JPS6229799B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a speech recognition device that specifies a speaker. First, referring to FIGS. 1 and 2, the outline of a speech recognition device will be explained, problems with the conventional recognition device will be clarified, and the object of the present invention will be explained. In FIG. 1, an audio signal input from a microphone 10 is amplified by an amplifier 11, analyzed by a frequency analysis section 12, normalized for energy (amplitude), and normalized by an audio pattern conversion section 13. , features are extracted and converted into speech patterns. By connecting the switch 14 to the 15 side in advance, the output from the voice pattern converter 13 is stored in the registered pattern area 18. The voice words desired to be recognized are uttered in sequence and their voice patterns are stored in the registration area 18 in sequence. In the pattern matching unit 17 that connects the switch 14 to the switch 16 side, pattern matching is performed between the registered pattern in the registered pattern area 18 and the output from the audio pattern conversion unit 13, and each registered pattern is matched with the input pattern. (output from 13) is calculated. In the determination unit 19, the one with the highest degree of similarity is selected from these degrees of similarity, and the validity of this degree of similarity is examined. If it is valid, the category number of the registered pattern with the greatest degree of similarity is output as a result, which means that the input speech has been recognized. In order to perform speech recognition, the speech signal must be analyzed and converted into a speech pattern by some means. In the case of FIG. 1, frequency analysis is used as the analysis means. This analysis section has been a problem in the past in order to downsize or reduce the cost of speech recognition devices. FIG. 2 shows an example of the frequency analysis section of FIG. 1. 20, 21, 22 are respectively,
These are band filters with different center frequencies. The outputs of these bandpass filters are passed through a rectifier circuit 23 and a low-pass filter 24 to convert them to values close to the DC level, and a multiplexer 25 controlled by a control signal 27 converts the outputs of each low-pass filter into Connect with the A/D converter 26,
Convert analog signals to digital signals that are easy to process. Bandwidth filters are generally 8 or 16
In order to realize FIG. 2, the size of the entire device becomes large, which poses problems in terms of miniaturization and difficulty in adjusting the band filter. Another method for realizing frequency analysis is FFT (Fast Fourier Transform), but due to the complexity of processing, it is difficult to process in real time, and if you try to realize it, there will be problems such as cost. Alternatively, it is possible to implement the diagram in FIG. 2 using a digital filter, but it is difficult to do so, as is the case with FFT. For FFT and digital filters,
This is because multiple multipliers are required. There are other methods than frequency analysis, or methods for obtaining and analyzing autocorrelation coefficients, linear prediction coefficients, etc., but all of these methods require multiple multipliers, reducing the size of the entire device and reducing costs. In terms of downs, it was a problem. There has been a desire for a means that can analyze audio signals and convert them into audio patterns through simpler processing. This is also the purpose of the present invention, and it is an object of the present invention to realize miniaturization or cost reduction of a speech recognition device through simpler processing. Embodiments of the present invention will be described below with reference to the drawings. Fourier transform is a means of converting a signal into the frequency domain. However, there are other types of transformations using orthogonal function systems other than trigonometric function systems. The present invention uses a Walsh function system other than a trigonometric function system to analyze audio signals. The transformation using the Uoluyu function system is characterized in that the transformation matrix has only +1 and -1 elements, and its operations can be realized only by addition and subtraction. Furthermore, there is a high-speed Walsh-Hadamard transform as a conversion method using the Walsh function system, which is much simpler than conventional analysis methods, making it possible to downsize and cost down the entire speech recognition device. become. Hereinafter, the already well-known Walsh function and conversion method will be briefly explained. First, the Walsh function will be explained. The Walsh function only takes values of +1 and -12
Value function, normalized interval with period T, -1/2θ<
At 1/2 (θ=t/T), it is defined by the difference equation as follows. Wal (2j+p・θ)=(-1) [ ^j/2 ] ^+p {wal
(j・2(θ+1/4)] +(−1) ^j+p wal[j・2(θ−1/4]} p=0or1, j=0, 1, 2...... wal[0・θ ]=1 −1/2θ＜1/2 wal[0・θ]=0 θ＜−1/2・θ1/2
...(1) The even function of the Walsh function defined by the above equation is
cal(i・θ). If an odd function is expressed as sal(i·θ), the following relationship exists between them. cal(i・θ)=wal(2i・θ) sal(i・θ)=wal(2i−1・θ)……Formula (2) The Walsh function system defined above is
In order to form a complete orthonormal function system, the function x(θ)
