JPS62195699A - Continuous numerical voice recognition - 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 (Field of Industrial Application) The present invention relates to a Ma digit speech recognition method in speech recognition.

(Prior Art) Continuous digit speech recognition methods have been developed and put to practical use in order to recognize and judge input speech of consecutive digits uttered continuously.
This type of recognition method disclosed in No. 95 uses a DP matching method especially for numbers among continuous speech recognition methods and limits the number of digits to recognize continuous input of reading numbers. It is.

This conventionally proposed continuous digit speech recognition method will be briefly explained with reference to FIG.

In Figure 2, 1 is J! W syntax numeral audio input terminal.

2 is an A/D converter, 3 is a buffer, 4 is a feature extraction unit, 5
6 is a standard pattern memory, 6 is a distance calculation section, 7 is a continuous DP processing section, 8 is a logic operation processing section, 9 is a power analysis section, 10 is a limiter, and 11 is an output terminal.

This method calculates the short-term power from the input audio at regular intervals, and then expands the time-series value of the short-time power with an orthogonal function.
Mora (syllables) that compose input speech by orthogonal function expansion
Find the number. Then, if the input voice is a continuous number, the number of digits of 5 is estimated by dividing the value obtained by adding 1 to the obtained mora number and dividing the value by 2, rounding down the decimal parts, and performing recognition processing.

(Problems to be Solved by the Invention) However, in the above-mentioned conventional method, firstly, the continuous DP
Since the matching method is used, the scale of the matching section etc. becomes extremely large, making it unrealistic from an economic point of view.

Second, since there is a limit on the number of digits, it has the advantage that the number of matchings is reduced and the recognition processing time is shortened to some extent compared to a method without a limit on the number of digits. In this case, the number of digits is limited, and the range of use for S-pronunciation is extremely narrow (for example, when inputting a telephone number, ``Zero-ni-igo-ohachi''). Furthermore, related to this second point, thirdly, various units (
For example, "yen" for monetary amount, rKm for distance unit
J, etc.], it is always uttered with a scale such as "O million thousand O hundred O ten O J", so it includes not only consecutive numbers with a standard reading but also other consecutive numbers. Naturally, it is difficult to recognize ordinary consecutive numbers using the above-mentioned algorithm dedicated to "reading numbers". Therefore, when considering continuous digit recognition methods that meet the needs of users in the future, it is thought that a system that has both mechanisms and can immediately respond to input speech will be required.

In view of the above-mentioned problems, an object of the present invention is to detect in a short time whether the input voice has a scale or a reading of the numbers when a number of arbitrary digits is uttered continuously, and after detection, It is an object of the present invention to provide a continuous number recognition method that enables highly accurate recognition processing corresponding to each number.

(Means for solving the problem) In order to achieve this purpose, in the continuous digit speech recognition method of the present invention, especially when digits of N digits (N is any positive integer) are continuously uttered, , the input form recognition is determined by detecting whether the input voice has a scale (00,000,000,000) or whether the number is read in a straight line, and in making this determination, the voice determined by the start frame and the end frame is used. Block the audio power data of the section for each word candidate Next, detect the end frame of this audio power data, and from the time of detection, only the first and second blocks of words are detected from this end frame to the start frame. Recognition processing according to the form is performed in accordance with the determination result obtained at the 0th order, which calculates the distance between the candidate speech pattern and the position standard pattern.

Furthermore, when the determined input form is scale, when recognizing the input voice of consecutive numbers with scale, the distance calculation between the voice pattern and the standard pattern is performed by referring to a rule table that stores rules specific to scale. conduct.

(Function) According to such a configuration, when an M-syllable digit voice with an arbitrary number of digits is input, the voice pattern of the word candidates from the end frame side to the second block from the end frame side among the word candidates of the voice power data of the voice section is input. Since only the matching with the standard pattern is performed, the matching time is shortened.

Furthermore, since the input form of the input voice is determined and speech recognition processing corresponding to the determined input form is performed, recognition accuracy is high, and recognition processing can be performed regardless of the number of digits, so the range of application is extremely wide.

