JPH08202388A - Voice recognition device and voice recognition method - Google Patents

Abstract

PURPOSE: To improve a recognition rate by supplementally using plural voice recognition processing means which are relatively easily obtainable even if recognition systems to satisfy stringent requirements and specifications at a high degree are not established and gradually autonomously adapting these means to a change in working environment and to actual recognition objects. CONSTITUTION: This voice recognition device has the plural voice recognition processing means 1A, 1B having the respectively intrinsic recognition systems, a recognition result consolidating means 2 which totalizes the respective recognition results of these voice recognition processing means 1A, 1B, rearranges these results by the degree of reliability of the respective voice recognition processing means 1A, 1B to be prescribed references, makes overall evaluation thereof and estimates and outputs the final results and a consolidative learning means 3 which consolidates the data for learning by integrating the respective voice recognition processing means 1A, 1B and the recognition result consolidating means 3 based on feedback data from users and knowledge possessed by a learner.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is effectively applied to, for example, various ticket vending machines and the like, and the feature amount of voice data of a target area given as an input is extracted using a computer, and the extraction is performed. The present invention relates to a voice recognition device and a voice recognition method for recognizing and outputting input voice data based on the feature amount.

[0002]

2. Description of the Related Art A conventional general voice recognition apparatus, as shown in FIG. 11, passes a voice input for template creation through a template creation apparatus 100 to create a standard template 101, and a single standard template 101 is provided. The voice recognition processing device 102 performs a recognition process of a new voice input and outputs the recognition result. As for learning, as shown in FIG. 11, a method of giving feedback when the standard template 101 is created,
As shown in FIG. 12, there has been known a method of providing learning feedback using a new voice input.

[0003]

However, in the above-described conventional voice recognition apparatus and voice recognition method, what value or set of values can be considered as the feature amount of the voice data of the target area, and , It is necessary to first establish a speech recognition processing means of a high-level and strict specification recognition method such as a calculation method for calculating the feature amount from input speech data, the target to be recognized is complicated, and As a feature to be output as a recognition result requires a high degree of sophistication, it becomes very difficult to first determine the feature amount calculated at the time of recognition. As a result, if the quality of the standard template 101 is poor, the recognition rate will decrease over time, and in order to improve the quality, there is a problem that a great deal of labor and advanced technology are required.

Regarding learning, the above-mentioned conventional method is limited to learning by a single module and is not effective learning such as sequential learning. There is a problem that it is inevitable that the recognition rate will be lower than the required specifications due to the deviation from the assumption at the time of construction and the change over time of the device itself.

The present invention has been made in view of the above circumstances, and does not establish a recognition method satisfying a high level and strict requirement specification, but complements a plurality of voice recognition processing means which are relatively easy to obtain. It is a main object to provide a voice recognition device and a voice recognition method which can improve the recognition rate by using the voice recognition device.

Another object of the present invention is to make it possible to gradually and autonomously adapt to changes in the operating environment and actual recognition targets, and to improve the recognition rate over time. is there.

[0007]

In order to achieve the above main object, a voice recognition apparatus according to the invention of claim 1 provides a plurality of voice recognition processing means and a recognition result of each of the plurality of voice recognition processing means. A recognition result integrating means for collectively evaluating the results according to a predetermined standard and outputting a final result is provided.

In order to achieve the above-mentioned other object, a voice recognition apparatus according to the invention of claim 2 has a plurality of voice recognition processing means and the recognition results of each of the plurality of voice recognition processing means as a predetermined criterion. And a recognition result integrating means for comprehensively evaluating and outputting a final result, and an integrated learning means for integrating data for collectively learning the plurality of voice recognition processing means. Is.

As the recognition result integrating means in the voice recognition device according to claim 1 or 2, as in claim 3,
Among the recognition results of each of the plurality of voice recognition processing means, a recognition selection means that selects a recognition result of the recognition processing means satisfying a predetermined criterion, comprehensively evaluates the selected recognition results, and outputs a final result. Preference is given to using.

Further, in the voice recognition device according to any one of claims 1 to 3, it is preferable that a plurality of voice segment cutout means are provided in front of the plurality of voice recognition processing means, as in claim 4. As in claim 5, it is desirable that different types of speech recognition processing means are used.

