JPS58224399A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

TECHNICAL FIELD OF THE INVENTION The present invention relates to a speech recognition device, and in particular to the registration of speech parameters in its speech dictionary.

Background of the technology Generally speaking, speech recognition uses a speech dictionary that stores data (speech parameters) indicating the characteristics of speech, calculates the speech parameters from the input speech, compares them with the speech parameters in the dictionary, and finds a match (if the distance is The input speech is recognized as an in-dictionary speech (word) corresponding to an in-dictionary speech parameter (within a dark value). For voice dictionaries, it is preferable to use a type with σ that can handle unspecified speakers, but this type has a low recognition rate, and to increase this the amount of stored data would be enormous, making comparisons cumbersome and time-consuming. Because of these problems, the current situation is to use one dedicated to each speaker. The speech dictionary for each speaker allows the speaker to input the words they plan to use through a microphone and key in the corresponding words.
The recognition device analyzes the voice input, that is, performs frequency division, sampling, feature extraction, etc. to obtain voice parameters, stores them in a memory together with the keyed-in words, and creates them.

Prior Art and Problems One important application of voice recognition is voice control of equipment, devices, systems, etc. Voice control does not require any keyboard or other operations, so there is no need for a specialized operator.
Although it has the advantage of being convenient for remote control, there is currently a risk of erroneous recognition, and if erroneous recognition occurs, it may lead to a serious accident. Therefore, commands used for voice control in particular need to have a high recognition rate. In voice recognition, as described above, the distance 1i11 is calculated by 1 and the answer is one within the dark value (the difference is within the allowable range), but two or three gaps may appear. Of course, there must be only one answer, and the rest are wrong. If the allowable range is made smaller, the number of i'IJi cases in which two or three answers appear will decrease, but the number of cases in which there are no answers will increase. In order to increase the recognition rate while leaving the dark value as is, it is sufficient to increase the distance between the audio parameters.

Purpose of the Invention The present invention focuses on this point, and when creating a speech dictionary, it is arranged so that important sounds such as control commands are placed at a greater distance from other sounds. . Specifically, the speaker is warned to register the input voice with a short distance by replacing it with a synonymous and different word, thereby increasing the distance between the important voice and other voices to avoid mistakes. This is an attempt to prevent recognition from occurring.

Structure of the Invention The present invention provides a speech recognition device that includes a human processing section that extracts speech parameters of human speech, a speech dictionary that stores the speech parameters, and a control section that compares the input speech parameters and the speech parameters in the dictionary to find corresponding ones. , an area is provided in the voice dictionary for specifying the minimum allowable distance between the voice parameter and other registered voice parameters, and furthermore, when registering the voice parameter, a warning prohibits the registration if the mutual distance between ゛ζ is within the minimum distance. Embodiment 1 of the invention, which is characterized in that it has a generation function, will be described below with reference to embodiments. FIG. 1 shows an outline of the configuration of a speech dictionary according to the invention.

The main part is the audio parameter 10, which indicates the characteristics of the audio of one word, such as "Kanagawa" or "Kawasaki."

Specifically, the sound produced by saying "Kanagawa" into a microphone is recorded in the audio frequency band of ~200 to 5000 KHz.
The frequency is divided through about 16 band-pass filters that divide it into channels, and the output of each channel, which lasts about 1 degree, is sampled with a clock cycle of about 10m5, and the power, fluctuation amount, high frequency power...・No.1
．． Second. Third...Audio parameters are obtained and arranged for each sampling time. In other words, in the figure, it is simply referred to as "voice parameter", but for more details, refer to Section 1.
~ nth audio parameters, each of which is sampled at the sampling time t1. t2... Something that changes at tm.

Item I2 is the voice or word corresponding to voice parameter 10, which is Kanagawa or Kawasaki in the above example, but to give a specific example, it is the address of a table that stores these. When this table is accessed using this item as an address, input signal data for audio or display output for Kanagawa and Kawasaki can be obtained. Category 14 is the category of the item in question; in this example, since the items are Kanagawa and Kawasaki, it is a code indicating the prefecture, city, town, village, etc.). For example, category 1 is a group of voice commands such as start and end, category 2 is a data set of prefecture names, and category 3 is a data set of municipal names, which are used to facilitate and ensure voice recognition. . For example, "Japan" and "Nihon" are easily misrecognized, but the category of the item "Japan" is
The category ``Country name'', 12 items'' stores information indicating the ``number of items'', and when inputting ``Japan'' by voice, if you also input the category ``Country name'' using keystrokes, the voice 872
m, only the speech parameters in the dictionary for the category are extracted and compared, and "Japan" can be easily and quickly recognized as the result.

