JPS6226038B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device, and particularly aims to improve the recognition rate of a word speech recognition device based on a pattern matching method.

A word speech recognition device based on the principle of pattern matching usually has a configuration including a speech input section 1, a feature extraction section 2, a recognition processing section 3, a registered pattern memory 4, and an input pattern memory 5, as shown in FIG. It has two operating modes: mode and recognition mode. In the registration mode, the word sounds to be recognized are registered in advance, and the features of the sounds extracted by the feature extraction unit 2 from the registered sound which is the output of the audio input unit 1 including the microphone are extracted as a time series pattern. The pattern is stored in a registered pattern memory 4, which is also called a standard pattern memory. Since the input audio is an analog quantity and is to be digitally processed, the feature extraction section 2 naturally has sampling and analog-to-digital conversion functions. In the recognition mode, after a voice feature pattern similarly extracted from the input voice is stored in the input pattern memory 5, the recognition processing unit calculates the degree of similarity between this input pattern and the registered pattern stored in the registered pattern memory 4. 3, the registered pattern with the highest degree of similarity is identified, that is, recognized as matching the input voice, and an appropriate output is performed accordingly. The configuration of the recognition processing section 3, registered pattern memory 4, and input pattern memory 5 shown in FIG. 1 having such functions is realized by a computer system centered on a CPU.

It is well known that physical quantities such as a frequency spectrum, an autocorrelation function, the number of zero crossings, and an α parameter are used as a method for extracting voice features. Among these methods, the method of extracting the frequency spectrum of the voice using a large number of band-pass filters has a relatively simple configuration and can obtain a high recognition rate, so it is often used in practical applications of this type of device.

FIG. 2 is a specific example of a speech recognition device that analyzes a frequency spectrum using a filter. The audio input section 1 is made up of a close-talk type microphone consisting of two directional and non-directional microphones 11 and 12 connected in opposite phases as shown in the figure, a microphone amplifier 13, and an AGC circuit 14. . Such a differential connection increases the signal-to-noise ratio of the control command voice within the directional characteristic range compared to the voice outside the directional characteristic range. The AGC circuit 14 keeps the output level substantially constant regardless of the magnitude of the input audio level, and has an attack time that is sufficiently faster than the rise of the audio signal and a recovery time that is appropriately slow. The feature extraction unit 2 includes a group of M bandpass filters 21-1, 21-2, .
M, and a group of low-pass filters 22-1, 22-2, . . . 22-M having a cutoff frequency of several tens of Hz, which are connected in series to each filter and detect each output envelope. With this configuration, the audio input section 1
Each filter component of the audio signal input from the input audio signal is sequentially sampled by the analog multiplexer 23 at appropriate time intervals (in most cases around 10 msec), converted to a digital code by the analog-to-digital converter 24, and then input to the I/O code of the microcomputer 5. The data is once taken into the buffer memory 27 via the /O port 26. The amount of data is, for example, the number of filters is 8, the audio input time is 1.6 seconds, the sampling time interval is 10 msec, and the analog-to-digital converter is 24.
When the number of bits in is 8, 1.6/0.01×8×8=10240 bits=1.28KB (B: byte). Since this data corrects the expansion/contraction of the time axis, only sampling data that equally divides the start and end of the voice into N pieces for each filter output is extracted and taken in as a feature pattern. For example, when N=32, N×8×8=2048 bits=256B is registered pattern memory 28 in registration mode.
In the recognition mode, the input pattern memory 29
are stored respectively. These memories are usually random access memories (RAM), and the address of the registered pattern memory 28 is specified by the ROM 30 that stores the CPU program and the input control unit 31.

The number of registered patterns is determined by the number m of speakers and the number r of words that each speaker can register. For example, when m=4 and r=8, a registered pattern memory of 4×8×256B=8.192KB is prepared.