Further, as the voice recognition processing means in the voice recognition apparatus according to the second aspect of the present invention, as in the sixth aspect, the voice recognition means having the standard pattern, the short-term learning, the voice recognition means having the standard pattern, and the long-term learning. It is composed of three types of voice recognition means having a standard pattern. Further, as the integrated learning means, the short-term learning standard pattern reflects the learning result for a first predetermined period, and the long-term learning standard pattern is the first learning pattern. It is desirable that the learning result is reflected over a second predetermined period that is longer than the predetermined period, and here, as in claim 7, the integrated learning means is configured so that each of the three types of speech recognition processing means has It is desirable that the recognition results are collectively evaluated according to a predetermined standard, and the final result is output.
As described above, it is possible to provide a pattern evaluation means for evaluating the use frequency of each of the short-term learning standard pattern and the long-term learning standard pattern on a predetermined scale, and erasing unnecessary learning standard patterns based on the evaluation. Especially desirable.

A speech recognition method according to a ninth aspect of the present invention is characterized in that the recognition results obtained by each of the plurality of speech recognition processing means are collectively evaluated according to a predetermined standard and a final result is output. is there.

Further, the voice recognition method according to the invention of claim 10 is characterized in that the recognition results obtained by each of the plurality of voice recognition processing means are collectively evaluated according to a predetermined standard, and a final result is output. It is a thing.

In the speech recognition method according to claim 9 or 10, the recognition result satisfying a predetermined criterion is selected from the recognition results obtained by each of the plurality of speech recognition processing means, and those are selected. It is preferable to comprehensively evaluate the recognition results and output the final result.

Further, in the voice recognition method according to any one of claims 9 to 11, as in claim 12, there is provided a plurality of voice section cutout means for cutting out the input voice into a plurality of sections.
It is preferable that the voices of a plurality of sections cut out by the plurality of cutout means are recognized by a plurality of voice recognition processing means, and the plurality of voice recognition processing means have different types. It is desirable to use one.

As the speech recognition processing means in the speech recognition method according to the tenth aspect of the invention, as in claim 14, the speech recognition means having a standard pattern, the speech recognition means having a short-term learning standard pattern, and the long-term learning standard pattern. Using the three types of speech recognition means having, the short-term learning standard pattern reflects the learning result over the first predetermined period,
Moreover, it is desirable that the long-term learning standard pattern reflects the learning result over a second predetermined period that is longer than the first predetermined period, and here, as in claim 15,
It is desirable to collectively evaluate the recognition results obtained by each of the three types of speech recognition processing means according to a predetermined standard and output a final result. Furthermore, as in claim 16, the short-term learning standard pattern and the long-term learning standard pattern are obtained. It is particularly desirable to evaluate the use frequency of each learning standard pattern on a predetermined scale and eliminate unnecessary learning standard patterns based on the evaluation.

[0017]

According to the inventions of claims 1 and 9, when voice data is input, a plurality of voice recognition processing means output respective recognition results in accordance with their own recognition methods, and the plurality of recognitions are performed. The results are aggregated and collectively evaluated according to a predetermined standard, and the final result is output. In this way, a plurality of recognition methods that are relatively easy to obtain are used in a complementary manner, and the results obtained by these recognition methods are collectively evaluated to comprehensively evaluate the final speech with a high recognition rate that satisfies the required specifications. It is possible to output the recognition result.

According to the second and tenth aspects of the present invention, while outputting the final voice recognition result as described above, based on feedback from the user based on the result, the above plurality of By collectively learning the voice identification processing means so as to improve the recognition suitability in the current input voice data, the voice recognition processing means is gradually and autonomously adapted to a change in the operating environment or an actual recognition target, thereby recognizing the recognition rate. It is possible to improve over time.

Further, according to the inventions of claims 3 and 11, when the recognition results of the plurality of speech recognition processing means are collectively evaluated, the recognition results of the recognition processing means satisfying a predetermined criterion are selected. By comprehensively evaluating the selected recognition result and outputting the final result, it is possible to output the final speech recognition result having a higher recognition rate.

