JPS6332399B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voice recognition system that is small and inexpensive.

A conventional speech recognition system is shown in FIG. 1st
In the figure, 1-1, 1-2, ..., 1-m are input signal lines for input audio signals of the speech recognition system, and 10
20-1,..., 20n are recognition unit input signal lines, and 30-1,..., 30-n are recognition units (for example, feature extraction unit 30a, similarity calculation unit 30b, maximum similarity (consisting of a determination unit 30c), 4
0-1, . . . , 40-n are configured like standard pattern memories storing characteristic parameters of registered words.

In general, voice recognition systems such as telephone answering systems need to process many line inputs simultaneously, but the inputs do not always occur at the same time. Therefore, the recognition unit selection circuit 10 selects and connects a recognition unit that is not in operation at the time an input request occurs. For example, when an input audio signal is generated on the input signal line 1-m and the recognition unit 30-1 is vacant, the input signal line 1-m is selected by the recognition unit selection circuit 10 via the recognition unit input signal line 20-1. recognition unit 3
Connected to 0-1. Therefore, a configuration can be achieved in which the ratio of the number of input signal lines to the number of recognition sections is m>n, and the number of recognition sections can be reduced. However, standard pattern memories 40-1, .
Then, the input audio signal is analyzed to extract feature parameters, and the similarity calculation unit 30b calculates the similarity between the feature parameters and the feature parameters for each registered word stored in the standard pattern memory. The maximum similarity determining unit determines the registered word with the highest degree of similarity, and outputs the code corresponding to the registered word as the determination result.

In this way, a standard pattern memory is connected to each recognition unit, and the standard pattern memory is output in synchronization with a standard pattern memory reference request from the recognition unit, which increases the size of the recognition system. The drawback was that it became expensive.

The purpose of the invention is to eliminate these drawbacks,
A standard pattern memory is commonly used by each recognition unit, and will be explained in detail below.

Figure 2 shows the first example of the speech recognition system according to the present invention.
An example of this is shown. In Figure 2,
1-1, 1-2,..., 1-m are input signal lines, 10
are recognition unit selection circuits, 20-1,..., 20-i...,
20-n is the recognition unit input signal line, 30-1,...,3
0-i..., 30-n are recognition units. These recognition units include a feature extraction unit 30a and a similarity calculation unit 30, respectively.
b, a maximum similarity determination unit 30c. 5
0 is a standard pattern memory that stores feature parameters of registered words, and each recognition unit 30-1,...,30
-i..., 30-n.

FIG. 3 is a time chart showing the operation timing of the voice recognition system shown in FIG. 2. Hereinafter, the operation of the first embodiment will be explained based on FIGS. 2 and 3.

Since the purpose of the recognition unit is to calculate the degree of similarity between the input speech signal and the features of the registered words, the degree of similarity may be calculated in any order for each registered word in the standard pattern memory.

Therefore, first, input signal line and recognition section selection circuit 1
0, the input speech signal is input to the recognition unit 30-i, the feature extraction unit 30a extracts the feature parameters of the input speech, and then the similarity calculation unit 30b
In order to calculate the degree of similarity between the feature parameters of each registered word in the standard pattern memory 50,
The characteristic parameters of each registered word in the standard pattern memory 50 are referred to. For example, if the number of registered words in the standard pattern memory 50 is five, the code number and feature parameter of each registered word are always output from the standard pattern memory 50 in sequence.

FIG. 4 shows an example of the configuration of the standard pattern memory 50 used in the speech recognition system of the present invention shown in FIG. is stored,
Moreover, "1" is set in the first bit as a code number FLAG indicating that this is the code number storage address. The characteristic parameters of the registered word are sequentially stored in each of the remaining addresses in the area that follows. The first bit of each address where this feature parameter is stored is set to "0". The characteristic parameters of the registered words stored in the standard pattern memory 50 include, for example, frequency-analyzing the voice to be registered using a plurality of frequency-compatible hand-pass filters, and sampling the outputs of the group of band-pass filters at a standardization cycle of about 10 ms. Use the converted version.

