature parameter extraction section 3 of the embodiment in FIG. 1, FIGS. 7 and 8 are diagrams for explaining the recognition section 4 of the embodiment in FIG. It is a figure for explaining the effect of a figure example. 1...Microphone, 2...A/D converter, 3
...Feature parameter extraction unit, 4...Recognition unit, 8...
...Output device.

Claims (1)

[Claims] 1 A unit that analyzes a speech signal of an unknown input monosyllable and compares the input feature parameter time series extracted from this speech signal with the registered feature parameter time series registered in advance to recognize the unknown input monosyllable. In the syllable speech recognition method, a distance calculation is performed according to a dynamic programming matching method in which the beginning of the word is known for each of the input feature parameter time series and the registered feature parameter time series, and the distance calculation is performed according to the dynamic programming matching method. Or, find the minimum cumulative distance at two different midpoints uniformly determined in advance for all syllables between the beginning and end of each word in the registered feature parameter time series,
The vowel of the unknown input syllable is determined based on the minimum cumulative distance between the midpoints of the endings, the candidate syllables are narrowed down based on this vowel, and the minimum cumulative distance of the midpoints of the endings of the narrowed down syllables is determined. A monosyllabic speech recognition method characterized in that the syllable that minimizes the syllable is the recognition result.