Further, in executing the above-mentioned voice recognition processing, a plurality of voice section cutout means for cutting out the input voice into a plurality of sections are provided, and a plurality of pieces cut out by these means are provided. By recognizing the voice of the section by the plurality of voice recognition processing means, among the plurality of cutout means whose superiority or inferiority changes according to various conditions, the most accurate cutout means at the time of voice input is efficiently used. The accuracy of voice segment extraction is improved, and as a result, the voice recognition rate can be improved.

In executing the above-mentioned voice recognition processing, the number of recognizable vocabularies is voice-recognized by using a plurality of voice recognition processing means of different types as in claims 5 and 13. It becomes easy to increase the number of processing means without increasing the calculation cost.

Further, according to the inventions of claims 6 and 14, in the initial state, the voice recognition means having a standard pattern in which voice data registered in advance at the time of factory shipment is stocked is predetermined. At the same time as performing voice recognition processing,
Executes a learning processing algorithm such as extracting feature data and creating a short-term learning standard pattern and a long-term learning standard pattern by clustering when it is stocked above a certain amount, and thereafter, the standard pattern and short-term learning standard pattern. By executing a normal learning processing algorithm such as performing a predetermined speech recognition process using three types of speech recognition means having a long-term learning standard pattern, the learning result of each pattern is efficiently reflected, and the standard pattern Irrespective of whether it is good or bad, and without requiring a large amount of calculation such as a neural network,
It is possible to realize high-speed and flexible voice recognition in response to various changes in the situation and improve the recognition rate.

Further, in the above-mentioned voice recognition, as in claims 7 and 15, the recognition results of each of the plurality of voice recognition processing means are collectively evaluated according to a predetermined standard, and the final result is output. As a result, it becomes possible to output the final voice recognition result having a high recognition rate that satisfies the required specifications.

Particularly, in the above speech recognition, as in claims 8 and 16, unnecessary use standard patterns are evaluated by evaluating the use frequency of each of the short-term learning standard pattern and the long-term learning standard pattern by a predetermined scale. Is eliminated and only the effective learning standard pattern is left, so that the power consumption of the computer can be reduced.

[0025]

Embodiments of the present invention will be described below with reference to the drawings. Embodiment 1 FIG. 1 is a schematic configuration diagram of a voice recognition apparatus according to Embodiment 1 provided with two voice recognition processing means. In FIG. 1, 1A and 1B are two voice recognition processing means, and these voice recognition processing means are provided. The processing means 1A, 1B use the standard templates 1a1, 1b1 in which the knowledge data used for recognition are stocked and the knowledge data of these standard templates 1a1, 1b1 to output the respective recognition results according to their own recognition methods. It is composed of processing units 1a2 and 1b2. Reference numeral 2 is a voice section cutout unit that cuts out the input voice in a predetermined section, and the voice of the section cut out by this is the voice recognition processing unit 1 of the two voice recognition processing units 1A and 1B.
It is input to a2, 1b2.

Reference numeral 3 is a recognition result integrating means, which collects the recognition results of the two voice recognition processing means 1A, 1B and also serves as a predetermined reference.
Based on the reliability of B (described later), a comprehensive evaluation is performed and the final result is estimated and output. Reference numeral 4 denotes an integrated learning means, which collectively learns each of the voice recognition processing means 1A and 1B and the recognition result integration means 3 based on feedback data from the user and knowledge possessed by the integrated learning means 4 itself. Integrate the data to do so.

Next, the voice recognition operation according to the first embodiment having the above configuration will be described with reference to the flowchart of FIG. For example, the voice input through a microphone or the like is cut out by the voice section cutout unit 2 over a predetermined section, and the voices of the cutout sections are then recognized by the two voice recognition processing units 1A and 1B.
It is input to 1b2 (step S10). The input voice is subjected to recognition processing according to the recognition method unique to each of the voice recognition processing units 1a2 and 1b2 (step S11), and the result is output.

