JPS6075894A - Dictionary updating 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 Eleventh Order I The present invention relates to a dictionary updating method in an acoustic recognition device.

Voice pattern matching method that does not use DP method (dynamic programming),
Alternatively, as a method of absorbing variations caused by speakers, a matching method is considered that involves superimposing a pattern that connects the peaks of the time-frequency pattern of the voice with a broad pattern (
Autumn National Conference of the Acoustical Society of Japan (October 1983). An example of this method is shown in FIG. In Fig. 1, 1 is a microphone, 2 is a filter group, 3 is a voice section detection unit, and 4 is a binarization (
161) section, 5 is a 24fi conversion (threshold value 2) section, 6 is a register, 7 is an addition section, 8 is a dictionary section, 9 is a similarity calculation section,
Reference numeral 10 denotes a result display section. First, when registering a dictionary, words to be registered are uttered in order, and each word is uttered multiple times. The sound j1j input from the microphone L is converted into a frequency pattern, outputted with a pause for each voice, and then binarized into the Oll by a certain amount of time (threshold value 2) and stored in the register 6. A second utterance of the same word is then made,
A similar binarized pattern is added to a pattern already stored in the register and stored in the register again. In this way, patterns in which one word is uttered multiple times are added up and registered as many times as there are words in the dictionary.

Next, during recognition, after the speech is detected, it is binarized using a different threshold (threshold 1) than when creating the dictionary.

If threshold value 1>threshold value 2, then the strength of the input pattern or the width of "1" after binarization will be narrower than in the dictionary pattern.

The input pattern is superimposed on each pattern of each word in the dictionary, the degree of similarity is calculated from the degree of overlap, and the dictionary word with the maximum degree of similarity is taken as the recognition result. In general, human speech changes from moment to moment, and changes due to factors such as physical condition, so it is necessary to update the dictionary while using it. For convenience, we will explain the case where Kaga's dictionary is updated three times for a time of l samples.The calculation as shown in FIG.
When updating with 0 pattern, add patterns (a) and (b) and make pattern (C) from 1 or more parts.
A possible method is to subtract ``a'' and update the term a(d) such that the maximum value becomes ``3''. However, in such a method, rlJ of the second frequency sample, which is information specific to the pattern (b) (due to a change in g, etc.), is not reflected in (d). In other words, there is a drawback that the dictionary will not be updated even if it is updated. Alternatively, as shown in FIG. 3, it is conceivable to conversely draw rlJ (b) from the part of the current dictionary pattern that is greater than or equal to "1", and then add an incoming turn (d).

In this case, as shown in FIG. 3, it appears at first glance that the rlJ of the second sample of the frequency of the human cut turn is updated and saved. However, when it is updated next time, the "l" part in (a) is first removed, and the currently updated part becomes O, which has the disadvantage of not being reflected afterwards. Also,
A method as shown in FIG. 4 is also conceivable. In Figure 4, 1
1 is an input pattern register, 12 is an addition section, 13 is a subtraction section, 14 is a dictionary section, 15 is a binarization section, and 16 is a register that converts the corresponding dictionary pattern (Fig. 5 (a)) into two values. After converting into a pattern like (b) and storing it in a register, adding the human cover pattern and the corresponding dictionary pattern, (d), !
The pattern (e) is updated by subtracting the pattern (b). According to this method, the 44 characteristic part A of the input kabataan is stored in the updated dictionary, but 1. For example, the pattern of "l" in the dictionary such as B is "1" even in the updated dictionary. The disadvantage is that even if the value is updated, it will never become greater than "1".

The present invention has been made in order to solve the above-mentioned drawbacks, and in particular, to provide a dictionary updating method for obtaining a dictionary that can maintain a high recognition rate in a speech recognition device. It was made for the purpose of

EMBODIMENT OF THE INVENTION The structure of the present invention will be described below based on examples.

