JPS59124389A - Word voice recognition system - 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] Technical Field of the Invention The present invention relates to a word speech recognition method, and in particular, registers a plurality of sounds for each word in advance, and compares all registered word sounds with unknown input. When the distance is calculated, a predetermined number of items whose calculated distance is less than a threshold value are taken out, the average value for the specific word is determined, and this average value is defined as the distance between the unknown input word voice and the registered word. This invention relates to a system in which recognition results are determined based on the following information.

Conventional technology and problems When recognizing word sounds, the feature parameters of known word sounds are registered in a dictionary by a specific speaker in advance, and this is done by comparing these with the feature parameters of unknown input words. .

Therefore, before recognizing a word sound, a specific speaker inputs a known word sound through a microphone and extracts its characteristic parameters. As a characteristic parameter, 1, for example, 16 types of pandox filters are provided, and the audio signals are converted into outputs for each specific frequency range using these pandox filters, which are A/D converted. The average value over a sampling period of 157,728 is taken as a parameter. Then, these parameters are registered in the registration section. In this case, as a parameter, not only the average output of each filter within such a sampling time but also a formant frequency, LPC, etc. can be used. This registration is not performed only once for the same word, but is performed an appropriate number of times, such as 2 times, 4 times, or 5 times. Therefore, the word S1. S2. For example, when S3... is registered four times, parameters 81 to S1 are registered for 81 as shown in FIG. 1, and parameters 82 for S2 are registered.
~S2 is registered.

For S3, parameters 83 to S3 will be registered.

If the special pregnancy of the currently unknown input word sound X is indicated as the first X, then X and each word S1. S2. From step S3, a fixed number of 9, for example 3, objects that are close to each other are selected, the average distance is determined, and the comparison is performed based on this. At this time, the average distance dx-1 between X and 81 is obtained by the following equation (1): dx-1=+[(XtS;
, s; ):]------(11 Here, for example, (X, S
l) indicates the distance between X and parameter S1, and the same applies to the others.

In this way, the average distance dx-2 between X and 82,
The average distance d knee 3 from

dx−2= , [(X, 82 )+(X S 2
) + (Xl 82 ) ) ”...(2) dz-3=
'((X)8S)+(Xl S”s)+(X
,s, )]・(3) And their average distance d
The smallest size of x-19dx-2, dx-z will be selected as a candidate for , the candidate for X is Sl.

S2. Obtained in the order of S3.

By doing this, 9For example, even if the closest thing to It has the effect of reducing

By the way, the above KNN (K-Nearest N
In the above method, it is necessary to register the feature parameters by inputting the voice for the same word multiple times, but for some reason during this registration, for example, the human's throat is abnormal at that time, or If noise occurs and there is a parameter extracted under such an abnormal condition at the time of registration, the overall average value will change greatly due to this abnormal parameter, resulting in a decrease in recognition accuracy.

OBJECTS OF THE INVENTION An object of the present invention is to improve the problems caused by the presence of characteristic parameters in such abnormal conditions.

A threshold value is set for the parameters when calculating the average value, and abnormally large values are excluded from the average value calculation target. To provide a word speech recognition method in which a word is excluded from a recognition target when a threshold value is exceeded.

Structure of the Invention In order to achieve this object, the word speech recognition method of the present invention includes a registration unit that registers a plurality of sounds for each word in advance;
In a word speech recognition device equipped with a matching unit that compares and determines all the registered word speech and the unknown input word speech, the distance between the unknown input word speech and the registered word is determined by the registration of multiple words that are the same word. An average value calculation unit that calculates the distance between the word sound and the unknown input word sound, and extracts those whose distance is less than a predetermined threshold value in descending order of distance, and a predetermined number of them, and calculates their average value. The method is characterized in that the average value thereof is defined as the distance between the unknown input word voice and the registered word, and the recognition result is determined based on the defined distance.

Embodiment of the Invention An embodiment of the present invention will be described based on FIGS. 2 to 4.

FIG. 2 is a block diagram of an embodiment of the present invention, FIG. 3 is a detailed diagram of the average value calculation section, and FIG. 4 is an explanatory diagram of its operation.

In the figure, 1 is a preprocessing section, and 2 is a parameter extraction section.

3 is a switching section, 4 is a registration section, 5 is a DP verification section, 6 is an average value calculation section, 7 is a threshold value output section, and 8 is a determination section.

The preprocessing unit 1 performs preprocessing for extracting feature parameters, such as keeping the input level of the input voice constant or removing noise.

The parameter extraction unit 2 is equipped with, for example, 16 types of band-pass filters, and extracts parameters by obtaining the average output of each filter at each sampling period.

These are A/D converted and output as digital signals. Of course, if the parameters in this system are to use formant frequencies, a formant extraction means is provided, and if LPC is used, a corresponding configuration is provided.

The switching unit 3 sends the parameters output from the parameter extraction unit 2 to the registration unit 4 and stores them in the case of specific speaker registration, and also sends them to the DP verification unit in the case of verification. It operates to transmit.

The DP comparison section compares the unknown input word speech parameters with all the registered word speech parameters stored in the registration section 4 using a dynamic programming method, and outputs the matching results.

As explained with reference to FIG. 1, the average value calculation unit 6 calculates the average distance between the unknown input and the group of specific word sounds, but in this case, the distance (X, SQ) is larger than the threshold H output from the threshold output unit 7, it is not used and is excluded from the average value calculation target. Note that the above Sn□ is the n-th parameter in the word Sm.