Next, each of the above speech recognition processing means 1A, 1
The recognition results output from the speech recognition processing units 1a2 and 1b2 of B are aggregated by the recognition result integration unit 3 (step S1).
Along with 2), the correction distance of the recognition result from each of the voice recognition processing means 1A and 1B is obtained using the reliability as a predetermined reference, and the recognition result is comprehensively evaluated from the correction distance to estimate and output the final result. (Step S13). Based on the output final result, feedback data is input from the user to the integrated learning means 4 (step S14),
Based on the feedback data from the user and the knowledge possessed by the integrated learning means 4 itself, the reliability of the recognition method in each of the voice recognition processing means 1A, 1B is updated (step S15) and the current input is made. The speech recognition processing sections 1a2, 1a2 of the speech recognition processing means 1A, 1B are arranged so as to improve the adaptability of speech recognition.
The learning process of 1b2 is performed (step S16).

Next, each of the voice recognition processing means 1A, 1
When the learning process of the speech recognition processing units 1a2 and 1b2 of B is completed, the learning process of the integrated learning unit 4 is performed so that the final result matches the feedback data from the user according to the fitness of the recognition after the learning. I do.

It should be noted that the updating of the reliability used in the comprehensive evaluation by the recognition result integrating means 3 and the reliability of the recognition method by the integrated learning means 4 in the above-described voice recognition device will be described in detail. Each of the above speech recognition processing means 1A, 1
The reliability (S) of each B is: S = hit / bat (1) where bat is the number of voice recognition executions hit: Among the number of voice recognition executions, the recognition result includes the correct answers up to the third place It is calculated by the number of times (3rd place recognition). As the reliability (S) is closer to 1, the importance of the voice recognition processing means is higher, and as the reliability is closer to 0, the importance of the voice recognition processing means is lower. The reliability of the final recognition result is improved by teaching (learning process) the recognition result integrating means 3 with this reliability (S).

An example of obtaining the reliability data of each of the voice recognition processing means 1A and 1B by using the above equation (1) is shown below.
It becomes like Table 1.

[0032]

[Table 1]

Using the reliability (S) described above, a distance (distance;) which is a general method showing the degree of similarity between the input voices obtained by the respective voice recognition processing means 1A, 1B and the templates 1a1, 1b1. d) correct dist
ance; cd) is calculated by cd = d ÷ S (2). The distance (d) and the correction distance (cd) of each of the voice recognition processing means 1A and 1B are calculated by using the equation (2).
Table 2 shows an example of the calculation.

[0034]

[Table 2]

By the way, in each of the two speech recognition processing means 1A and 1B in the above-mentioned first embodiment, the respective templates 1a1 and 1b1 are different voicing methods for the same word (by gender, age, region, etc.), Further, when the templates are created in different formats, by using the ones corresponding to the respective templates 1a1, 1b1 as the voice recognition processing units 1a2, 1b2, the recognition results are improved by integrating the respective recognition results. Can be achieved. Further, if the voice recognition processing units 1a2 and 1b2 of the two voice recognition processing means 1A and 1B in the first embodiment are the same and the respective templates 1a1 and 1b1 are created for different words, recognition is possible. It is possible to increase the number of words, that is, the number of recognized vocabulary words, without increasing the calculation cost. Further, the templates 1a1 and 1b1 of the two respective voice recognition processing means 1A and 1B in the above-described first embodiment are created by different voicing methods (by gender, age, region, etc.) for different words and in different formats. If the template is
By using the ones corresponding to the respective templates 1a1, 1b1 as the speech recognition processing units 1a2, 1b2, the number of recognizable words, that is, the number of recognized vocabularies, is calculated by the number of the speech recognition processing units 1a2, 1b2. It can be increased without causing a rise.

Embodiment 2: FIG. 3 shows a voice recognition processing means 3
FIG. 3 is a schematic configuration diagram of a speech recognition apparatus according to a second embodiment provided therein, in which FIG. 1A, 1B, 1C are three speech recognition processing means, and these three speech recognition processing means 1A, 1C.
B and 1C output the respective recognition results according to their own recognition methods using the standard templates 1a1, 1b1, 1c1 in which the knowledge data used for recognition are stocked and the knowledge data of these standard templates 1a1, 1b1, 1c1. Speech recognition processing unit 1a2, 1b2, 1c2
Consists of Since the other configurations are the same as those of the first embodiment shown in FIG. 1, the same reference numerals are given to the same portions, and the description thereof will be omitted.

