JPH0346838B2 - - Google Patents

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a speech recognition device, and in particular, the present invention relates to a speech recognition device, in particular, when an analysis frame is in a compressed format during training (at the time of registration).
The present invention relates to a compressed DP type speech recognition device that recognizes input word speech by performing time normalization, so-called pattern matching, on a specific speaker's word, between a standard pattern registered in the system and an input speech pattern input each time a recognition process is performed.

[Prior art]

Multiple word sounds uttered by the identified speaker are analyzed at a predetermined analysis cycle, that is, each analysis frame, to find a standard pattern regarding the distribution of extracted feature parameters, and then this is compressed using the DP method. Then, the vector with the minimum vector distance between the two, i.e., the one with the minimum distortion of the recognized speech, is determined by tracing the DP path. The desired compressed DP type speech recognition device for a specific speaker alone is currently well known.

Such a compressed DP type speech recognition device uses a compressed standard pattern that should be time-normalized with the input pattern, so it is possible to reduce the standard pattern memory that stores the standard pattern and perform time normalization. It has the feature that the amount of processing required for this process can be reduced, and the scale of the hardware can also be correspondingly simplified.

The above-mentioned time normalization is performed when the input pattern and the standard pattern are the voice patterns of the same person, and when assuming a normal usage environment, the largest variation factor is the variation in speaking speed. The purpose is to eliminate the different and complex extensions and contractions of emergency forms that occur in the vowel and consonant parts. The purpose of this time normalization is to find the optimal mapping function between the input pattern and the standard pattern, and to align the time axis of the standard pattern with the time axis of the input pattern. This is done by implementing the DP method using the pattern spacing as an evaluation measure and finding the one that minimizes this.

Although such compressed DP type speech recognition devices have various features as mentioned above, since rectangular approximation is used as the standard pattern compression method, the ratio between the amount of calculation and the reduction in the amount of distortion obtained is The disadvantage is that there is an inevitable limit to compression efficiency.

[Purpose of the invention]

The purpose of the present invention is to eliminate the above-mentioned drawbacks, and provide a compressed DP type speech recognition device that targets word speech of a specific speaker, in which DP is achieved using optimal trapezoidal approximation as a standard pattern compression means, and time regularization is achieved. This is done by making the input pattern correspond to a compression standard pattern, compressing it and then normalizing it, or by expanding the standard pattern to match the input pattern and performing speech recognition processing using a DP matching method to find the DP path. An object of the present invention is to provide a speech recognition device with significantly improved compression efficiency.

[Structure of the invention]

The device of the present invention performs pattern matching based on time normalization between a standard pattern registered in a format in which the analysis frame is compressed using the DP method and an input pattern of word speech by the specific speaker. In a compressed DP type speech recognition device that performs speech recognition, there is a standard pattern compression means that compresses a standard pattern based on optimal trapezoidal approximation using the DP method, and a standard pattern compression method that compresses an input pattern to fit the standard pattern and then converts it to a standard pattern. Time normalization is performed using the pattern length, or time normalization is performed by stretching the standard pattern to match the input pattern, and in this time normalization, the amount of distortion with respect to the standard pattern is used as an evaluation criterion to find a DP path that minimizes this. and a time normalization means for calculating the time.

〔Example〕

Next, the present invention will be explained in detail with reference to the drawings.
FIG. 1 is a block diagram showing an embodiment of a speech recognition device according to the present invention.

The embodiment shown in FIG. 1 includes an acoustic analyzer 1, a switch 2, a compression processor 3, a standard pattern memory 4, a pattern matcher 5, a minimum distance searcher 6, and the like.

In voice recognition of specific speaker word speech, it is first necessary to store in advance standard patterns related to a plurality of words uttered by a specific speaker, and this is carried out as follows.

In other words, the acoustic analyzer 1 is an LPF (Low Pass
Filter), A/D converter, LSP (Line
Spectrum Pairs (Line Spectrum Pairs) with a built-in analyzer, etc., input audio to an LPF with a predetermined cutoff frequency.
After the data is railtared, it is sampled at a predetermined sampling frequency, converted to digital data, and then sent to an LSP analyzer.

