JPS59211098A - 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] fa) Technical Field of the Invention The present invention relates to a method for recognizing speech by comparing an input feature parameter time series of input speech of a speaker with a plurality of registered feature parameter time series registered in advance. Relating to a recognition device.

(bl Technology background) In recent years, speech recognition devices have been used to convert the input speech of a limited number of registered speakers into an appropriate input feature parameter time series and register it in a registered speech dictionary. A speaker speech recognition device that identifies whether the input speech is from a specific speaker by comparing it with a registered speech dictionary converts the input speech of the specific speaker's words into a time series of appropriate input feature parameters. With the advancement of voice recognition technology, such as specific speaker voice recognition devices, which register the input voice uttered by a specific speaker to be recognized with the registered voice dictionary and display the recognition results as characters, This has expanded to include the field of interaction with machines.

As a speech recognition method, unknown input speech is frequency-analyzed, the analog signal resulting from the analysis is converted into a digital signal, the converted digital signal is converted into a time series, the speech section is determined by the dark value that determines the speech section, and each The input feature parameter time series representing the features of the phoneme is extracted as a time series, and the extracted input feature parameter time series is compared with a plurality of registered feature parameter time series registered in advance to determine the closest distance between the matching results. is configured to obtain recognition results for unknown input speech by selecting . Various methods have been studied because the format of the feature parameter time series, the registration method and matching selection method, etc. affect the difficulty and recognition rate of input speech recognition. ing.

(cl. Prior Art and Problems) Conventional speech recognition methods have a registered speech dictionary in which multiple time series of input feature parameters of input speech uttered by a specific speaker are registered, and the beginning and end of the input speech to be recognized are determined. The recognition result is output by comparing the input feature parameter time series of the voice section with the time series of a plurality of registered feature parameters in the registered speech dictionary. After cutting out a voice section, distance calculations are performed for each of a plurality of registered feature parameter time series registered in advance using all input feature parameter time series of each voice section.

The configuration of this type of speech recognition method will be explained.

FIG. 1 shows a block diagram of a circuit configuration of a conventional speech recognition device. The procedure for registering the speaker's voice in advance is to input one input voice uttered by the speaker from the microphone 1, and filter the input voice through a group of band filters 2 in the voice band 200.
About Hz to 5KHz is divided into a group of 10 to 20 (7) channel filters, the output of each channel filter is extracted at 5 to 30m5 cycles, and the feature parameter time series extractor 3 converts it into an input feature parameter time series of digital information. , are stored in the input feature parameter time series buffer 4. The input feature parameter time series stored in the input feature parameter time series buffer 4 is processed by a speech segment cutting circuit 5 to cut out speech segments using dark values that determine the speech segments at the beginning and end of a word, and register them to represent the characteristics of the phonemes of the input speech. It is registered in the registered feature parameter time series dictionary section 6 as a feature parameter time series. The above procedure is repeated for the number of input voices to be registered, and a plurality of registered feature parameter time series are registered in the registered feature parameter time series dictionary section 6.

Next, the procedure for recognizing the speaker's voice is to input the input voice uttered by the speaker from the microphone 1, and store the input voice in the input feature parameter time series buffer 4 according to the above-mentioned procedure. . The input feature parameter time series stored in the input feature parameter time series buffer 4 is segmented into a plurality of speech segments and cut out by a speech segment extraction circuit 5 using a threshold value that determines the speech segment. The voice section extraction circuit 5 may be of a type in which only the darkness value is changed, operating according to the same algorithm, or may be of a type combining circuits with different algorithms themselves.

FIG. 4 shows an example in which a plurality of voice segments are determined by the dark value that determines voice segments in the voice segment cutout circuit 5.
Let the three-level thresholds be TL, TM, and TH, and when - times of voice, for example, "Aomori" is input, the darkness value will cause 1 "Aomori" at the dark value TL level and "1 Aomori" at the threshold TM level.
The three voice sections of "Omo" and "Omo" are determined by the dark value TH level. Therefore, in FIG. 1, when a dark value of n(llil) is set, n input feature parameter time series for the speech section are output, and a plurality of registered feature parameter time series dictionary sections 6 registered in advance are output. The matching selection circuit 7 sequentially matches the m registered feature parameter time series and the input feature parameter time series n for each voice section nxm times, calculates the matching distance nXm times, and calculates the matching distance nXm times. Among them, the closest matching distance is selected and outputted to the recognition terminal 8 as the recognition result.