The voice recognition operation according to the second embodiment having such a configuration is almost the same as that of the first embodiment, and the reliability data of the respective voice recognition processing means 1A, 1B, 1C in this case are described above. Table 3 shows an example obtained by the equation 1).
become that way.

[0038]

[Table 3]

Table 4 shows an example in which the distance (d) and the correction distance (cd) of each of the voice recognition processing means 1A, 1B, 1C are obtained by using the above equation (2). .

[0040]

[Table 4]

Looking at Table 4, the speech recognition processing means 1A
Only No. 1 Although the distance is not the shortest, since the reliability (S) is high, it is ranked first in the integrated result of the comprehensive evaluation. In this way, according to the principle of majority voting, the words that should be recognized as the first place may fall below the second place. Based on such a result, the recognition result integrating means 3 sequentially discards the recognition results of the speech recognition processing means having a low reliability (S) and a poor integration result, and the speech recognition processing means having a high reliability (S). It is possible to further improve the reliability of the voice recognition result by selecting only the recognition result and having a recognition selection function of comprehensively evaluating the selected recognition results and outputting the final result.

Third Embodiment: FIG. 4 shows a third embodiment in which two voice section cutting means and two voice recognition processing means are provided.
2 is a schematic configuration diagram of a voice recognition device according to the first embodiment. In the figure, the difference from the first embodiment is that two voice segment cutouts are provided before the two voice recognition processing units 1A and 1B to cut out input voices in predetermined sections. Means 2A and 2B are provided, and the voices of the sections cut out by these two voice section cutout means 2A and 2B are used as the two voice recognition processing means 1A and 1B.
The other components are the same as those of the first embodiment shown in FIG. 1, and therefore the same parts are designated by the same reference numerals and their description is omitted. To do.

In the case of the third embodiment having the above configuration, the superiority or inferiority changes depending on various conditions such as parameter setting.
Of the two voice section cutout means 2A and 2B, the voice section cutout accuracy can be improved by efficiently using the most accurate cutout means at the time of voice input, and as a result, the voice recognition rate can be improved. Can be improved.

Fourth Embodiment: FIG. 5 is a block diagram showing a concrete configuration of each of the two voice recognition processing means 1A and 1B of the voice recognition apparatus in the first embodiment. In FIG. 5, 5 is A / The D converter converts the input voice into digital data. A time normalization / feature extraction unit 6 extracts feature data from the input voice that has not been converted into digital data. Reference numeral 7 denotes a standard pattern creation unit, which uses the voice feature data extracted by the time normalization / feature extraction unit 6
Three types of standard patterns (corresponding to the template of the first embodiment) are created. As an area for stocking the voice feature data in order to create such three kinds of standard patterns, as shown in FIG. 6, an effective learning data area Temp and two learning data areas Temp1,
Three types of Temp2 are prepared. Of these, two learning data areas Temp1 and Temp2 accumulate voice feature data for creating a short-term learning standard pattern, and the effective learning data area Temp uses a short-term learning standard pattern and a long-term learning standard pattern described later as standard data. In doing so, the contribution, which is the frequency of use, is evaluated, and the voice feature data with high contribution is stored as long-term effective data.

As shown in FIG. 7, the three types of standard patterns (templates) are standard patterns Default created by clustering voice data registered in advance at the time of factory shipment and the learning data area Temp1.
Alternatively, the short-term learning standard pattern Learn1 created by clustering the speech feature data accumulated in Temp2 as short-term learning standard data and the above-mentioned speech feature data having a high contribution are clustered from the valid learning data area Temp as long-term effective data. There are three types of long-term learning standard pattern Learn2 created by the above.