In addition, in the case of voice power data of place value digits, recognition processing is performed while referring to a rule table storing rules specific to place value, so distance calculation and determination can be simplified.

(Embodiments) Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is an explanatory diagram of the consecutive number recognition method of the present invention, and is a block diagram showing one embodiment of the configuration of an apparatus for carrying out this method.

In FIG. 1, 100 is an input terminal, 200 is a frequency analysis section, and 300 is a logarithmic conversion section. 400 is a voice section determining section, 500 is a spectrum converting section, and 600 is a resampling section. Reference numeral 700 denotes an input form recognition determination section, which includes a blocking section 710, a distance calculation section 720, a "place" standard pattern memory 730, a form determination section 740, and each of these sections 710.7
The control section 760 instructs and controls the processing at 20, 730, and 740. 900 is a scale recognition processing section, distance calculation section 91O1 standard pattern memory 92
0, a determination section 930, and a rule table 940. t
ooo is a stick reading recognition processing unit, and a distance calculation unit 1010
, a standard pattern memory 1020, and a determining section 103'C, and 810 is an output terminal.

In such a configuration, continuously uttered numeric sounds are input to the input terminal 100, and the input audio signal input from the input terminal 100 is input to the frequency analysis section 200 to generate a quantum signal corresponding to a plurality of frequency bands. The zero-frequency analyzed signal is sent to the logarithmic conversion unit 300 to obtain logarithmic spectrum information and full range puff information.

The spectral information and power information are sent to the voice section determination section 400, and only the spectral information is sent to the spectral conversion section 500. The spectrum conversion unit 500 normalizes the utterance intensity and sound source characteristics by subtracting the least squares approximation straight line of the voice spectrum from the analysis data that varies depending on the speaker.

This voice section determination unit 400 is configured to perform a start frame (STFR) and an end frame (E
DFR) is detected. The determined start frame (ST
FR) and the end frame (EDFR) and the end frame (EDFR), the spectrum conversion unit 500
It is sent to the re-sampling unit 600 at the same time as the information sent from.

The resampling unit 600 normalizes the time axis of the audio power data. The method of normalizing the time axis is a conventionally known technique, and the linear matching method is a method of dividing the speech section into a fixed number of equal intervals in time determined by the conditions of the recognition device and resampling the same.

λ side US (hereafter) The voice power data resampled by the resampling unit 800 is first sent to the block unit 710 in the input form recognition determining unit 700.The input form recognition determining unit 700 receives the voice power data as voice power data. The 8-digit N number to be input (however,
N is any positive integer. This is the part that determines the type of input, such as whether it is a simple digit or a simple number such as ``ichi ni san yeon.'' This determination process will be explained below.

Figure 3 is a voice power diagram showing, as an example, the state of voice power when pronouncing 4-digit numbers with scale and in stick reading.The horizontal axis shows the time axis, and the vertical axis shows the magnitude of the voice power. be. In the same figure, the upper figure has scales and is ``Rokusengohyakunanajuuni'' (6,557,72
) and the power when uttering ``Rokugo (+) Nanani (i) J (6572)'' are shown in the figure of the f stage in stick reading.

Since the input form is determined using voice power data in units of divided blocks, the blocking unit 710 divides and divides into blocks at locations that are likely to be word sections.

The broken lines in FIG. 3 are called pointers, and are the positions of the boundaries between blocks of words, and the area between the pointers is the sound pattern of word candidates such as numbers and places. Also, the audio power of place value numbers.

The arrows in the audio power section of the stick numbers indicate the positional relationship between the respective pointers.

Now, in the distance calculation unit 710, the audio pattern (the audio pattern of the word candidate) of the audio power data that has been blocked by the blocking unit 710 and the “r” standard pattern memory 730
Perform distance calculations with the "place" standard pattern stored in
In this case, the "place" standard pattern memory 730 stores standard "place" patterns such as "e fistffl, 10,000, 1,000, 100, skewers".

Also, depending on the digit, for example, the number that comes before the digit of ``hundred'', the way in which ``hyaku'' is pronounced changes, such as ``byaku'' or ``byari'', so this modified pattern is also stored.