As described above, the code number and characteristic parameters of each registered word are sequentially output from the standard pattern memory 50 to each recognition unit. Now recognition part 3
There is an input audio signal in 0-i, and when its feature extraction is completed, a reference request is made to the standard pattern memory, but the recognition unit 30-i side reads the code number storage address of the first address of each registered word in the standard pattern memory. When the first bit of the word is monitored and the code number FLAG is detected, the code number of the registered word and the following feature parameters are taken in. Now, when a reference request for a registered word is generated from the recognition unit 30-i while the standard pattern memory 50 is outputting the feature parameters of the registered word S ₁ , monitoring of the code number FLAG is started, but the next registered word S ₂ is output. Upon detection of the code number FLAG, the code number is first fetched, and the subsequent feature parameters are sequentially fetched, and the similarity calculation unit calculates the degree of similarity with the input voice. As an example of similarity calculation, the feature parameters of input speech are
(k, l), and the feature parameter of the α-th registered word of the standard pattern is S〓(k, l). However, k is the number assigned to the bandpass filter output, l is the number assigned to the sampled data, and if the similarity is R〓, then R〓= _l 〓 _K 〓 | (k, l) −S〓(k,l)|

After the similarity with the registered word S ₂ has been calculated in this manner, the similarity with each registered word is subsequently calculated in the order of the registered word S ₃ →S ₄ →S ₅ →S ₁ .

In addition, in order to determine whether the similarity calculation with all registered words in the standard pattern memory has been completed, for example, the code number of the registered word whose similarity is to be calculated first is sent to a code number matching means (not shown). Store it and check it every time the code number is input.
When the output corresponding to each registered word from the standard pattern memory goes through a cycle and the same code number as the first code number is input again, the code number matches with the code number matching means and the similarity with all registered words is calculated. Make sure that it is determined that the calculation is completed.

Similarly, from the standard pattern memory, the registered word S ₃
While outputting the feature parameters, other recognition units 30-j
When a standard pattern memory reference request occurs in (not shown), the recognition unit 30-j reads the input voice in the order of each registered word S ₅ → S ₁ → S ₂ → S ₃ , starting with the next registered word S ₄ . and the registered word.

In this way, when the similarity calculation unit 30b finishes calculating the similarity between the input speech and all registered words in the standard pattern memory, the maximum similarity determination unit 30c next determines which registered word it is most similar to. The code number of the registered word with the highest degree of similarity is output as the determination result.

As explained above, in the first embodiment, the feature parameters of each registered word are always sequentially output from the standard pattern memory 50, and the feature parameters of all registered words are synchronized on the recognition unit side with the input voice. Since the similarity calculation is performed, there is an advantage that the standard pattern memory can be used in common.

In the first embodiment, a case has been described in which one standard pattern memory 50 is connected to a plurality of recognition units and used in common, but standard pattern memories divided into multiple groups are connected to a plurality of recognition units. It can also be used in common.

FIG. 5 is a diagram showing a second embodiment of the present invention. In Figure 5, 1-1, 1-2,..., 1
-m is the input signal line, 10 is the recognition section selection circuit, 20
-1, . . . , 20-2, . . . , 20-n are recognition unit input signal lines, and 30-1, 30-2, . 60-1, 60-2, 60-3 are standard pattern memories grouped according to the content of registered words, 70-1, 70-2,..., 70-n
is a multiplexer, 80 is a recognition control unit, 90-
1, 90-2, . . . , 90-n are determination result output lines of each recognition unit, and 91-1, 91-2, . . . , 91-n are multiplexer switching signal lines.