The above is the conventional speech recognition processing procedure. In this method, the amount of calculation for the matching distance between the input feature parameter time series of the speech interval n1lffi and the m registered feature parameter time series is nxm times, and the amount of calculation increases as the number of recognition categories increases. This method has disadvantages in that the verification process requires a large amount of time and the rate of misrecognition is high.

(dl OBJECT OF THE INVENTION The present invention aims to solve the above-mentioned conventional drawbacks.

(el) Structure of the Invention The above object is to recognize an unknown input voice by comparing a human feature parameter time series extracted from the unknown input voice with a plurality of registered feature parameter time series registered in advance. In the recognition device, a speech segment cutting circuit that cuts out a plurality of speech segments from the input feature parameter time series, a first selection circuit that selects the speech segment cut out from the speech segment cutting circuit, and the first selection. a second selection circuit that selects a recognition result using the input feature variation time series of the speech section selected by the circuit; The input feature parameter time series of the voice section is converted into a roughly sampled phase input feature parameter time series, and the first selection circuit selects the plurality of registered feature parameters so as to correspond to the phase input feature parameter time series. The closest speech section is selected by comparing with the time series, and the second selection circuit compares and selects the input feature parameter time series of the selected speech section and the plurality of registered feature parameter time series. This is achieved by the present invention, which is configured to recognize this.

By providing a plurality of speech segment extraction circuits and a two-stage selection circuit according to the present invention, it is possible to reduce the amount of matching distance calculation compared to the conventional method, and at the same time, it is possible to reduce the amount of matching distance calculation compared to the conventional method. series, 2
It recognizes various variations such as the strength, duration, and accent of the input voice, and can achieve a high recognition rate even if the voice becomes hoarse due to the speaker's physical condition that day. There are benefits that can be recognized.

ffl Embodiment of the Invention An embodiment of the present invention will be described below. FIG. 2 is a circuit configuration block diagram of a speech recognition device according to the present invention, and throughout the invention, the same objects are designated by the same symbols as in FIG. 9 is a first selection circuit, 10 is a registered feature parameter time series dictionary section, 11 is an AND gate circuit, and 12 is a second selection circuit.

In this circuit configuration, the voice registration processing procedure is the same as the conventional method shown in FIG. 1, so a description thereof will be omitted.

The speech recognition processing procedure according to the present invention is the same as that shown in FIG. 1 up to the point where the input speech of the speaker to be recognized is inputted from the microphone 1 and the input feature parameter time series is stored in the input feature parameter time series buffer 4.

This input feature parameter time series is cut out as a human feature parameter time series of n speech sections by setting n dark values in the speech section extraction circuit 5, and then in the first selection circuit 9, the n human feature parameter time series is cut out as a human feature parameter time series of n speech sections. sequence, one channel out of one channel of bandpass filter group 2 (however, j
<i) into a time series of n phase input feature parameters, and preregistered feature parameter time series dictionary unit 1
One of the registered feature parameter time series of m(lliI registered in One voice section is selected from n. Only one voice section selected from among these n is input to the a side of the AND gate circuit 11. At the same time, the voice section cut out with the same darkness value is input to the a side of the AND gate circuit 11. The input feature parameter time series of all i-channels is manually input to the b side of the AND gate circuit 11 of the same system.Only the systems input to both the a and b sides of this AND gate circuit 11 are input from the a side of the AND gate circuit 11. The input feature parameter make time series is output.Next, the other one registered feature parameter time series among m is compared with the phase input feature parameter time series of n speech sections, and the speech section is determined in the same manner as described above. is selected, and the input feature parameter time series of the selected voice section is output to the a side of the ant gate circuit 11. However, the input feature parameter time series is thinned out in the time direction from the input feature parameter time series 0 as the phase input feature parameter time series. It is also possible to use a parameter compressed version with a fixed number of bytes, etc. In this way, the phase input feature parameter time series of n speech sections is compared with the m registered feature parameter time series m times. , a total of m speech sections are selected, and only the input feature parameter time series of the selected speech sections are sequentially output to the C side of the AND gate circuit 11 and input to the second selection circuit 12. The input feature parameter time series of the selected voice section and the m registered feature parameter time series registered in the registered feature parameter time series dictionary section 10 are
In order, the second selection circuit 12 calculates the matching distance m times, selects the closest matching distance, and outputs the registered feature parameter time series selected as the recognition result to the recognition terminal 8. In the method of the present invention, the amount of calculation of the matching distance calculated by the first selection circuit and the second selection circuit is (j/i) < (11/n) satisfying (n j/ i +l) m < (nxm). ), the amount of calculation is reduced by setting i and j such that j<i.