In FIG. 5, a pattern collating unit 8 collates the voice characteristic data stored in the above three types of standard patterns with the actual input voice characteristic data. A result selection unit 9 selects the matching result of the pattern matching unit 8 and outputs the recognition result. A contribution degree evaluation unit 10 evaluates the use frequency of each of the short-term learning standard pattern Learn1 and the long-term learning standard pattern Learn2 on a predetermined scale. With the above configuration, the voice recognition processing means 1A, 1B
Each of them consists of three types of speech recognition means having a standard pattern Default, speech recognition means having a short-term learning standard pattern Learn1, and speech recognition means having a long-term learning standard pattern Learn2. The speech recognition processing means 1A, 1B are provided with a recognition result integrating means and an integrated learning means for comprehensively evaluating the recognition results output from the pattern matching parts 8 and outputting the final results. Since they are the same as those described, their illustration and description are omitted.

As a contribution level evaluation method in the contribution level evaluation unit 10, a high score is assigned to high-quality data by referring to the partition matrix of the voice data that is close in distance to the standard pattern. It is possible to use a method of attaching, a method of simply using the number of times used for voice recognition as a contribution degree (score), a method of using a hit rate as a contribution degree (score), etc. Explaining an example using the rate, the contribution degree (score)
Is score = hit ÷ bat (3) where bat: learning data area Temp1 or Temp
Number of utterances of voice feature data (word) accumulated in p2 hit: 1 out of number of utterances of voice feature data (word)
It is calculated by the number of times it is recognized as a rank, and a specific example is shown in Table 5.

[0048]

[Table 5]

Next, the learning algorithm of the voice recognition processing means according to the above-mentioned fourth embodiment will be explained. The learning processing algorithm from the initial state is as shown in the flowchart of FIG. 8. When voice is input through a microphone or the like (step S20), the input voice is converted into digital data by the A / D conversion unit 5. After being converted,
The feature data is extracted by the time normalization / feature extraction unit 6 (step S21). Then, the standard pattern Defau
The recognition process is performed using only the voice recognition means having lt (step S22), and if it is no in the determination (step S23) as to whether or not it is recognized in the first place, the extracted feature data is the learning data. The data is stored in the area Temp2 (step S24). Then, the learning data area Tem
When p2 is full (step S25), the short-term learning standard pattern Learn1 is created by clustering the learning data area Temp2 (step S26).

After that, the process shifts to a normal learning processing algorithm. This normal learning processing algorithm is as shown in the flowchart of FIG. 9. When a voice is input (step S30), the input voice is converted into digital data by the A / D converter 5, The feature data is extracted by the time normalization / feature extraction unit 6 (step S
31). Then, for the extracted feature data, the speech recognition means having the standard pattern Default, the short-term learning standard pattern Learn1 and the long-term learning standard pattern L
A recognition process using earn2 is performed (step S3).
Along with 2), the contribution levels of the short-term learning standard pattern Learn1 and the long-term learning standard pattern Learn2 are evaluated (step S33). Then, the above recognition processing is 1
It is determined whether or not the position was recognized (step S34),
In the case of no, the extracted feature data is stored in the learning data area Temp1 (step S35), and when the learning data area Temp1 is full (step S3).
In 6), the short-term learning standard pattern Learn1 having a high contribution is stored from the learning data area Temp2 to the effective learning data area Temp, and the rest is deleted (step S37).

At the same time, the filled learning data area Te
Another learning data area Temp2 in which a blank is generated from mp1
The extracted characteristic data is transferred to and stored (step S38),
The short-term learning standard pattern Learn1 is created by clustering the learning data area Temp2 (step S39). By repeating this, when it is determined that the effective learning data area Temp is full (step S40), the effective learning data area Temp is clustered to create the long-term learning standard pattern Learn2 (step 41). .

In the operation of step 33 in the flow chart of FIG. 9, the short-term learning standard pattern Lear is
The algorithm at the time of the contribution evaluation when the contributions (score) of n1 and the long-term learning standard pattern Learn2 are in the states of Table 5 described above is as shown in FIG. That is, it is determined whether or not the number of utterances of the characteristic data (word) exceeds a predetermined value (const1) (step S50), and when the number of utterances exceeds the predetermined value (const1), the hit rate is a predetermined value (const2). It is determined whether or not it is below (step S51), and if the hit rate is below a predetermined value (const2), the learning data area T
The characteristic data stored in emp1 and Temp2 are deleted (step S52). This makes it possible to reduce the power consumption of the computer as much as possible.

