JPH02118699A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To prevent the lowering of a recognition rate even when a background noise is fluctuated by inputting the contents of a template memory to store a first feature parameter and a second feature parameter and calculating the similarity of the voice. CONSTITUTION:In a first spectral memory 2, the time serial spectrum of a registered voice to be registered when the background noise is sufficiently small, and in a second spectral memory 3, the time serial spectrum of the noise just before a recognition subject voice input time is stored. When the voice and background noise are statistically independent, an additionality is realized between both in the dimension of the spectrum and a simultaneously, the background noise is fluctuated, by adding the two time serial spectral by means of an adding part 4, the real time serial spectrum is limited. By using the time serial spectrum imitated in such as way as the template, the background noise conditions at the time of preparing the template can be approximated. Thus, the wide lowering of the recognition rate can be prevented.

Description

DETAILED DESCRIPTION OF THE INVENTION FIELD OF INDUSTRIAL APPLICATION The present invention relates to a registration type speech recognition device in which the voice of a speaker is registered in advance.

BACKGROUND OF THE INVENTION In recent years, registered speech recognition devices have been increasingly used in industry.

A conventional registration type speech recognition device will be described below with reference to the drawings. FIG. 2 is a block diagram showing a conventional registration type speech recognition device. In FIG. 2, the output of the spectrum calculation section 1' is connected to the analysis section 2', and the output of the analysis section 2' is connected to the template memory 3' and the collation section 4' via a switch 5'. Further, the output of the template memory 3'' is connected to a matching section 4'.

Regarding the registration type speech recognition device configured as shown below,
The operation will be explained below.

The spectrum calculation section 1' calculates a time-series spectrum from the voice uttered by the user, and inputs the result to the analysis section 2'. The analysis section 2' receives this input, performs a predetermined analysis, and outputs the obtained characteristic parameters to the switch 5'. The switch 5' inputs the characteristic parameters as the first output to the template memory 3' if the uttered voice is the one at the time of registration, and -1- records if the uttered voice is the one at the time of recognition. The feature parameters are output as second outputs to the matching section 4'. During recognition, the matching unit 4' receives the output of the template memory 3' and the second output of the switch 5', performs template matching, and determines the optimal recognition result from among a plurality of registered word templates.

Problems to be Solved by the Invention In template matching type speech recognition, it is desirable that the background noise conditions at the time of template 1 registration and the background noise conditions at the time of inputting the speech to be recognized are the same. However, in the configuration described above, if the background noise is fluctuating, the background noise conditions at the time of template registration and the background noise conditions at the time of inputting the recognition target speech are no longer the same, and There was a problem in that the recognition rate was significantly low.

In view of the above problems of conventional speech recognition devices, the present invention ensures a good recognition rate even when the background noise is fluctuating, as long as it can be assumed that the background noise is sufficiently small even during speech registration. The purpose of the present invention is to provide a speech recognition device that can perform the following tasks.

Means for Solving the Problems The registration type speech recognition device of the present invention includes a spectrum calculation unit that calculates a time-series spectrum from an input signal, and a first storage unit that stores the time-series spectrum of registered speech obtained from the spectrum calculation unit. a spectrum memory; a second spectrum memory that stores the time-series spectrum of noise obtained from the spectrum calculation unit; and an addition unit that adds the contents of the first spectrum memory and the second spectrum memory. and extracting the first feature parameter by analyzing the time-series spectrum addition result output from the adding section, or analyzing the time-series spectrum of the recognition target speech obtained from the subekle calculation section. a template memory that stores the first feature parameter; and a speech analyzer that inputs the contents of the template memory and the second feature parameter to calculate the similarity of speech. It consists of a collation section.

In the present invention, with the above-described configuration, the first spectral memory stores a time-series spectrum of registered speech registered when the background noise is sufficiently small, and the second spectral memory stores a recognition target. The time-series spectrum of noise immediately before the speech input point is memorized, and the speech and background noise are statistically independent, additivity exists between them in the dimension of the spectrum, and the background noise is the length of the word to be recognized. to ratio to C
If the uplink is fluctuating with a large time constant, the true time-series spectrum at the time of inputting the speech to be recognized is simulated by adding these two time-series spectra, and the time-series spectrum simulated in this way is used as a template. By using this as a template, it is possible to approximate the background noise conditions at the time of creating the template and the background noise conditions at the time of inputting the speech to be recognized, and it is possible to prevent a significant drop in the recognition rate.