As shown in FIG. 3, the average value calculation unit 6 includes a word identification unit 109
Section holding section 11. Reading section 12. Sequence organizer 13. Count section 14. Comparison part 15. It is composed of a calculation section 16 and the like. Then, when the DPP matching results are sequentially output from the DP verification unit, they are sent to the word identification unit 10 for each same word sound.
, and temporarily holds it in the classification holding section 11 .

Then, these are taken out for each word by the reading unit 12, and as shown in the flowchart shown in FIG. do.

In this case, the number of pieces for which the average value is calculated by the KNN method is determined to be 3. For example, if = 3, the counting unit 14 sequentially sends this to the calculation unit 16, and the counting unit 14 calculates the third value. When the third value is sent out, the comparator 15 is controlled to compare this third value with the threshold H sent out from the threshold output section 7, and if this is less than H, a calculation instruction signal "1" is output to the calculation section 16. Then, the average distance of the three items sent so far is calculated and outputted to the determination unit 8.

However, if this third word exceeds the threshold H, the comparison unit 15 outputs "0" to the calculation unit 16 to indicate that this word is not to be recognized without calculating the average value. The signal is output from the calculating section 16 to the determining section 8. For example, as described above (when k=3 and the threshold value is H, the calculation instruction signal "1" is output from the comparator 15 for each of Sl and S2, but for S3, the third parameter S3
Since the distance between the two and
6 and is output to the determination unit 8.

In this way, the determination unit 8 determines candidates in order of decreasing average distance output from the average value calculation unit 6 without displaying that they are not recognized by the average value calculation unit 6, and outputs them as determination results. I will do it.

Next, the operation of the present invention will be explained with reference to FIG. 2 and other figures.

(1) First, a specific speaker utters each word multiple times as described above, passes through the preprocessing unit 1, extracts the characteristic parameters by the parameter extraction unit 2, and then passes through the switching unit 3. Register in the registration section 4. It is preferable that this input be performed discontinuously rather than by uttering the specific word several times in succession.

(2) Next, when the unknown input word speech X is input from a specific speaker, the preprocessing unit 1 performs the preprocessing as described above, the parameter extraction unit 2 extracts the characteristic parameters, and the parameter is extracted via the switching unit 3. and is transmitted to the DP verification unit 5. Distances between each word registered in the registration unit 4 and a plurality of parameters are calculated, and the distance is output to the average value calculation unit 6.

(3) The average value calculation unit 6 uses the word identification unit 10 to identify the same words in the output from the DP matching unit 5, temporarily stores them in the classification holding unit 11, reads them out in the reading unit 12, The order arrangement unit 13 arranges the words in the same word in descending order of distance and sequentially outputs them to the counting unit 14. The counting section 14 sequentially sends this to the calculation section 16, and when the third one is sent, controls the comparison section 15 to compare this third one with the threshold value H.
Have them compare. If this is less than or equal to H, the comparison section 15 outputs "1" and causes the calculation section 16 to calculate the average distance of these three distances and output it to the determination section 8. However, when it exceeds H, the comparison section 15 outputs "0" and the calculation section 1
6, instead of calculating the average distance, it outputs that this word is not a recognition target. Furthermore, if there is noise or mispronunciation, it will not be possible to extract a predetermined number of all registered words, and the recognition result will be rejected.

(4) The determination unit 8 selects candidates in the order of distance calculated by the average value calculation unit 6 in order of decreasing distance, and outputs these as determination results.

In the above explanation, an example has been explained in which the number of times of registration input of the same word is 4, but the number of registration inputs is of course not limited to this, and can be selected as appropriate. Further, the number of targets for distance calculation in the average value calculation section is not limited to three, and may be two, for example.

An appropriate number smaller than the number of registrations can be selected.

, Effects of the Invention According to the present invention, when an abnormal voice is mixed in the registered voices in the conventional KNN method, the average distance becomes large due to this erroneous voice, resulting in erroneous recognition. This kind of erroneous recognition can be prevented by excluding words that do not reach the threshold value or less from the recognition target.

Not only can recognition accuracy be improved.

At the same time, calculation time can also be reduced.

[Brief explanation of the drawing]

Fig. 1 is a schematic explanatory diagram of the KNN method, Fig. 2 is a configuration diagram of an embodiment of the present invention, Fig. 3 is a detailed diagram of the average direct calculation dj part of the strip, and Fig. 4 is an explanatory diagram of its operation. . In the figure, 1 is the preprocessing section, and 2 is the parameter extraction sound. 3 is a switching section, 4 is a registration section, 5 is a DP collation section, 6 is an average value calculation section, 7 is a threshold value output section, and 8 is a half-circle constant blind. Patent applicant Fujitsu Ltd. Representative Patent Attorney Akira Yamatani

Claims (3)

[Claims] (1) In a word speech recognition device that includes a registration unit that registers a plurality of sounds for each word in advance, and a matching unit that compares and determines all the registered word sounds and unknown input word sounds, an unknown input word As the distance between a voice and a registered word, the distance between a plurality of registered word voices that are the same word and an unknown input word voice is calculated, and those distances that are less than or equal to a predetermined threshold are selected in descending order of distance. An average value calculation unit is provided to extract a predetermined number of words and calculate their average value, define the average value as the distance between the unknown input word voice and the registered word, and determine the recognition result based on the defined distance. A word speech recognition method characterized by the following: (2) The word speech recognition method according to claim 1, wherein registered words that cannot be retrieved in a predetermined number are not subject to recognition result determination. (3) The word speech recognition method according to claim 1, wherein if a predetermined number of all registered words cannot be extracted, the recognition result is rejected.