In the recognition process, similarly, the audio signal is once input to the buffer memory 27 in the recognition mode, and from this, the contents of the input pattern memory 29 that stores the data of N sample points normalized on the time axis, and m×r registered patterns are input. This is done by pattern matching the contents of Various algorithms have been proposed for calculating the distance between the input pattern and the registered pattern, but since this invention has little to do with the subject matter, the explanation will be given using the simplest method, the Thiebishiev distance, as an example.

A time series of eight filters [fij ^(R) ] (i: filter number 1 to
8, j: sample points 1 to N) and the same filter time series [fij] of the input speech pattern, the Thievishev distance D is defined by the following equation.

In other words, this is the input pattern fij and the registered pattern
fij ^(R) , which is the sum of the absolute values of the differences between the corresponding data, and the input pattern is assumed to match the one that is the smallest among the Tievisiev distances obtained for a maximum of m × r registered patterns. It is something to be considered. For temporary storage of these calculation results,
The CPU 25 has a buffer memory 27, a registered pattern memory 28, an input pattern memory 29, and a recognition processing memory 32.

Now, as mentioned above, in a word speech recognition device using the pattern matching method, it is easy to imagine that the recognition rate will vary depending on the number of words registered in the registered pattern memory 28. If the number of registered patterns is small, it will be easy to identify the input pattern and a high recognition rate will be obtained; conversely, if the number of registered patterns is large, the distance between the correct registered pattern and other patterns will be smaller as a result of similarity calculation. This increases the possibility that separation will not be possible with the necessary distance and margin.

FIG. 3 shows actual measured values of the relationship between the number of registered words and the recognition rate for a word speech recognition device based on the above principle and having a registered word capacity of 34 words. This experiment was conducted by five male adults using meaningful word speech (1.6 seconds or less) consisting of several syllables, and the average recognition rate was shown. According to this result, the recognition rate is 100% when the number of registered words is 5 or less, and the maximum
The rate dropped to 95.8% for 34 words. The reason why the recognition rate decreases with the number of registered words is that the distance between the registered patterns becomes closer as the number of registered patterns increases, and the present invention solves this problem by adjusting the voice in the registration mode. When registering, the distance between the registered patterns that have already been registered is calculated in recognition mode as well, and this result is output and displayed in some form to maintain an appropriate distance between the registered patterns. It is something.

A specific embodiment of the present invention will be described below based on FIG. 4. This figure only shows the input/output control section of the recognition processing section, but the parts other than those shown, such as the voice input section, feature extraction section, etc., are omitted because they are the same as those in FIG. 2 of the conventional example. In this Figure 4, 25
is a CPU, and this CPU 25 has a buffer memory 27, a registered pattern memory 28,
Input pattern memory 29, recognition processing memory 3
2. A program memory 30 and an I/O port 26 are connected. The registered pattern memory 28 is provided with a memory portion for storing m×r feature patterns [fij ^(R) ] corresponding to the number m of speakers that can be registered and the number r of words that can be registered for each speaker. In the above example, the number of memory bits allocated to each memory portion is the product of the number of filters, the number of AD converter bits, and the number of sampling points (=N). I/
The O port 26 includes a data line with a feature extraction unit (not shown) or an output control unit 33 that performs control output based on recognition results, and an input control unit 31 that performs voice registration control and display related to registration recognition. It is connected.

The input control unit 31 includes a switch 34 for specifying registration and recognition modes, a switch group 35-1, 35-2...35-m for specifying m speakers, and a switch group 36 for specifying r registered words. -1,36-2...
36-r, and m+r speaker designation LED group 3
7. Registered word designation LED group 38, registration completion indicator 39 that indicates that registration is completed, when the level of the input voice is too low, the input time exceeds the specified time, or the corresponding registration as a result of recognition calculation As shown in the figure, a registration failure indicator 40 is provided to display when a pattern cannot be specified. In addition, this input control section is equipped with a power switch, a switch for erasing registered contents, etc., but these are omitted in the specific example shown in FIG.