In general, in a recognition device, the fewer the number of recognized words, the higher the recognition performance. Therefore, in question-and-answer format recognition systems, the range of answers can be limited depending on the question, so the words to be recognized are divided into several groups depending on the type of question, and it is possible to divide the words to be recognized into several groups depending on the type of question. However, it can actually take the form of limiting the number of words to be recognized at one time. This will be explained below based on FIG.

Each standard pattern memory 60-1, 60-2, 6
The code numbers and characteristic parameters of each registered word stored in each standard pattern memory, which are sequentially output from 0-3, are output from each multiplexer 70-1,...,70.
- is entered in n. When the determination result of the recognition unit for a certain input voice is input to the recognition control unit 80, the recognition control unit 80 transmits the recognized result to the system control unit (not shown), and then the system control unit , and issues a command to a voice response unit (not shown) to send a response to the input side according to the recognition result. Next, before the recognition unit receives the next answer input from the input side, the system control unit connects the standard pattern memory storing the registered word corresponding to the content of the next answer input to the recognition unit via the recognition control unit. In order to do this, the multiplexer is switched, a predetermined standard pattern memory is connected to the recognition unit, and similarity with all registered words in the standard pattern memory is calculated in the same manner as in the first embodiment. With such a configuration, the standard pattern memory can be used efficiently and recognition accuracy can also be improved.

In the first and second embodiments described above, the memory structure of the standard pattern memory is such that the code number of each registered word is stored at the first address of the registered word, and the characteristic parameters are stored at the following address. However, the memory structure is not limited to this particular format, and any method may be used as long as the storage location of the characteristic parameters of each registered word can be determined.

Further, although the feature parameters stored in the standard pattern memory are frequency-analyzed parameters, it is clear that they may be autocorrelation parameters, general phoneme feature parameters, or the like.

Furthermore, in this embodiment, feature parameters are always output from the standard pattern memory, but when a reference request occurs and all feature parameters have been referenced, if no recognition unit references the standard pattern memory, It is clear that the configuration may be such that the output of the standard pattern memory is stopped.

In the present invention, the recognition unit executes the recognition operation in synchronization with the output of the standard pattern memory, so the standard pattern memory can have a minimum configuration.
It can be used in various recognition systems such as voice recognition systems.

[Brief explanation of the drawing]

Figure 1 is a configuration diagram of a conventional speech recognition system;
The figure is a block diagram of the first embodiment of the speech recognition system of the present invention, Figure 3 is a time chart explaining the embodiment of the present invention, and Figure 4 is an example of the memory structure of the standard pattern memory used in the present invention. The figure shown in FIG. 5 is a block diagram of a second embodiment of the speech recognition system of the present invention. 1-1, 1-2, ..., 1-m...input signal, 1
0... Recognition unit selection circuit, 20-1,...,, 20-
n...Recognition unit input signal, 30-1,...,30-n
... Recognition unit, 40-1, ..., 40-n... Standard pattern memory, 50... Standard pattern memory, 60-
1,60-2,60-3,...standard memory divided into groups, 70-1,70-2,...,70-
n...Multiplexer, 80...Recognition control unit, 90-
1, 90-2,..., 90-n... Judgment result output line, 91-1, 91-2,..., 91-n... Multiplexer switching signal line.

Claims (1)

[Scope of Claims] 1. A plurality of audio signal input systems, a plurality of audio recognition units, and the audio recognition unit that is in an inactive state is selected in response to an input request to each of the audio signal input systems, and a recognition unit selection circuit that interconnects the voice signal input system and the selected voice recognition unit; and a recognition unit selection circuit that stores characteristic parameters of each of a plurality of registered words, and a a commonly used standard pattern memory, the feature parameters of each registered word are commonly output from the standard pattern memory to each of the speech recognition sections, and the speech recognition section is synchronized with the output of the standard pattern memory. A voice recognition system characterized by performing a recognition operation of a voice input signal inputted from the voice signal input system. 2. The speech recognition system according to claim 1, wherein the standard pattern memory is divided into a plurality of groups according to the contents of the registered words.