In order to match these standard patterns with the speech patterns of word candidates, the blocking unit 710 first extracts the end frame (E DFR) of a block of speech power data separated by pointers with small power values. To detect. By detecting this end frame, it is possible to confirm the end of the consecutive utterances.

Next, the pointer is searched from the point at which this end frame is detected, going back in the direction of the start end frame (STFR). This is done by matching the candidate voice pattern with the "rank" standard pattern.

Then, the minimum value of the sum of each distance obtained by calculating the distance with the "place" standard pattern and the voice power data of the voice section for which the distance was calculated are output to the form determining section 740, where the minimum value of the total distance is Depending on the value, it is determined whether the input consecutive numbers have a scale or are in stick reading.

In this way, in this method, the starting frame (S T F
Distances from the "place" standard pattern are calculated for all the blocks between pointers from R) to the end frame (EDFR).

However, when the actual value is uttered, "place" always appears in the first or second block counting from the end frame (EDFR) (for example, ni→jiyuu→
If we focus on the fact that Nana→Hiichi<War...) is coming,
It is possible to distinguish between scale numerals and stick digits by examining only the first digit from the end frame side.
Furthermore, as a method for shortening the matching time, a method may be considered in which only the first and second blocks counted from the end frame (EDFR) of the block are matched with the "place" standard pattern and the determination results are output.

Incidentally, the control section 760 performs, for example, the detection of the end frame, the search for the pointer, the instruction and control of specifying blocks to be matched to shorten the matching time, etc. in the input form recognition and determination section 700.

In this judgment unit 740, if the judgment result is a scale number, it sets a flag (FLAG) = 1, adds this flag = 1 to the voice power data of the voice section that is sent, and also if the judgment result is a decimal number. If it is a flag (FLA
G) A flag=O is added to the voice power data sent as =O, and the result is sent to the next stage recognition processing section 800.

In the recognition processing section 800, the voice power data for which the flag is 1, that is, the input voice is determined to have scale, are sent to the scale recognition processing section 800.

The distance calculation unit 910 performs distance calculation with the standard pattern memory 920. This distance calculation will be explained below.

First, FIG. 4 shows the contents of the standard pattern memory 920. The contents shown in FIG. In addition to the standard patterns for digits in the digit standard pattern memory 730 used for distance calculations in the form determining section 700, standard patterns for each number from 1 to 9 are also provided as shown in FIG. There is. Furthermore, as is clear from this Figure 4, ``1 Kuchi (tsu)'' only exists in the 100's place.
"Ha (tsu)" only exists in the thousand or hundred places, etc.
There is a characteristic pattern that differs from the sound pattern of numbers when spoken alone. Furthermore, in the case of consecutive numbers with scale,
In addition to the characteristic patterns described above, there are certain specific rules as shown in FIG.

FIG. 5 shows an example of the relationship between numbers and scale readings from the thousands digit to the - digit. In this figure, the readings of the numbers are shown in Roman letters. In addition, the place within the loco is the place, and the reading of the transformed place is prescribed. Also, the solid line connected to the loco is a number that connects to the reading of each digit.For example, the thousands digit can be read as “SEN” and (ZEN)" and the number pronounced as "r3" (SAN) is the only one that exists. In addition, the hundreds place is further classified, and it can be seen that there are some special patterns in the number patterns themselves.

The rule table 940 in the scale recognition processing unit SOO stores the feature rules for reading continuous numbers with scale shown in FIG. Therefore, the distance calculation section 810 matches the word candidates already blocked by the blocking section 710, sends the distance calculation result to the determination section 930, and sends the distance calculation result to the determination section 930.
30 refers to rules stored in the rule table 840 during calculation and separates digits and numbers to shorten distance calculation time and judgment time and improve recognition accuracy.

As described above, a comprehensive range is determined using only the limited combinations, and the category with the smallest range is output to the output terminal 810 as a recognition result.

On the other hand, the voice power data indicating that the flag=O1, that is, the input voice is determined to be W5 reading, is sent to the stick reading recognition processing section 1000.