is similar to the Fourier series expansion using a trigonometric function system,
Series expansion is possible using the Walsh function system. Approximating this series term to N(2 ⁿ ) terms results in the following. FIG. 3a shows the Walsh function system in the case of N=8 (=2 ³ ) and the correspondence between the Walsh function system and the trigonometric function system. cal is trigonometric function cs
, sal corresponds to sin. In addition, in the case of the Walsh function, what corresponds to the "frequency" in the trigonometric function system is the "alternating number", and the alternating number spectrum is expressed by the following formula. S _i =√ _i ² + _i ² s _p = a _p ......Equation (4) In other words, from the input signal, find a _i and b _i shown in Equation (3), and from Equation (4) above, Once the alternating number spectrum is obtained, the input signal can be transformed into the alternating number domain in the same way as the input signal is transformed into the frequency domain in the case of Fourier transform. On the other hand, the rows of the Hadamard matrix defined by the following equation can correspond to the Walsh function. Figure 3c shows the matrix for N=8 according to equation (5),
(natural form Hadamard matrix), and FIG. 3b shows the 8th-order Walsh-Hadamard matrix based on equation (1). The correspondence between the rows of the Hadamard matrix and the Walsh function is indicated by 30. Letting the input vector (N) and the output vector be 1B (N), the Hadamard transformation is expressed as 1B (N) = 1/nH (n) 〓 (N) ......Equation (6), and the Hadamard matrix is By associating the line with the Walsh function, 1B of the above equation
(N) can be thought of as a vector whose elements are a _i and b _i . Figure 4 shows a permutation matrix T(N) for performing this correspondence, and this matrix is used as the input vector 〓(N) or the output vector 1B
By multiplying by (N), a _i and b _i can be obtained in the order of the number of alternations. Note that an algorithm based on Gray code has been considered to obtain this permutation matrix. The above will be summarized and explained according to FIG. As shown in the figure, the audio signal 50 is sampled to obtain a time sequence 51 of x(0) to x(7). This is multiplied by the permutation matrix shown in FIG. 4 to obtain 52. 52
By multiplying by the natural form Hadamard matrix of Figure 3c, the respective coefficients are obtained in the order of the number of alternations, as shown in 53, and according to equation (4), the number spectrum of alternations as shown in 54 is obtained. You can finally ask for it. Even if the time series 51 is multiplied by the Walsh-form Hadamard matrix shown in FIG. 3b without suddenly being multiplied by a permutation matrix, each of the series 53 can be obtained in the order of the alternating numbers. However, when trying to perform the fast Hadamard transformation shown below, the amount of calculation is greatly reduced by multiplying by a permutation matrix and transforming by a natural form Hadamard matrix. FIG. 6 shows the 8th order fast Hadamard transform. There are various methods to perform the Hadamard transformation of natural forms at high speed, and I will show one example. Only the arrow in Figure 6 (61 etc.)
62 indicates addition, and 62 indicates subtraction.
In this case, the amount of calculation for addition and subtraction is 24 (3 x 8) times,
If Sumiko's method is not used, 64 (8 x 8) times are required, which is a huge reduction. The amount of calculation in the N-th case is Nlog ₂ N. As mentioned above, the already well-known Walsh function and
The conversion method was explained. In the explanation, mainly N
=8, but even if N is increased by a multiple of 2, the same processing method will be obtained. Speech recognition using the Walsh-Hadamard transform will be described below with reference to FIG. The audio signal input from the microphone 71 is amplified and sent to the A/D converter 73, and depending on the sampling time from the oscillator 75, the audio signal is sequentially converted to a digital signal and sent to an area 74 or an area 76 for processing. It's crowded. 89, 88, a double switch, when switch 89 is on the _i1 side, switch 88 is on _{the 2} side, and when 89 is on the _i2 side, 88 is on _{the 1 side.}
Be on your side. The control of these switches is controlled by oscillator 7.
5, and every time a certain number N are input into the area or inside, a control signal 8 is sent.
1 switches the switch. Sorting section 7
7 is a place where the contents of area 74 or area 76 are multiplied by a permutation matrix as shown in FIG. The rearranged input vectors are subjected to Hadamard transformation in a high-speed Hadamard transformation unit 78 (processing similar to that shown in FIG. 6). From the output of 78, the alternation number spectrum as shown in equation (4) is determined by 82. Processes 77, 78, and 82 are performed using area 74 or area 76. When rearranging, Hadamard transform, and spectrum are being calculated for one area, the output from the A/D converter is input to the other area, and switches 88 and 89 are used to switch between these. It's because I'm getting used to it. If the sampling time is 100μs・N=256 (256th order Hadamard), then 25.6ms (100μ×
256), the alternation number spectrum of the audio signal is found. Including the DC component, there are 82 spectra of 129 points.