[0053]

As described above, according to the inventions of claims 1 and 9, a plurality of voice identification processing means are used,
Even if the object to be recognized is complicated, it can be used to compliment the advanced and strict requirement specifications by complementarily using the relatively easy recognition methods that these have and collectively evaluating the results. This has the effect of significantly improving the recognition rate without establishing a method.

Further, according to the inventions of claims 2 and 10, while outputting the final voice recognition result as described above, based on feedback from the user based on the result, the plurality of voices are output. By collectively learning the identification processing means so that the recognition suitability in the current input voice data is improved, the identification processing means is gradually and autonomously adapted to changes in the operating environment and the actual recognition target, thereby improving the recognition rate. This has the effect of improving over time.

According to the invention of claims 3 and 11, the recognition result of the recognition processing means satisfying a predetermined criterion is selected from the recognition results of the plurality of speech recognition processing means, and the selected recognition is performed. It is possible to comprehensively evaluate the results and output the final result, and it is possible to further improve the recognition rate.

Further, in executing the above-mentioned voice recognition processing, as in claims 4 and 12, the input voice is cut out by a plurality of voice section cutout means and is recognized by a plurality of voice recognition processing means. Among a plurality of cutting means whose superiority or inferiority changes depending on various conditions, the most accurate cutting means at the time of voice input is efficiently used to improve the voice segment cutout accuracy, and as a result, the voice recognition rate is further improved. It is possible to improve.

Further, in executing the above-mentioned voice recognition processing, by using a plurality of voice recognition processing means of different types as in claim 5 and claim 13, the number of recognizable vocabulary is recognized by the voice recognition. It can be easily increased by the number of processing means without increasing the calculation cost.

Further, according to the inventions of claims 6 and 14, the voice recognition means having a standard pattern in which the voice data registered in advance at the time of factory shipment is stocked and the voice using the standard pattern. By using three types of speech recognition means, that is, a speech recognition means having a short-term learning standard pattern and a long-term learning standard pattern created by storing the feature data extracted at the time of recognition processing, the learning result of each pattern can be efficiently obtained. It is possible to realize fast and flexible voice recognition in response to various situation changes regardless of whether the standard pattern is good or bad and without requiring a large amount of calculation such as a neural network. Thus, the recognition rate can be improved.

Further, in the above speech recognition, as in claims 7 and 15, the recognition results of each of the plurality of speech recognition processing means are collectively evaluated based on a predetermined standard, and the final result is output. By doing so, it is possible to obtain a high recognition rate that meets the high and strict requirement specifications.

Particularly, in the above-mentioned speech recognition, as in claims 8 and 16, unnecessary use of the learning standard pattern is evaluated by evaluating the use frequency of each of the short-term learning standard pattern and the long-term learning standard pattern by a predetermined scale. , And leave only the effective learning standard pattern, the power consumption of the computer can be reduced.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram of a voice recognition device according to a first embodiment of the present invention.

FIG. 2 is a flowchart illustrating an operation of the voice recognition device according to the first embodiment.

FIG. 3 is a schematic configuration diagram of a voice recognition device according to a second embodiment of the present invention.

FIG. 4 is a schematic configuration diagram of a voice recognition device according to a third embodiment of the present invention.

FIG. 5 is a voice recognition device 2 according to the first embodiment shown in FIG.
It is a block diagram which shows the concrete structure of each one speech recognition processing means.

FIG. 6 is a configuration diagram of voice feature data according to the fourth embodiment illustrated in FIG. 4.

FIG. 7 is a configuration diagram of a standard pattern of Example 4 shown in FIG.

FIG. 8 is a flowchart illustrating a learning processing algorithm from the initial state according to the fourth exemplary embodiment illustrated in FIG.

FIG. 9 is a flowchart illustrating a normal learning processing algorithm of the fourth embodiment illustrated in FIG.

FIG. 10 is a flowchart illustrating an algorithm for evaluating a contribution degree according to the fourth exemplary embodiment illustrated in FIG.

FIG. 11 is a schematic configuration diagram of a conventional voice recognition device.

FIG. 12 is a schematic configuration diagram of a conventional speech recognition learning device.