In the present invention, the following operations are performed to achieve the above object. That is, recognition is performed by the speech recognition device, and for those for which a correct answer is obtained, the "0" part in the recognized input pattern is replaced with a negative unit and added to the dictionary pattern to be updated. The negative part of the result is again “0”
'', and if the value exceeds the maximum value of the dictionary pattern, a dictionary update pattern is created by replacing it with the maximum value. The basic conditions for this method are as follows, but we will explain them here for convenience.
It is assumed that a dictionary is created by adding 3 2-binarization patterns (by saying the same word 3 times and adding 37 each of the 2-binarization patterns). That is, the values in the dictionary pattern are integers from O to 3, and the maximum value of the dictionary pattern is 3.

(Condition) If the O of the dictionary pattern and the O of the input pattern overlap, the update pattern is O.tt l tt, Q tt O// //, 2 // Q tt l tttt 3 tt
Q // 2 tt l/ Oll l ll 1 〃 tt l // l // 2// lt 2 tt l tt 3 // lt 3 tt l tt 3 // To achieve this, the configuration of the update method is described above. It was like this.

According to the above condition, when a dictionary pattern as shown in FIG. 7(a) is updated with a human cover turn as shown in FIG. 7(b), a pattern as shown in FIG. 7(f) is obtained.

The L order of undiscovered and light is shown in Fig. 6, and also shown in Fig. 7(a).
, how the pattern in (b) changes to Fig. 7(f)
Figures 7(C) to (f) show how the pattern changes. First, it is determined whether there is a 0 in the position of the person, and the 0 is replaced with 1, which is a negative position (Fig. 7 (C)). This and the dictionary pattern to be updated are added (FIG. 7(d)).

The negative part in the ' pattern of FIG. 7(d) is returned to rQJ (FIG. 7(e)). After that, the maximum value of the pattern is corrected. In this case, the maximum value is 3 because the dictionary is made up of the sum of three patterns. Since other word dictionaries are also structured in this way, all dictionary patterns must always be configured in a rotating configuration. Therefore, by rewriting all the parts "4" and "-" in the pattern of FIG. 6(e) to "3", the pattern of FIG. 7(f), that is, the updated pattern is obtained. This is (f) in Figure 7, which was created as described above.
This shows that the desired update pattern is created in accordance with the above.

Prevention - Eleven As is clear from the above explanation, according to the present invention, as the input pattern changes from moment to moment, the dictionary pattern changes accordingly, and therefore, even if the state of the speaker's voice changes, the recognition rate cannot be improved. High 6' without lowering! A dictionary for maintaining knowledge-V can be obtained.

[Brief explanation of drawings]

FIG. 1 is an electrical block diagram for explaining an example of a conventional voice matching method, FIGS. 2 and 3 are electrical block diagrams for explaining an example of a dictionary updating method, and FIG. , an electrical block diagram illustrating another example of the conventional voice matching method, ・Fifth
The figure is a diagram for explaining an example of the dictionary updating method, FIG. 6 is a flowchart for explaining an embodiment of the present invention, and the 71st sword is a signal pattern diagram for that purpose. 1... Microphone, 2... Filter 〇f, 3... Voice section detection section, 4... Binarization (threshold l) section, 5...
Binarization (14J value 2) section, 6... register, 7...
Addition unit, 8... End aft, 9... Similarity calculation unit, 1
0...Result output section, 11 Large kaputar resistor, 1
2... Addition section, 13... Subtraction section, 14... Dictionary section, 15... Binarization section, 16... Register.

Claims (1)

[Claims] Convert speech into time-frequency patterns and store some of them as a dictionary. When unknown speech is input, this unknown speech is converted into time-frequency patterns in the same way as above and stored in the dictionary. If a correct answer is obtained in a speech recognition device that calculates similarity by overlapping each pattern with each pattern, it changes the "0" part in the human cover turn to a negative unit by 1&, and updates the dictionary pattern to be updated. A dictionary update method characterized in that an update pattern is created by adding the negative part to "0" again, and replacing the part exceeding the maximum value of the dictionary pattern with the maximum value.