It is calculated by In step 79, the total square sum of the spectrum or the square root of the total square sum, excluding the DC component, is calculated, and the audio power is accumulated every 25.6 ms. The speech word cutting section 80 detects whether it is a sound period or a silent period from the speech power every 25.6 ms, and sends a signal 91 to the compression section 83 indicating whether or not to activate the compression section 83. . Further, this 80 generates a signal 92 for removing silence before and after the pronounced spoken word and cutting out the spoken word.
is sent to the input pattern area 90. Some spoken words have periods of silence, so
(For example, there is a silent period between ``to'' and ``kiyo'' in ``towokiyou''.) In general, the existence of a silent period does not mean the end of the spoken word. The audio word ends after the silent period continues for a certain period of time. From the time the sound period appears, the compression section 8
The output from the compression unit 83 is stored in the input pattern area 90, and the output from the compression unit 83 is stored sequentially until the end of the audio word. Processing is performed within the input pattern area using signal 92 so that the aforementioned silence near the end of the audio word is removed and only the audio word is ultimately stored within the input pattern area. . The information in the alternation number spectrum section 82 cannot be used as a voice pattern as it is, and is subjected to compression processing in step 83. The operation of the compression unit 83 is such that compression of the amount of information is independent of the magnitude of the audio power.
This is normalization of the amount of information. When N=256 and sampling time=100 μsec, an alternating number spectrum of 129 points is obtained every 25.6 ms, but these vary depending on the loudness of the sound and have a large amount of information. Therefore, the aforementioned compression and normalization are necessary. FIG. 8 shows a more detailed block diagram of the compression section 83. The channel dividing unit 95 divides N spectra (excluding the DC component spectrum) into M spectra, calculates the sum of the divided spectra, and converts the N spectra into M channel information. do. In Figure 9, N=256・Sampling time=100μ
This shows the case where the channel is divided into M channels. for example,
on channel 8

Perform [Formula] and set the value of channel 8. S _i follows equation (4). The normalization unit 96 is a normalization unit for converting information unrelated to audio power, and follows the formula below. L _k is an integer value, and by doing so, it becomes independent of voice power. For example, if K is 16, the maximum value of L _k will be 16, and L _k
Therefore, it is sufficient to prepare 4 bits. In this way, the amount of information every 25.6 ms becomes 4 bits x 16 channels = 64 bits. If the A/D converter is 8 bits, the amount of information is 8 bits x 256 pieces.
This means that it is compressed to 64 bits and normalized to something that has nothing to do with audio power. The method of division is equal division with a fixed division width, or centering on the geometric mean of the divided areas.
Logarithmic division can be considered so that the ratio of adjacent center values is almost constant, but the latter division is preferable. This is because if the sampling time is 100μs・N=256, the spectral resolution is 39Hj, and a variable number spectrum can be found in units of 39Hj.On the other hand, the first formant of speech exists in the low alternation number region, so if it is divided into equal parts, This is because the amount of information of a channel calculated from a high alternation number region becomes extremely small, making it difficult to distinguish the characteristics of the audio signal. Also, the DC component is due to amplifier drift or
It is better not to use it for power calculation or spectrum division based on offsets, etc. Note that when dividing a spectrum into channels, it is more natural to obtain the sum of squares of the spectra and then take the square root, rather than simply adding the spectra. However, in this case, processing becomes complicated. In experiments, simple addition of spectra was sufficient. Alternatively, it is possible to simply calculate the sum of squares and normalize without taking the square root, but in either case, the experimental results (sum, sum of squares, square root of the sum of squares) are not very different. Ta. Summation is advantageous in terms of ease of processing. Furthermore, if such compression is performed in a place where the audio power is low, a stable audio pattern cannot be obtained in terms of the S/N ratio. Therefore, voice presence and silence are detected, and areas where voice power is low are treated as silent periods without compression, and the voice pattern is set to a certain predetermined value (for example, all zeros). Noise is actually superimposed on the silent section of the audio signal, and if normalization processing is performed on this noise as it is, this noise will be amplified to the same level as the original signal component. Moreover,
Since the position of occurrence of this noise component on the spectrum is not constant, it is generally different from the position of occurrence on the spectrum of the noise component in a pre-registered voice pattern. For this reason, when matching the audio pattern based on the input audio signal with the audio pattern registered in advance, a very large distance difference may appear, causing misrecognition. This problem can be solved by setting the interval to a constant value. The compression section 83 has been described above. The subsequent processing is similar to that described in FIG. 1 of the conventional example. By setting the switch 87 to the e side, the input pattern is transferred to the registered pattern area 84 and the voice word is registered. At the time of recognition, set the switch 87 to the R side and input pattern area 90.