Examples Examples of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a registration type speech recognition device in one embodiment of the present invention. In FIG. 1, l is a spectrum calculation unit that calculates a time-series spectrum from an input signal, 2 is a first spectrum memory, 3 is a second spectrum memory, and the output of spectrum calculation unit 1 is is input to the spectrum memory 2, the second spectrum memory 3, and the first input terminal of the switch 8. 1st
The output of the spectrum memory 2 and the output of the second spectrum memory 3 are input to the adder 4, and the time-series spectrum addition result output from the adder 4 is input to the second input terminal of the switch 8. . The output of switch 8 is sent to analysis section 5
The output of the analysis section 5 is input to the switch 90. Further, the first output of the switch 9 is input to the template memory 6, and the output of the template memory 6 and the second output of the switch 9 are both input to the matching section 7.

The operation of the registration type speech recognition device configured as described in 1- above will be explained below.

The registered speech uttered when the background noise is sufficiently small is converted into a time series spectrum by the spectrum calculation unit 1, and the first
is stored in the spectrum memory 2 of. During voice recognition,
The spectrum calculation unit l recalculates the time-series spectrum of the background noise every time the statistical properties of the background noise change, and updates the second spectrum memory 3. The adder 4 re-adds the output of the first spectrum memory 2 and the output of the second spectrum memory 3 each time the update is performed, and generates a time-series spectrum simulating the case where background noise is superimposed on the registered voice. (This time series spectrum will be referred to as a simulated registered time series spectrum from now on) and this signal is used as the second input to the switch 8. The switch 8 always operates to output the second input except when the recognition target voice is input, and the analysis unit 5 always outputs the second input except when the recognition target voice is input due to the function of the switch 8. Analyze time series spectra. The switch 9 always selects the first output terminal except when the speech to be recognized is input.
The updated simulated registration time-series spectrum is always inputted to the input device 1, except when the speech to be recognized is inputted.

On the other hand, when the speech to be recognized is input, the switch 8 selects the first input terminal r, and the switch 9 selects the second output terminal, and by this operation, the time-series spectrum of the speech to be recognized is sent to the matching section 7. is input. The collation unit 7 receives the second output of the switch 9 and the output of the template memory 6, and compares the characteristic parameters analyzed from the time-series spectra of the speech to be recognized and the plurality of simulated registered time-series spectra. Performs template matching between multiple feature parameters and outputs the optimal recognition result.

Note that the effect of the present invention remains the same as long as additivity holds true for the spectrum, but in this example, the Fourier
Auto power spectrum shall be used. Also, various methods can be considered for the processing content in the analysis section 5, but in this embodiment, the most -
・Perform logarithmization of spectral power. In the case of HH<joint 7, a DP matching method using the Thiebishev distance is used.

Tokorotte 5teven Bolt
-L) by Mr. ans'a'c't, io"ns ;=
;/rc'o'u"st!'cs, shore e5h,
'S'T20, April 1979) is a well-known noise removal method.
As long as linearity is maintained in the spectral processing process, Bolt's spectral subtraction method and the spectral addition method of the present invention are equivalent. On the other hand, if nonlinearity exists in the spectrum processing process, a difference will occur between the two. In general, speech recognition devices often employ analysis methods that take auditory characteristics into consideration, and in this embodiment, logarithmization processing is also performed in the analysis section 5. This is not equivalent to processing based on the spectral addition method. When the spectral subtraction method is used, the subtraction result may be negative when the estimated noise is subtracted from the noise-containing input audio signal, and a problem occurs when the result is logarithmized. In the above-mentioned literature, as a countermeasure against such a case, the spectrum in the negative frequency domain is forcibly set to zero, but this is just a convenient method, and the spectrum after processing is distorted. On the other hand, in the spectral addition method of the present invention, the problem of logarithmization of negative numbers as in the case of L does not occur, and a spectrum without distortion can be used for subsequent processing.