[Explanation of symbols]

 1A, 1B Speech recognition processing means 2, 2A, 2B Speech segment cut-out means 3 Recognition result integration means 4 Integrated learning means 10 Contribution evaluation part Default standard pattern Learn1 Short term learning standard pattern Learn2 Long term learning standard pattern

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Masatomo Hashimoto 68 Izumi-cho, Takano Izumi-cho, Sakyo-ku, Kyoto-shi, Kyoto Berge 56-606

Claims (16)

[Claims] 1. A plurality of voice recognition processing means, and a recognition result integrating means for collectively evaluating the recognition results of the plurality of voice recognition processing means according to a predetermined standard and outputting a final result. Characteristic voice recognition device. 2. A plurality of voice recognition processing means, a recognition result integrating means for collectively evaluating the recognition results of the plurality of voice recognition processing means according to a predetermined standard and outputting a final result, and the plurality of voices. A speech recognition apparatus comprising: integrated learning means for integrating data for collectively learning recognition processing means. 3. The recognition result integrating means selects a recognition result of a recognition processing means satisfying a predetermined criterion from among the recognition results of the plurality of voice recognition processing means, and comprehensively evaluates the selected recognition results. The voice recognition device according to claim 1 or 2, further comprising a recognition selection unit that outputs a final result. 4. A plurality of voice segment cutout means are provided in front of the plurality of voice recognition processing means.
The voice recognition device according to any one of 3 above. 5. The voice recognition device according to claim 1, wherein the plurality of voice recognition processing means are of different types. 6. The voice recognition processing means comprises three types, a voice recognition means having a standard pattern, a voice recognition means having a short-term learning standard pattern, and a voice recognition means having a long-term learning standard pattern, and the integrated learning. A means reflects the learning result on the short-term learning standard pattern over a first predetermined period, and reflects the learning result on the long-term learning standard pattern over a second predetermined period longer than the first predetermined period. The voice recognition device according to claim 2. 7. The voice recognition device according to claim 6, wherein the integrated learning means collectively evaluates the recognition results of the three types of voice recognition means by a predetermined standard and outputs a final result. . 8. A pattern evaluation means for evaluating the use frequency of each of the short-term learning standard pattern and the long-term learning standard pattern by a predetermined scale and erasing unnecessary learning standard patterns based on the evaluation. Item 6 or 7
The voice recognition device described in. 9. A voice recognition method, wherein the recognition results obtained by each of the plurality of voice recognition processing means are collectively evaluated according to a predetermined standard and a final result is output. 10. The recognition results obtained by each of the plurality of voice recognition processing means are collectively evaluated according to a predetermined standard, a final result is output, and the plurality of voice recognition processing means are collectively learned. Characteristic voice recognition method. 11. A recognition result satisfying a predetermined criterion is selected from the recognition results obtained by each of the plurality of voice recognition processing means, and the selected recognition result is comprehensively evaluated to output a final result. 10. The voice recognition method according to 10. 12. A method according to claim 9, further comprising a plurality of voice section cutout means for cutting out the input voice into a plurality of sections, wherein the plurality of voice sections cut out by the plurality of cutout means are recognized by a plurality of voice recognition processing means. The voice recognition method according to any one of 1 to 11. 13. The plurality of voice recognition processing means,
The voice recognition method according to claim 9, wherein different types are used. 14. As the voice recognition processing means, three types of voice recognition means having a standard pattern, voice recognition means having a short-term learning standard pattern and voice recognition means having a long-term learning standard pattern are used, and the short-term learning standard is used. The learning result is reflected in a pattern for a first predetermined period, and the learning result is reflected in a second predetermined period, which is longer than the first predetermined period, for the long-term learning standard pattern.
The voice recognition method described in 0. 15. The voice recognition method according to claim 14, wherein the recognition results obtained by each of the three types of voice recognition means are collectively evaluated according to a predetermined standard and a final result is output. 16. The method according to claim 10, wherein the use frequency of each of the short-term learning standard pattern and the long-term learning standard pattern is evaluated by a predetermined scale, and unnecessary learning standard patterns are deleted based on the evaluation. Speech recognition method.