Now, in the registration mode in which the mode changeover switch 34 is operated on the registration side, voice registration is performed by operating one of the m speaker designation switch groups 35 and selecting one of the corresponding speaker designation LED groups 37. While lit, r registered audio designation switches 36 are turned on.
(The corresponding registered word designation LED lights up) to input the registration control command voices in sequence, and the characteristic pattern of each registered voice is determined by the speaker designation switch 35 and registered voice designation switch 36. Addresses are sequentially stored in the registered pattern memory 28 associated with each address, but in the present invention, in the registration mode as well as in the recognition mode, the characteristic pattern of the registered voice and the previously registered registration pattern are stored. Calculate the similarity between That is, the characteristic pattern of the registered voice is temporarily stored in the input pattern memory 29 via the buffer register 27, and distance calculation, that is, similarity calculation, between the memory contents and the contents of the registered pattern memory 28 is performed. As a result, when the difference [D] between the characteristic pattern input for registration and the already registered pattern is greater than a certain specified value, the contents of the input pattern 29 are stored in the registered pattern memory specified by the input control unit 31. 28 and at the same time, the registration completion indicator 39 lights up. Similarly, if the difference [D] between the input pattern for registration and any of the already registered patterns is less than the specified value, the pattern is not transferred to the registered pattern memory and a message indicating that registration is not possible is displayed. The registration impossible indicator 40 is lit up. At the same time, input pattern memory 29
The speaker number and the registered word of the registered pattern similar to the content of, that is, [D] is below the specified value, are displayed on either of the respective LED groups 37, 38, and this already registered characteristic pattern is changed. It would be more practical to suggest that. In this case, registration can be performed again after re-inputting the registered voice corresponding to the registered pattern. Furthermore, even if the registered input pattern is similar to an already registered voice pattern, if this is the same word from another speaker and the corresponding recognition output is the same, then the unregistration indicator 40
The pattern is stored in the registered pattern memory 28, assuming that the registration has been completed successfully without the pattern being turned on.

As described above, in the conventional registration mode, voice registration was unilateral voice registration of the speaker without any feedback, but according to the present invention,
At the time of registration, the degree of similarity with already registered voices is calculated, and by doing so, it is possible to maintain the distance between the registered feature patterns of the registered voices at a certain level or higher, which reduces the recognition rate of this type of speech recognition device. can be improved.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing the concept of a speech recognition device, Fig. 2 is a block diagram showing a specific example of a speech recognition device, Fig. 3 is a relationship curve diagram between the number of registered words and recognition rate, and Fig. 4 is a diagram of the present invention. This is a block diagram showing the configuration of the main parts of the device, in which 1 is a voice input section, 2 is a feature extraction section, 3 is a recognition processing section, 4 is a registered pattern memory, and 5 is an input pattern memory.

Claims (1)

[Scope of Claims] 1. A microphone that converts audio into an electrical audio signal, an amplification circuit that amplifies the audio signal from the microphone, and characteristics of the audio signal amplified by the amplification circuit. an A/D conversion circuit that converts the feature pattern extracted by the extraction means into a digital signal; and an input pattern that stores the feature pattern of the voice digitized by the conversion circuit each time the voice is input. A storage means, a registered pattern storage means in which a plurality of voice characteristic patterns are stored, a similarity calculation means for calculating the similarity between the characteristic patterns of the input pattern storage means and the registered pattern storage means, and the similarity calculation means and a registration control means that selects characteristic patterns suitable for registration at the time of voice registration, and the registration control means operates the similarity calculation means. Then, the degree of similarity between the characteristic pattern of the input pattern storage means inputted for registration and the characteristic pattern of the registered pattern memory that has already been registered is determined, and when the degree of similarity is greater than a predetermined value, the characteristic pattern of the input pattern storage means is determined. The characteristic pattern of the input pattern storage means may be transferred and stored in the registered pattern memory, but when the degree of similarity is smaller than a predetermined value, the characteristic pattern of the input pattern storage means may be prohibited from being transferred and stored in the registered pattern memory, and the display means may display that registration is not possible. Featured voice recognition device.