The alternating layer calculation unit 1010 performs distance calculations on the standard pattern stored in the standard pattern memory 1020 one digit at a time for each block-formed numerical word candidate, and sends the result to the determination unit 1030.

The standard pattern memory 1020 differs from the standard pattern memory 920 in that it stores only the reading pattern of @added numbers that corresponds to the reading of numbers.The 9 judgment unit 1030 outputs the total distance and category name as a recognition result to an output terminal. Output from 810. Further, the detailed operations of the distance calculating section 1010 and the determining section +030 of the stick reading recognition processing section 1000 are described in the "voice pattern matching method" proposed in Japanese Patent Application No. 59-58440 filed by the applicant of the present invention. By limiting the slope of the speech steady-state bus, the transient bus is limited to the range where the rate of speech is expected.

In addition, in response to this, a bus setting process tentatively called optimization is repeated a plurality of times.

As can be understood from the above explanation, according to the method of the present invention, when a continuous number voice is input, after determining whether it is a number with a scale or a number with a scale,
It is sufficient to have a configuration that performs the corresponding voice recognition processing, and the specific configuration is not limited to the above-described embodiment.

In addition, the main components 400 and 5 shown in FIG.
00, 800, 710, 720, 740, 7
80,! 310, 930, 1010. Processing 1030 can be performed by a CPU (central processing unit).

(Effects of the Invention) As is clear from the above description, according to the present invention, when arbitrary numbers are uttered continuously, it is possible to detect in a short time whether the input voice has a scale or is a reading of the numbers, Since the recognition process corresponding to each detection is performed after each detection, the recognition process of the uttered continuous numeric voice can be performed with high accuracy, and therefore, the effect of improving the recognition rate and simplifying the calculation process can be expected. Furthermore, since it has the functions of both consecutive number recognition methods, such as the one-place reading method, it can be applied to continuous number recognition devices that meet the needs of users in the future, and has a wide range of applications.

[Brief explanation of drawings]

Fig. 1 is a block diagram for explaining an embodiment of the continuous digit speech recognition method of the present invention, Fig. 2 is a block diagram for explaining a conventional continuous digit speech recognition method, and Fig. 3 is a block diagram for explaining the conventional continuous digit speech recognition method. Voice power diagram when uttering a 4-digit number. Figure 4 is a diagram showing the contents of the standard pattern memory. FIG. 5 is a diagram showing the relationship (rule) between numbers and scale. 100... Input terminal, 200... Frequency analysis section 300... Logarithmic conversion section, 400... Speech interval determination section 500... Spectrum conversion section 600... Re-sampling section 700... Input form Recognition determination unit 710...blocking unit, 720...distance calculation unit 7
30... "place" standard pattern memory 740... form determining section, 760... control section 80
0... Recognition processing unit, 810... Output terminal 900
... Place recognition processing section 910 ... Distance calculation section 920 ... Humiliation pattern memory 930 ... Judgment section, 940 ... Rule table 1000 ... Bar reading recognition processing section 1010 ... Distance@S Section 1020... Standard pattern memory 1030... Judgment section. Patent applicant: Oki Electric Industry Co., Ltd.
// /A-i no Ren” scolding roasted Ryuto recognition method 5 17) S destination B
Moon map figure 2

Claims (1)

[Claims] (1) Recognize whether the input form of continuous numbers in the input voice is continuous numbers with place value or consecutive numbers with bar reading using the place standard pattern, and perform recognition processing of the input voice according to the obtained judgment result. When recognizing an input voice consisting of an arbitrary number of consecutive digits uttered continuously, the voice power data of the voice section determined by the start frame and end frame is divided into blocks for each word candidate in the recognition judgment of the input form. After that, the end frame is detected, and from the time of detection, from the end frame to the start frame side, distance calculation is performed between the voice pattern and position standard pattern of word candidates only in the first and second blocks, Furthermore, in recognition processing of input speech whose input form is determined to be scaled, a distance calculation between the speech pattern and the standard pattern is performed using a rule table storing rules specific to scale, and a final recognition result is output. A continuous digit speech recognition method characterized by the following.