LSP analyzer is LPC (Linear Prediction)
It also has a linear predictive coefficient (linear prediction coefficient) analyzer, and uses known techniques from LPC coefficients such as PRRCOR (partial autocorrelation coefficient), which are analyzed and extracted by the LPC analyzer for each time frame of a preset analysis cycle, that is, each analysis frame. For example Newton
An LSP coefficient sequence of a predetermined order is determined for each analysis frame by a method of solving higher-order equations using the iterative method, and this is sent to the switch 2. The LSP coefficient obtained in this way is a parameter representing the resonance characteristics of the vocal tract, and is a parameter based on the line spectrum frequency of the transfer function of the vocal tract filter when the glottis is virtually completely open and completely closed, and is a characteristic parameter treated in the frequency domain. It is also well known that

During standard pattern training (registration), the switch 2 is switched to the connection state shown by the dotted line, and therefore the LSP parameters regarding the words of the specific speaker are supplied to the compression processor 3.

The compression processor 3 performs frame compression processing using optimal trapezoidal approximation of this LSP parameter as follows.
Implemented using DP method.

It is well known that in addition to optimal linear approximation, methods such as optimal rectangular approximation and even optimal trapezoidal approximation are being used in frame compression processing in fields such as variable-length frame type linear predictive pocoders.
Among these optimal approximations, optimal rectangular approximation is often used as a basic means of compression utilization in speech recognition devices. This is because the expected reduction in the amount of calculation as a result of compression is significant compared to optimal linear approximation, but on the other hand, due to the nature of optimal rectangular approximation, there is a limit to the degree of approximation that can be obtained, and therefore the amount of distortion is also less than optimal linear approximation. This will increase significantly compared to the previous year.

On the other hand, although the optimal trapezoidal approximation does not reach the same level of computational complexity as the optimal rectangular approximation, the degree of approximation is much greater, and therefore the amount of distortion can be improved to almost the same as the optimal linear approximation, significantly improving compression efficiency. be able to.

FIG. 2A is a principle diagram for explaining the principle of optimal rectangular approximation, and FIG. 2B is a principle diagram for explaining the principle of optimal trapezoidal approximation.

In FIG. 2A, for example, LSP parameters are extracted as feature vectors from input speech a for each analysis frame. In optimal rectangular approximation, K frames of LSP parameter vectors that are continuously supplied in this way are collectively treated as one new processing section, and the maximum number M (1 < M <K) feature parameters and the analysis frame that each of the M feature parameters should represent is selected, and the series of analysis frames approximated by such selection is shown in Figure 2A. This results in a variable length frame based on the optimal rectangular distance shown in b.

The maximum number M of feature vectors to be set for each processing section described above can be arbitrarily set between 1 and K in consideration of compression efficiency. In this way, the maximum number M of feature vector groups set for each segment is determined by using the DP method, and what combinations of M feature vectors each represent which analysis frame. In this case, DP is performed using the distortion caused by rectangular approximation as an evaluation measure, and this distortion is calculated using the representative M
It can be easily determined by repeating the method of determining for each processing section which analysis frame each group of feature vectors represents and minimizes the distance between this feature vector and its rectangular approximate feature vector.

However, such optimal rectangular approximation using DP has the above-mentioned problem regarding the limit of compression efficiency. Therefore, in this example, the optimal trapezoidal approximation is
We aim to significantly alleviate this problem by using the DP method.

Optimal trapezoidal approximation uses the fact that an excessive portion of audio information that changes rapidly has a nearly constant length of time, usually about 20 mSEC, to calculate this excessive portion to a time length that corresponds to a preset number of analysis frames. Expressed as This is an optimal approximation that uses a trapezoidal function instead of a rectangular function, and the degree of approximation is essentially higher than the optimal rectangular approximation. In such optimal trapezoidal approximation, while selecting a group of representative feature vectors that minimize the vector spatial distance from the original feature vector using the DP method, the selected representative feature vectors are separated by the above-mentioned constant time length so-called slope interval. Processing is basically done by a method of expressing the data, and it has recently been widely used in fields such as variable-length frame vocoders, but in this embodiment, so-called piecewise approximation, in which processing is performed for each processing section, is used. This method is characterized by treating each word that should be registered as a standard pattern as one processing section, performing optimization using trapezoidal approximation using the total amount of distortion as a guideline, and therefore not setting a fixed number of frames to be selected. .

FIG. 2B is a diagram of the principle of optimal trapezoidal approximation having such characteristics, where curve c is an approximate trapezoid with one-word speech by a specific speaker, trapezoid d is an approximate trapezoid with one-word speech c as one processing section, and point P ₁ , _P2 , _P3 , _P4, etc. indicate representative feature parameter groups, and variable length frame sections ₁ to ₄ are represented by these representative feature parameters.
A slope section of a predetermined length of time is set between them. Determining the optimal trapezoidal approximation is nothing but finding a trapezoid that minimizes the area indicated by the diagonal line formed by the trapezoid d and the one-word sound c using the DP method. Furthermore, as is clear from FIG. 2B, the optimal trapezoidal approximation obtained in this way has a much higher degree of approximation than a rectangular trapezoid, and therefore requires far fewer representative feature vectors to be set, improving compression efficiency. becomes.