Pattern matching is performed between the pattern within and the registered pattern within the registered pattern area 84. The determining unit 86 receives the output from the pattern matching unit 85, selects the category and number of the registered pattern with the most similarity, and examines whether the similarity is appropriate. If it is valid, the category number is output and the input speech word is recognized. Equation (4) is used as an alternating number spectrum, but this calculation requires square and square root calculations.
Also, the calculation of power also requires squares and square roots. While the Hadamard part calculations are simply addition and subtraction, these calculations require calculations of squares and square roots, which complicates the processing and poses a problem. For the alternation number spectrum, the following formula (8) is defined as the sum of absolute values, and the calculation of the alternation number spectrum part in Figure 7 is performed using this formula.As for the power, the sum of the alternation number spectra is determined and the experiment is performed. I did it, but
There was almost no difference in the recognition results compared to equation (4). S _i = | a _i | + | b _i | ......Formula (8) Also, by converting into absolute values in this way, the alternation number spectrum can be obtained without going to the trouble of processing the above formula (8). Get better. All you have to do is convert the values after the Hadamard transform to absolute values for all values, and then process the channel division all at once.
It is sufficient to divide the spectrum using the spectrum number obtained by multiplying the spectrum number to be divided by 2 in the table below, and the processing becomes extremely simple.

[Table] However, when calculating the audio power from the absolute value,
There is a slight problem in detecting sound/silence. Since there is no clear correlation between the calculation of the power based on the sum of absolute values and the actual voice power, the presence or absence of voice is not detected correctly, resulting in failure to recognize it correctly. FIG. 9 shows processing using absolute values and solving the problems caused by the above-mentioned absolute values. The sum of the absolute values of the output from the Hadamard transform unit 78 is determined by 100, and without determining the alternation number spectrum of equation 8, channel division is suddenly performed by 95 as described above, and the normalization unit 96 converts the voice pattern into Convert. The power is calculated from the audio signal itself before the Hadamard transform, as shown in the figure, without using the Hadamard transform and taking the absolute value. A/
The power is determined from the D-converted audio signal x(n) using the following equation. By using this value, it becomes possible to detect voice presence and silence more accurately. In FIG. 9, other blocks are the same as in FIG. 7. However, by eliminating square calculations or square root calculations, it became possible to perform all processing with simple additions and subtractions, making it possible to miniaturize equipment and reduce costs. As described above, according to the present invention, the information compression means for compressing the information of the transform values obtained by the orthogonal transform means and converting them into speech patterns is configured to compress the absolute value of each spectrum of the transform values obtained by the orthogonal transform means. a channel information exchange means that divides the absolute value of each spectrum into M groups, calculates the sum for each divided region, and converts it into M channel information; and normalizing means for normalizing the channel information so that In addition, the normalization method avoids unnatural normalization of silent sections in the audio signal, resulting in a significant improvement in recognition accuracy.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a speaker-specific speech recognition device, Fig. 2 is a block diagram of the frequency analysis section of the device in Fig. 1, Fig. 3 is a diagram for explaining the Walsh-Hadamard function, Figure 4 is a diagram showing a permutation matrix for converting a natural-form Hadamard matrix into a Walsh-form Hadamard matrix, Figure 5 is a diagram to explain the conversion method to the alternating number domain, and Figure 6 is an example of fast Hadamard transformation. FIG. 7 is a block diagram of a speech recognition device according to an embodiment of the present invention.
FIG. 8 is a block diagram of the compression section of the same apparatus, and FIG. 9 is a block diagram of another embodiment. 71...Microphone, 72...Amplifier, 73...
A/D converter, 75... oscillator, 74... area, 76... area, 77... sorting unit, 7
8...Fast Hadamard transform unit, 79...Power calculation unit, 80...Speech word extraction unit, 82...Alternating number spectrum unit, 83...Compression unit, 84...Registered pattern area, 85...Pattern matching Department,
86... Judgment unit, 87, 88, 89... Switch, 90... Input pattern area.

Claims (1)

[Claims] 1. Orthogonal transformation means for performing orthogonal transformation of an audio signal;
Information compression means for compressing information of the transformed value obtained by the orthogonal transformation means and converting it into a speech pattern;
A pattern matching means performs pattern matching between the voice pattern and a pre-registered voice pattern, and the information compression means takes an absolute value for each spectrum of the transformed value obtained by the orthogonal transformation means, Channel information converting means divides the absolute value of each spectrum into M groups, calculates the sum for each divided region, and converts the sum into M channel information, and the total sum for each divided region becomes constant. normalizing means for normalizing the channel information so that the channel information is normalized, and the normalizing means replaces a silent section in the audio signal with a constant value.