As described above, this embodiment includes a spectrum calculation section 1 that calculates a time-series spectrum from an input signal, and a first spectrum memory 2 that stores the time-series spectrum of registered speech obtained from the spectrum calculation section 1. and a second section that stores the time-series spectrum of noise obtained from the spectrum calculation section 1.
a spectrum memory 3, an adder 4 that adds the contents of the first spectrum memory 2 and the contents of the second spectrum memory 3, and the output of the spectrum calculator l as the first input, and the output of the adder 4 as the first input. A switch 8 which takes the second input, an analysis section 5 which analyzes the output of the switch 8 and extracts feature parameters, a switch 9 which takes the output of the analysis section 5 as an input and has first and second outputs, and a switch A template memory 6 receives as an input the simulated registered time-series spectrum which is the first output of the switch 9, and a time-series spectrum of the speech to be recognized which is the second output of the switch 9 and the output of the template memory 6. By providing the matching section 7, the first spectrum memory 2 stores the time-series spectrum of the registered speech registered when the background noise is sufficiently small, and the second spectrum memory 3 stores the time-series spectrum of the registered speech registered when the background noise is sufficiently small, and the second spectrum memory 3 stores the time-series spectrum of the registered speech registered when the background noise is sufficiently small. The time-series spectrum of the noise immediately before is memorized, and the speech and background noise are statistically independent, additivity exists between them in the spectral dimension, and the background noise is proportional to the length of the word to be recognized. +1 if the fluctuation is with a larger time constant than
− The true time-series spectrum at the time of inputting the speech to be recognized is simulated by adding the two time-series spectra, and by using the time-series spectrum simulated in this way as a template, background noise at the time of template creation is It is possible to approximate the conditions and the background noise conditions at the time of inputting the speech to be recognized, and it is possible to prevent a significant drop in the recognition rate.

In addition, since the information on the registered speech and background noise is held in the dimension of spectrum rather than the time waveform, in order for the analysis section 5 to perform a predetermined analysis, the first spectrum memory 2 and the second spectrum memory are used. You only need a small amount of all three. To illustrate the average analysis conditions of sampling the signal at 10 kHz and performing 20th order spectrum analysis every 256 samples, 256 words of memory are required per point in time to hold the registered speech or background noise in the time waveform. However, in the case of spectrum, it only takes 20 edicts.

Furthermore, the amount of calculation in the spectrum calculation section 1 is calculated by the analysis section 5.
Usually, the amount of calculation is large compared to the amount of calculation in
If registered voices are held in the form of time waveforms, it becomes necessary to recalculate time-series spectra for each of a plurality of registered voices each time the statistical properties of the background noise change, which reduces the operating speed of the entire apparatus.

On the other hand, if the time-series spectrum is retained, the only thing that needs to be recalculated every time the statistical properties of the background noise change is the time-series spectrum of the background noise at that point, which reduces the operating speed of the entire device. It never declines.

Effects of the Invention As described in 2., the present invention comprises: a spectrum calculation unit that calculates a time-series spectrum from an input voice; a first spectrum memory that stores the time-series spectrum of the registered voice obtained from the spectrum calculation unit; a second spectrum memory that stores a time-series spectrum of noise obtained from the spectrum calculation section; an addition section that adds the contents of the first spectrum memory and the second spectrum memory; and the addition section that adds the contents of the first spectrum memory and the second spectrum memory. Analyzing the time-series spectrum addition result output from the spectrum calculation unit to extract the first feature parameter, or analyzing the time-series spectrum of the recognition target speech obtained from the spectrum calculation unit to extract the second feature parameter. an analysis unit for extracting, a template memory for storing the first feature parameter, and a template memory for storing the content of the template memory and the second feature parameter;
Since it is equipped with a speech matching unit that calculates the similarity of speech by inputting the feature parameters of Not only can it approximate the noise condition and prevent a significant drop in the recognition rate, but it can also retain information on the registered speech and background noise in the form of a spectrum, thereby retaining both in the form of a temporal waveform. It is possible to provide an excellent registration type speech recognition device that can achieve the desired purpose with a smaller memory capacity and a smaller amount of calculation compared to the previous example.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a registered speech recognition device according to an embodiment of the present invention, and FIG. 2 is a block diagram of a conventional registered speech recognition device. DESCRIPTION OF SYMBOLS 1... Spectrum calculation unit, 2... First spectrum memory, 3... Second spectrum memory,
4... Addition unit, 5... Analysis unit, 6... Template memory, 7... Collation unit, 8... Switch, 9...
・Switch, 1'...spectrum calculation section, 2'...
Analysis section, 3'...template memory, 4'...verification section, 5'...Mates general

Claims (1)

[Claims] a spectrum calculation unit that calculates a time-series spectrum from an input signal; a first spectrum memory that stores the time-series spectrum of registered speech obtained from the spectrum calculation unit;
a second spectrum memory that stores a time-series spectrum of noise obtained from the spectrum calculation section; an addition section that adds the contents of the first spectrum memory and the second spectrum memory; and the addition section that adds the contents of the first spectrum memory and the second spectrum memory. Analyzing the time-series spectrum addition result output from the spectrum calculation unit to extract the first feature parameter, or analyzing the time-series spectrum of the recognition target speech obtained from the spectrum calculation unit to extract the second feature parameter. Speech recognition consisting of an analysis section that extracts, a template memory that stores the first feature parameter, and a speech matching section that inputs the contents of the template memory and the second feature parameter and calculates the similarity of speech. Device.