The explanation will be given by returning to FIG. 1 again. The compression processor 3 performs such optimal trapezoid approximation processing on the extracted feature parameters and LSP parameters by analyzing each word voice uttered by a specific speaker, and sends these as a standard pattern to the standard pattern memory 4 for storage. .

With the standard pattern stored in this way, the switch 2 is switched to the recognition side, and the specific speaker utters which word sound is stored in the standard pattern memory 4 via the input terminal 101, and this is transmitted to the acoustic analyzer. 1 to extract the LSP parameters and then supply them to the pattern matcher 5.

The pattern matching device 5 is equipped with a spectral distance measuring device, an interpolator, etc., and performs a DP method that uses spectral distance as an evaluation measure to generate a standard pattern between a standard pattern and an input pattern compressed to match the standard pattern. Perform time normalization by length as follows.

The standard patterns read out from the standard pattern memory 4 one after another are set by the built-in interpolator of the pattern matching device 5 to set interpolated values between the DP-compressed representative feature vectors, and then by the built-in spectral distance measuring device. I measured it. Time normalization is performed using the standard pattern length through the DP method using spectral distance as the evaluation measure.

There are two ways to time normalize the DP compressed standard pattern and the input pattern.
A method of thinning and compressing the data to match the compressed standard pattern, and then normalizing the time using the standard pattern length, or a method of repeatedly generating representative feature vector intervals in response to the input pattern and stretching the standard pattern to achieve time normalization. However, in this embodiment, time normalization is performed using the former method. In order to align the time axes of the compressed standard pattern and the uncompressed input pattern, that is, to find the image function between the standard pattern and the input pattern and perform time normalization, the input pattern must be thinned out. There is no problem in either stretching the standard pattern to match the input pattern or conversely stretching the standard pattern to match the input pattern.

FIG. 3 is a pattern matching principle diagram showing the principle of pattern matching processing in the embodiment of FIG. 1. The description of the embodiment will be continued below with reference to FIG.

In FIG. 3, a standard pattern 1001 is one of the standard patterns formed by the above-mentioned optimal trapezoidal approximation and using the DP method, and an input pattern 1002 should be pattern matched with the standard pattern 1001, that is, time normalized. is the input pattern to be aimed at.

Now, considering the ij plane as shown in FIG. 3, the i direction corresponds to the standard pattern 1001, the j direction corresponds to the input pattern 1002, and the vertical lines indicated by black circles are actually measured LSP parameters. Pattern matching device 5
The interpolation LSP parameters shown by the X mark between these solid lines are set using the built-in interpolator as shown by the dotted lines.

In addition, the input pattern is obtained as a solid line perpendicular to the j direction, and the LSP parameter vector is obtained as a solid line perpendicular to the j direction at every analysis period t ₀ of the acoustic analyzer 1. The spectral distance between the LSP parameter vectors of both patterns for each of these corresponding positions is determined for every set of corresponding points, and
If the DP path that minimizes the distance between the two patterns is found using the DP method using this as an evaluation measure, these will indicate the spectral distance between the two patterns. However, when using this DP method to find a DP path that minimizes the distance between both patterns, the extreme time axis variation range between patterns that does not actually occur is eliminated, and the DP process is usually called a matching window. The processing is limited to the areas _l1 and _l2 .

In FIG. 3, for example, Q ₁ on the ij plane
The spectral distance shown by the arrow between the LSP parameter vector of the standard pattern and the corresponding LSP parameter vector of the input pattern is measured. The path in d at these spectral distances, 45
The degree line corresponds to a constant time length section, so-called slope section, in the optimal trapezoidal approximation, and the spectral distances shown by the straight line and analysis line including this are measured. This spectrum measurement is performed for the correspondence between the standard pattern and the input pattern for all vertical and horizontal intersection positions including the dotted line within the processing range limited by the lines _l1 and _l2 , and with the standard pattern length. The conditions for implementing this standard pattern length vary depending on the degree of DP compression of the standard pattern, but in the case of this embodiment, the implementation is performed for a total of six corresponding combinations as shown in FIG.