FIG. 3 shows another embodiment of the present invention, in which 13 is a first selection circuit, 14 is a selected input feature parameter time series buffer,
15 is a second selection circuit. The first selection circuit 13 compares the n phase input feature parameter time series described in FIG. , the selected input feature parameters are stored in the time series buffer 14. This selected and stored one input feature parameter time series and m(
The second selection circuit 15 calculates the matching distance r times for the registered feature parameter time series of IIil, selects the closest matching distance, and outputs it to the recognition &ff1i terminal 8 as a recognition result. The amount of calculation for the matching distance calculated by the first selection circuit and the second selection circuit is as follows: (j/i) By setting r which is i 1 , i and J, the amount of calculation is reduced.

Although the embodiments of the present invention have been described above using a frequency analysis method using a bandpass filter, the present invention can also be used in a speech recognition device employing an LPG analysis method or the like.

(gl As described in detail of the invention, by providing a plurality of speech segment extraction circuits and a two-stage selection circuit according to the present invention, it is possible to recognize the strength and weakness of the input speech in speech recognition that recognizes the input speech of the speaker. Since recognition can be performed in response to various variations such as time length and accent, it is possible to improve the recognition rate and reduce the amount of calculation for matching distance.

[Brief explanation of drawings]

FIG. 1 is a circuit configuration block diagram of a conventional speech recognition device.
FIGS. 2 and 3 are block diagrams of the circuit configuration of the speech recognition apparatus according to the present invention, and FIG. 4 is a relationship diagram for determining the speech section of input speech. In the drawing, 1 is a microphone, 2 is a group of bandpass filters 3 2 , 3 is a feature parameter time series extractor, 4 is an input feature parameter time series buffer, 5 is a speech section extraction circuit, 7 is a matching selection circuit, and 8 is a recognition giill. Terminal, 9 is the first selection circuit, 10 is the registered feature parameter time series dictionary section, 1
2 represents a second selection circuit, 13 represents a first selection circuit, 14 represents a selected input feature parameter time series buffer, and 15 represents a second selection circuit. 4

Claims (1)

[Claims] In a speech recognition device that recognizes unknown input speech by comparing an input feature parameter time series extracted from unknown input speech with a plurality of registered feature parameter time series registered in advance, the input feature parameter time series is a speech segment cutting circuit that cuts out a plurality of speech segments from a series; a first selection circuit that selects the speech segments cut out by the speech segment cutting circuit; A second selection circuit selects a recognition result using the input feature parameter time series, and the input feature parameter time series of the plurality of speech segments cut out by the speech segment extraction circuit is used to select the unknown input speech to be recognized. , convert it into a roughly sampled phase input feature parameter time series, and compare it with the plurality of registered feature parameter time series to correspond to the phase input feature parameter time series in the first selection circuit to select the closest voice. It is characterized in that a section is selected, and the input feature parameter time series of the selected speech section and the plurality of registered feature parameter time series are collated and selected by the second selection circuit for recognition. speech recognition device.