In the present invention, a column for minimum distance 16 is provided in the dictionary. This indicates the minimum allowable distance of the item (word) you are trying to register to other items or phonetic parameters in the dictionary.
is less than the minimum distance, the recognition device issues a warning and instructs to change the voice to a different word, although the other meanings are the same. For example, when trying to register ``manual'' in a place where ``direction'' is registered, if you say ``shudo,'' the distance will be below the minimum distance, so follow the warning and have the user change the word to ``tedo.'' In addition, in the present invention, the minimum distance is set large for important items such as control commands, so that erroneous voices are not registered. In this way, even if the darkness value is increased to a certain extent, there is no risk of erroneous recognition, and since the darkness value is large, even if the speaker's utterance changes slightly, it can be sufficiently recognized.

FIG. 2 shows an example application. During voice recognition, the speaker specifies that the category of the voice to be input is a voice command using a keyboard operation, etc., and then presses "
"Start" by voice input. The recognition device recognizes the input command and voice, and when it finds that it is a ``start'' instruction, it issues a prompt message ``Please proceed'' indicating that the input is OK, either by voice or by text on the CRT screen.

The speaker then uses the keyboard etc. to select Dataset 1.
, and says "Kanagawa" while indicating that the category of the human-generated voice is the name of the prefecture. The recognition device performs recognition using these input categories and voices, obtains ``Kanagawa,'' and asks for confirmation, ``It's Kanagawa.'' The speaker indicates the category of ``1 voice command'' and inputs [-yes] as a sound. The recognition device now knows that the recognition result is good, and outputs the prompt message 1~, please for the next voice input.

The speaker then specifies the next voice category as data set 2 and inputs "Kawasaki" by voice. This shall apply hereinafter.

By specifying a category, the expected value (input) can be limited for regular processing, so the comparison range can be reduced and the recognition rate can be improved. Then, it is effective to avoid misrecognition by setting a priority for each item (sound or word) within the same category based on the minimum distance, and setting a large minimum distance for important items so that there are no other registered items in close range. It is.

The minimum distance is input by the speaker into the recognition device along with the item and category at the time of registration. It would be advantageous to register items with large minimum distances, that is, important items, before registering other items. Furthermore, when using the factor of category, the minimum distance can be determined within the same category.

FIG. 3 shows an outline of the configuration of the recognition device, and 20 is a CRT display, 22 is a control section, 24 is an input processing section, and 26 is a voice dictionary. Voice is input to the processing unit 24 through a microphone, and the voice parameters are extracted as described above. The control unit 22 includes a host computer HOS'F and a keyboard KB.
It uses the category information input from the keyboard and the voice parameters extracted by the extraction unit 24, refers to the voice dictionary 26, performs the distance calculation described above together with the host computer, performs voice recognition, and performs the voice recognition as described above. It outputs guidance information, warnings, and paraphrase requests by voice, and also displays the CRT display 2.
0 visually.

There are several calculation methods for calculating the distance M, but the simplest method is to calculate the sum of the differences between the input speech parameters and the speech parameters in the dictionary. Also, since the audio parameters are in time series, the input audio parameters for the printer time -1, 2, . ...IPm and the audio parameters in the dictionary DPI, DP2...DP
m at the same timing (i=1.
2. ...Compare IPl and DPj as m)
Compare methods and timings (t, j
Both 1.2.・・・・・・m, but IP as i≠j
There is a DP matching method (comparing i and DPj).

In the present invention, any distance %1tKI calculation method may be used.

As described in detail, in the present invention, the voice dictionary is provided with a minimum ratio R section that specifies the minimum allowable distance from other items, and registration of other items that fall within the minimum distance is prohibited. Avoid misrecognition without changing the margin required for (the above-mentioned dark value),
Accurate speech recognition can be performed.

[Brief explanation of drawings]

FIG. 1 is an explanatory diagram showing an overview of the structure of a speech dictionary, FIG. 2 is an explanatory diagram showing the outline of speech recognition, and FIG. 3 is a block diagram showing an outline of a recognition device. In the drawing, 24 is an input processing section, 26 is a speech dictionary, 22 is a control section, and 16 is a minimum distance designation area. Applicant Fujitsu Ltd. Representative Patent Attorney Minoru Aoyagi Figure 1 Figure 2 Figure 3 0

Claims (1)

[Claims] A speech recognition device comprising: an input processing section that extracts speech parameters of input speech; a speech dictionary that stores the speech parameters; and a control section that compares the input speech parameters and the speech parameters in the dictionary to find corresponding ones. An area for specifying the minimum allowable distance between itself and other registered audio parameters is provided, and a warning generation function is provided to prohibit registration if the mutual distance is within the minimum distance when registering audio parameters. Characteristic voice recognition device.