In this way, standard pattern 1001 and input pattern 10
02, input pattern 1002 is converted to standard pattern 1.
001 and time-normalized using the standard pattern length is obtained as the DP path g, and the spectral distance between this time-normalized input pattern and all standard patterns is successively determined by the minimum distance searcher. 6.

Generally, the distance between two LSP parameter spectra is represented by the spectral distance D _sr shown in the following equation (1).

D _sr = 1/π∫〓 ₀ {S _s (ω)−S _r (ω)} ² dw ……(1) Equation (1) is also usually transformed into the following approximate equation (2). used.

D _sr = _N 〓 ^k=1 W _K {P _K ^(s) −P _K ^(r) } ² ...(2) In equations (1) and (2), s and r are analysis frames or processing sections (blocks) number, S _s (ω), S _r
(ω) is the logarithmic spectrum of the analysis frame or block s, r as a function of frequency ω, P _K ^(s) ,
P _K ^(r) is the analysis frame or block s and r
where W _K is the K-order LSP frequency spectral sensitivity.

Time normalization using the above-mentioned DP method, in other words, DP pattern matching, is based on the above-mentioned calculation basis, and calculates the space vector distance between re-patterns by thinning out the input pattern with respect to the standard pattern.
This calculation is performed on the input patterns for each of all standard patterns, and the results are sequentially supplied as spectral distance data to the minimum distance search unit 6 together with the designated number data of the standard patterns.

The minimum distance search device 6 stores the spectral distance data for each input pattern for each input standard pattern in the built-in memory, determines the magnitude relationship between them, searches for the one with the minimum value, and retrieves the minimum spectral distance data. The standard pattern information is supplied as a recognition result to the output terminal 601 from the standard pattern designation number data provided, and speech recognition is thus performed through DP pattern matching with the standard pattern by optimal trapezoidal approximation.

In addition, in the above-mentioned embodiment, the feature parameters of audio words that should be stored as standard patterns include
Although LSP parameters are used, it is clear that this can be done similarly using other feature parameters, such as the cepstrum, which is expressed by the inverse logarithmic transformation of the spectrum of word sounds. .

In addition, in this example, as a time normalization method, the input pattern is compressed to match the standard pattern.
This is an example of normalizing using the standard pattern length, but this is the same even if the compressed standard pattern is stretched to match the input pattern and time normalization is performed using DP. In this case, it can be easily implemented in a format in which the compressed standard pattern is read out and repeated in time to match the input pattern.

〔Effect of the invention〕

As explained above, according to the present invention, word speech by a specific speaker is recognized through time normalization of a standard pattern registered in a compressed format of an analysis frame and an input pattern of word speech uttered by a specific speaker. The speech recognition device is equipped with a standard pattern based on the optimal trapezoidal approximation of the compressed DP type obtained using the DP method, and in time normalization of the input pattern and the standard pattern, the input pattern is adjusted to match the standard pattern. The feature vector distance between the two patterns is used as an evaluation measure.
By providing a means for implementing the method based on the DP method, it is possible to realize a speech recognition device that can significantly improve compression efficiency and significantly reduce the memory capacity of standard patterns.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention;
Figure 2A is an optimal rectangular approximation principle diagram showing the principle of optimal rectangular approximation, Figure 2B is an optimal trapezoidal approximation principle diagram showing the optimal trapezoidal approximation principle, and Figure 3 is time normalization in the example of Figure 1. FIG. 2 is a time normalization explanatory diagram for explaining. 1...Acoustic analyzer, 2...Switcher, 3...Compression processor, 4...Standard pattern memory, 5...Pattern matching device, 6...Minimum distance search device.

Claims (1)

[Claims] 1 The analysis frame is dynamic programming (hereinafter referred to as dynamic programming).
A standard pattern of word speech by a specific speaker registered in a compressed format using a method (abbreviated as DP) and an input pattern of word speech by the specific speaker are pattern-matched by time normalization. In a compressed DP type speech recognition device that performs speech recognition for words, there is a standard pattern compression means that compresses a standard pattern based on optimal trapezoidal approximation using the DP method, and a standard pattern compression means that compresses an input pattern to match the standard pattern and then creates a standard pattern. Time normalization is performed using the pattern length, or time normalization is performed by stretching the standard pattern to match the input pattern, and in this time normalization, the amount of distortion with respect to the standard pattern is used as an evaluation criterion to find a DP path that minimizes this. 1. A speech recognition device comprising: time normalization means for obtaining time by: