JPS62119600A - Word voice recognition equipment - Google Patents

Abstract

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

Description

[Detailed Description of the Invention] [Summary] Among various input terminal devices, speech recognition technology is a technology for realizing a so-called voice input device that inputs voice as it is. In realizing a recognition device, it is relatively easy to recognize speech uttered by dividing it into words, and in particular, word recognition devices for specific speakers have been put into practical use in various fields.

In this type of word speech recognition device, when comparing the standard pattern of a word with the feature pattern of the input speech, the acoustic features of some phonemes change over time, so even if the word is the same from the same speaker, the acoustic features may change over time. Matching is achieved by correcting the expansion and contraction of time length.

This expansion/contraction method includes nonlinear and linear methods, and the nonlinear method is generally dynamic programming (dynamic programming D).
P) is used. In the DP method, the conversion function for time axis expansion/contraction is selected according to an optimization algorithm so that the error between the standard pattern and the input pattern is minimized, so recognition performance can be improved, but the amount of processing is increased. The problem is that it takes time. On the other hand, since the linear time warp matching method does not include an optimization algorithm, the amount of processing is small and the processing time can be reduced, but there is a problem that recognition performance deteriorates.

In such a word speech recognition device, the present invention divides a target word into subcents, and stores selection information for each subset as to whether to use the DP method or the linear time warp matching method in the form of a table in a memory circuit. . The present invention is characterized in that, for each subcent, a selection is made with reference to the table as to whether to use the linear time expansion/contraction matching method, which requires a small processing amount of matching, or to use the DP method, which requires a large processing amount of matching but has good recognition performance. In this way, you can consider in advance whether the number of words to be recognized is large or small, when misrecognition would be fatal, when misrecognition is not a big problem, or when it is easy to recover. By selecting either the DP method or the linear time warp matching method for each subset according to the above, it is possible to cope with cases where requirements for response time are severe and cases where requirements for recognition performance are severe.

[Industrial application field]

The present invention relates to a speech recognition device that is the basis for realizing a speech input device, and is particularly aimed at specific speakers who recognize speech uttered in units of syllables, words, etc. in lit and m units. 11'L A related to the configuration of a speech recognition device.

Furthermore, the present invention divides each voice into ``subsesoto'' and uses a linear time warping method that requires a small amount of processing for matching, or a DP method that requires a large amount of processing for matching but has good recognition performance. The present invention relates to a configuration of a word speech recognition device that can selectively utilize words.

[Traditional technique]

With the progress of integration technology, among various input terminal devices used as man-machine interfaces, voice input devices that input voice directly have been put into practical use. The use of a voice input device has the advantage that the input speed of input data to be information can be increased, and even a person who is not skilled in operating an input device can input input data by voice. Speech recognition technology is the basis for realizing this voice input device. In speech recognition technology, it is extremely difficult to recognize sentence sounds that are naturally uttered by humans because their acoustic characteristics are complexly modified by accents and intonation. utter it,
So-called discrete word recognition devices, which recognize individual syllables and words, were first put into practical use. The number of words to be recognized is usually on the order of several hundred words, but even if the number of words to be recognized is as small as this, it can be effectively used for product inspection in factories, etc. In such a discrete word recognition device, a word recognition device that detects a break for each word and recognizes the words in order,
If a standard pattern obtained by analyzing the voice of a specific person is used, a high recognition rate can be obtained for the voice input of that person. Therefore, when constructing a standard pattern, a speech recognition device for a specific speaker uses a learning function that is created for each speaker to recognize a specific speaker. A recognition rate of 99% or more can be obtained.

There is always a matching section in the recognition device that compares and matches the standard pattern with the characteristic pattern of the input voice currently input to the device. Here, the input sound 6 one-voice pattern is a time series created by extracting features of the input source sound at every fixed frame period. On the other hand, the standard pattern is stored in the dictionary unit as a word dictionary, and is a time series of features extracted from the source speech at regular frame intervals through the learning process. By inputting human speech and comparing and collating each pattern in the word dictionary, it is determined that the current input speech is a specific word.

Conventionally, there are two types of matching methods: linear and nonlinear. In other words, when matching li words, when comparing the input speech pattern and the standard pattern, even if the same word is produced by the same speaker, the time axis may be expanded or contracted. It is necessary to perform normalization. Generally, this expansion and contraction of the time axis is nonlinear expansion and contraction. Linear matching takes correspondence on the time axis at a constant expansion/contraction rate, which simplifies the processing method, but has the problem of lowering the recognition rate. On the other hand, when nonlinear matching is performed by adjusting nonlinear expansion and contraction, a conversion function for normalizing the time axis is selected so that the error between the cover pattern and the standard pattern is minimized.

When performing such optimization, a huge amount of calculation is required because a conversion function is selected so that the error becomes the minimum value for every combination of each time-series data of the input cover pattern and the standard pattern. Therefore, the general method for reducing this amount of calculation is to use dynamic programming, or DP matching, to significantly reduce the amount of calculation. In comparison, the amount of calculation is considerably large, and the time required for recognition is longer than in linear matching.

[Problem that the invention seeks to solve]

The present invention eliminates such shortcomings of conventional word speech recognition devices, divides each word of speech into subsets, and uses a linear time warping matching method that requires less matching processing for each subset, or a matching process. To provide a terminology speech recognition device that can deal with both cases with severe demands on response time and cases with severe demands on recognition performance by referring to a table and selecting one of the DP methods having a large amount of data but with good recognition performance. It is something to do.

[Means to try to solve problems]

According to the present invention, an acoustic analysis section inputs an audio signal, extracts speech features, and performs section detection; a dictionary section that stores word standard patterns analyzed in advance through the acoustic analysis section; a first matching unit that performs linear matching between the characteristic pattern of the audio signal outputted through the acoustic analysis unit and the standard word pattern of the dictionary unit; and the standard word pattern from the dictionary unit and the characteristics of the audio signal. a second matching unit that performs non-linear matching with a pattern; and a selection unit that selectively transfers the output of the acoustic analysis unit to the dictionary unit, the first matching unit, or the second matching unit; The acoustic analysis section, the dictionary section, the first and second collation sections, and a control means for controlling the selection means via a quadrupling processing section, and for each subset, a simple method with a small processing amount of collation is provided. It is constituted by a means for setting in advance in a table which one of the nine-way method by the first collation unit and the method by the second collation unit, which requires a large amount of verification processing but has good recognition performance, is to be used. This is achieved by providing a word speech recognition device featuring:

[For production]

Each word of the speech is divided into subsets, and a table showing a correspondence indicating whether to use the linear time warping matching method or the DP method is referred to for each subset.

〔Example〕

Next, the present invention will be explained with reference to the drawings.

The speech recognition device shown in Fig. 1 acoustically analyzes the input speech input 1, extracts the linguistic features of the words included in the speech input, and analyzes the linguistic features of the words included in the speech with respect to a specific speaker in advance. Standard patterns related to features are stored in the dictionary 7, and compared with the feature patterns of the currently input voice input, and recognition is determined based on the similarity.

When the voice input 1 inputted from the microphone is inputted to the preprocessing section 2, the high frequency portion is emphasized. Alternatively, if the subsequent processing is to be digitally processed, the analog audio input is converted into a digital signal in the preprocessing section 2 via an A/D converter.

The high-frequency emphasized audio input is acoustically analyzed in the parameter calculation unit 3, and in particular, the frequency spectrum envelope of the audio is calculated. The frequency spectrum I-envelope characteristic is analyzed using bandpass filters nY and an integer>jt smoothing circuit connected to each bandpass filter as shown in FIG. That is, the bandpass filter group BPF divides the audio frequency band into about 12 small bands. Rectifying the outputs of each of the 12 bandpass filters,
Furthermore, by smoothing, the amount of signal power that falls on each band component is output as a DC voltage value. n
The rectified outputs of the bandpass filters are n-dimensional vectors A.

A2...An, which represents the characteristics of the frequency spectrum envelope of the voice. Parameter calculation section 3
The output of is input to the section detection section 4, where the start and end of each word are detected using the dark power value. That is, the power of the input voice input is calculated, and if the calculated power exceeds the darkness value, it is the beginning of a word, and if the darkness value decreases from top to bottom, it is the end point of the word. In this way, each word is separated, and recognition processing can be performed for each word in turn. The output of the section detecting section 4 is inputted to the switching section 6, and the characteristic pattern of each word found by the parameter calculating section 3 and the section detecting section 4, that is, the characteristic pattern regarding the spectral envelope in particular, is sent to the dictionary section 7, the linear matching section 9 or It is selectively transferred to the DP verification unit 8. The standard pattern to be stored in the dictionary section 7 is a pattern obtained by acoustically analyzing words whose linguistic content related to a specific speaker is known through the preprocessing section 2, parameter calculation section 3, and interval detection section 4. be. The word standard pattern is expressed as a time series of feature parameters obtained by analyzing each recognized word over its entire duration. For example, if the duration length of word A is T, the standard pattern of word A within TA is usually recorded as time series data by time sampling the output of a bandpass filter. In other words, as shown in Fig. 3, each element of the matrix created by taking the duration length, or frame, on the horizontal axis and the number of channels in each band on the vertical axis is the output of the parameter calculation unit, that is, the spectral envelope of each channel. It is a value. A dictionary is constructed by preparing as many matrices as the number of words.

The feature of the present invention is that the linear matching section 9 or the DP matching section 8 selectively matches the similarity between the standard pattern configured in this way and the feature pattern of the human voice input 1 manually inputted from the microphone at the present time. There is.

Phonemes of speech that are input when comparing the similarity between the standard pattern stored in the dictionary section 7 and the characteristic pattern of the speech human input manually entered into the preprocessing section via the parameter calculation section 3 and the interval detection section 4 In some cases, the acoustic characteristics change over time. Moreover, even if the words are the same from the same speaker, there is an expansion or contraction in the temporal duration of the word, so by correcting the expansion or contraction of the duration, the standard pattern and the characteristic pattern of the voice input are the closest. It is necessary to compare. This is correction of the duration length and normalization of the time axis. For this normalization of the time axis, there are various methods of comparison and matching between the standard pattern and the characteristic pattern of the audio input. Now, the speech input to be recognized is analyzed by the parameter calculation unit 3 and the interval detection unit 4 using the same bandpass filter BPF used to analyze the standard pattern stored in the dictionary unit 7, and the output is The pattern obtained by sampling is X=x+ + x2 +...x
Ill. In other words, it is assumed that the input cover pattern X is composed of m time-series patterns. On the other hand, the same applies to the standard pattern, and the standard pattern Y is y+', Vz
. . . It is assumed that the pattern is composed of a time series pattern called yn. Note that each time-series pattern corresponds to a column of the matrix shown in FIG. 3, and therefore is expressed by a vector having each output of the bandpass filter as an element.

Now, when munching the input cover turn X and the standard pattern Y, the length of the input cover turn cannot be compared. In general, even voices generated by the same speaker are subject to expansion and contraction on the time axis, so it is necessary to normalize the time axis and compare them. Moreover, this expansion and contraction on the time axis is generally nonlinear expansion and contraction,
Either a non-linear matching method is adopted, which is performed according to non-linear expansion/contraction, or linear matching is forcibly matched on the time axis at a constant expansion/contraction rate.

The linear matching unit 9 uses a matching method that takes correspondence on the time axis at a constant expansion/contraction rate, and is a matching method that is fast in processing but lowers the recognition rate. On the other hand, nonlinear expansion/contraction matching adjusts nonlinear expansion/contraction to match the time axis, but dynamic programming, or DP matching, is used to perform this calculation.
The DP matching section 8 is a processing section based on this. For example, the time series pattern of input cover turn X is 8 from xl to xIl.
The standard pattern Y for that is yl to y.

When there are only five time-series patterns, linear matching and nonlinear matching determine the correspondence between each sample point as shown in FIGS. 4 and 5, respectively.

As shown in Figure 4 (al, (bl), in linear matching, when the subscripts of each sample of pattern Figure 4 (In bl, xl and xl are compared with the standard pattern yl, and
'31X6 is Ys.

By comparing X and x6 with y, the path of this correspondence relationship becomes a straight line, and therefore thinning is performed so that linear matching is achieved. The linear matching section 9 performs linear matching in this manner.

In the nonlinear matching shown in Fig. 5 (al, fbl), the correspondence relationship is nonlinear as shown in the fbl diagram. That is, xI + xl + x3 corresponds to y, and x4 corresponds to)'z + Xs and x6 correspond to y. , corresponding to X,
is ya + X a corresponds to y5.

In this case, the curve U becomes a nonlinear path. An optimal algorithm is used to select this route so that the optimal route is selected. This optimization algorithm generally uses the concept of least squares method, and a monotonically increasing function U is selected so that the error between the input pattern X and the standard pattern Y is minimized. Since all the correlations between the input time series pattern X and the standard time series pattern Y are calculated based on the least squares method, J! Determining an appropriate conversion function U is very time consuming. Therefore, dynamic programming (DP method) is generally used to significantly reduce the amount of calculation.

This DP method does not calculate the vector distance by combining the standard pattern and all the time series patterns of the human cover pattern, but it recursively calculates only the vector distance related to the neighboring time series patterns from the initial value of the transformation function U. The conversion function U is determined by sequential optimization. The I)P method, which selects the conversion function U that minimizes the error between the input pattern and the standard pattern, includes an optimization algorithm, so the amount of calculation is larger than the linear masoching method. , the recognition performance is very good because it is optimized with respect to time expansion and contraction. - Therefore, in the present invention, the linear time expansion/contraction matching method and the DP matching method are selected under the control of the switching unit 6 to optimize processing and recognition performance.

The recognition results obtained by the linear matching section 9 or the DP matching section 8 are transferred to the host computer 5 via the control section 10, where appropriate processing is performed. In the word speech recognition apparatus shown in FIG. 1, each part is controlled in accordance with control commands from the host computer 5 via the control unit IO.

As explained above, the linear time warp matching method is a method in which the amount of matching processing is small, but the recognition performance is inferior.
5 On the other hand, although the DP method requires a large amount of processing, it has good recognition performance. Therefore, in the word speech recognition device of the present invention, the switching unit 6 is controlled by commands from the host computer and operated based on a control signal from the unit 10, and either the linear time warp matching method or the DP method is adopted depending on the subcent to be recognized. There is a special problem in deciding whether to do so. Here, the subset refers to each subset divided by type, for example, as shown in FIG. Subset 1
stores words related to Japanese place names, and has a relatively large number of words, but Subset 2 is divided into 26 letters of the alphabet, and Subset 3 is divided into 10 letters of numbers 0 to 9. By specifying subsets, it is possible to reduce the number of matching targets. In a recognition device that can specify subsets in this way, a predetermined subset of the set of subsets to be recognized is matched using a simple linear time warping matching method that requires less matching processing, and in the case of other subsets, D, which requires a large amount of matching processing but has good recognition performance
I use the P method to check. That is, the term flowchart for the operation of the speech recognition device according to the present invention is +-L;
It is shown in FIG.

When the operation starts, first specify Liebsen 1.

For example, when recognizing the symbol string A 0038 depicted at the bottom of FIG. 6, subset 2 is selected because the first A relates to the alphabet set of subcent 2. Subset 3 is selected because the next four words are words related to numbers. In this way, the set amount of matching targets is reduced. After specifying the subset, input the audio. For example, a voice speaker inputs voice by reading out each word in sections such as A, O, 0, 3, and 8. After voice input, the number of words to be recognized is input from the host computer 5. That is, it is known in advance that the first word is a speech related to alpha vent of subset 2. Based on the subcent to be recognized, the present invention uses the table shown in FIG. 1 to determine whether to perform DP matching or linear matching. For example, the alphabet is subset 2, and DP matching is selected according to table reference as shown in FIG. DP
As a result of the comparison, a recognition result is output, and the host computer 5 executes recognition processing based on the recognition result. If recognition has been completed, the computer will stop, but if recognition has not been completed, it will be determined whether or not to change the subcent.

In the example in Figure 6, the letter A is followed by four numbers, so change the specification from subset 2 to subcent 3, input the voice again, and perform DP matching or linear matching using the subset in the same way. We are trying to determine and recognize what is happening. If a word related to a place name comes after a number, subset 1 will be selected, but in this case as well, the table will be referred to, and if the table in FIG. 8 is used, linear matching will be performed.

As described above, in the present invention, a set of subsets is created for all words depending on the type of word, and a table is used to indicate whether to use the DP method or the linear time expansion/contraction matching method for each subset. That is, the first
Inside the table 11 shown in the configuration diagram of the figure
Logical 1 indicating that the number and DP method are used, and logical 0 indicating that the linear time warp matching method is used,
As shown in the table shown in the figure, each subset is made to correspond. By referring to this table, either DP matching or linear matching is performed. In this way, a recognition device that can specify subsets can handle situations where the number of words to be recognized is large or small, situations where incorrect recognition would be fatal,
The correspondence between subsets and method selections is determined in advance, as shown in the table shown in FIG. 8, in accordance with situations in which erroneous recognition does not pose much of a problem or situations in which recovery is easy. This makes it possible to cope with both cases where the requirements for response time are severe and those where the requirements for recognition performance are severe.

〔Effect of the invention〕

In this way, the present invention stores the settings for each subcent in the storage circuit as a table, whether to select the linear time warping matching method with a small amount of matching processing, or to use the DP method, which requires a large amount of matching processing but has good recognition performance. However, by referring to this table when specifying a subset, a method can be selected, and there is an effect that it is possible to cope with cases where requirements for response time are severe and cases where requirements for recognition performance are severe.

[Brief explanation of drawings]

Figure 1 is a block diagram of the (9) word speech recognition device of the present invention;
Figure 3 is a block diagram of the parameter calculation section of the m-word speech recognition device of the present invention, Figure 3 is a standard pattern stored in the dictionary section of the word speech recognition device of the present invention, and Figure 4 is the word speech recognition device of the present invention. FIG. 5 is an example diagram of the non-linear expansion/contraction matching method (DP method) of the word speech recognition device of the present invention, and FIG. 6 is an example of the operation of the word speech recognition device of the present invention. FIG. 7 is a flowchart showing the flow of the operation of the word speech recognition device of the present invention. FIG. 8 is an embodiment of the reference table of the word speech recognition device of the present invention. DESCRIPTION OF SYMBOLS 1... Voice human power section, 2... Preprocessing section, 3... Parameter calculation section, 4... Section detection section, 5... Host computer, 6... Switching section, 1 7... - Dictionary section, 8... DP matching section, 9... Linear matching section, 11... Table. Frame Fig. 3 Fig. 4 (0) Fig. 4 (b) Fig. 5 (a) Fig. 5 (b) Suffix 1, P) B, C, D, ......Suffix set 3 (Seiro) (・chi, ni, san, --...Suffix set 4 Fig. 6 Fig. 8 Fig. 7)

Claims (3)

[Claims] (1) an acoustic analysis section that inputs an audio signal, extracts speech features, and performs section detection; a dictionary section that stores word standard patterns analyzed in advance through the acoustic analysis section; and the acoustic analysis section a first matching unit with a low processing amount for matching the feature pattern of the audio signal outputted through the word standard pattern of the dictionary unit; a second matching section that performs a large processing amount of matching against a standard pattern; and selectively transferring the output of the acoustic analysis section to the dictionary section, the first matching section, or the second matching section. and a control means for controlling the acoustic analysis section, the dictionary section, the first and second matching sections, and the selection section via a calculation processing section, for each subset, selecting the first and second matching sections. 1. A word speech recognition device comprising means for setting in advance in a table which of the method using the matching unit and the method using the second matching unit having better recognition performance is to be used. (2) The word speech recognition device according to claim 1, wherein the first matching unit performs linear matching. (3) The word speech recognition device according to claim 1, wherein the second matching unit performs nonlinear matching.