JPH10274994A - Similar word identification method and device - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To identify a similar word that is easily misrecognized when performing word recognition using a DRNN word model with high accuracy. SOLUTION: As one example of a method for identifying a similar word,
When a voice of a certain word is input, a DRNN output corresponding to the input word voice data is output from the word detection signal output unit 4 using a DRNN word model, and the input word voice data is encoded using a codebook 7. Convert to data. When a DRNN output indicating certainty or more certainty is output from the word detection signal output unit 4, the recognition processing unit 9 sets a predetermined section including the characteristic portion of the input word in the DRNN output, The coded code data is checked in the set predetermined section, and based on the result, an input word and a word similar to the input word are identified.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a similar word identification method and a similar word identification device for identifying words having similar pronunciations, and more particularly to a dynamic random current (DRNN) as one of voice recognition techniques for unspecified speakers. Neural Network
s) A similar word identification method and a similar word identification device in a speech recognition technology using a word model.

[0002]

2. Description of the Related Art As one of speech recognition techniques for unspecified speakers, there is a speech recognition technique using a DRNN word model. An application has been filed in JP-A-6-119476.

[0003] This DRNN word model is designed so that when a feature vector sequence of a certain word is input as time-series data, an appropriate output for the word is obtained.
The weight and bias between each unit are determined in accordance with a predetermined learning rule, whereby an output similar to the teacher output for the word is obtained for the voice data of the word spoken by a certain unspecified speaker. ing.

[0004] For example, "good morning" of a certain unspecified speaker
When the time series data of the feature vector sequence of the word “good morning” is input, in order to obtain an output close to the ideal output (teacher output) for the word “good morning”, the feature vector of the word “good morning” at each time Is given to the corresponding input unit, and is converted by the weight and bias set according to the learning rule. This is input as a time series data.
The time series processing is performed for all the feature vector strings of one word for each time. In this way, for speech data of a word spoken by an unspecified speaker, an output close to a teacher output for the word is obtained.

[0005] As described above, the learning rule for changing the weight of the DRNN speech model prepared corresponding to all the words to be recognized so that an appropriate output is obtained for each word is described in IEICE technical report: IEICI sp92-125 (1993-0
It is described on pages 17 to 24 of 1).

The speech recognition using the DRNN word model previously learned for some words will be briefly described with reference to FIG.

[0007] The speech recognition technology based on the DRNN method is as follows.
For example, a word registered in advance as a recognition target word (in this case, “good morning” or “weather”) from a continuous voice such as “good morning, good weather” is used as a keyword, and these keywords are input. This is to obtain a value indicating which part in the sound is present and with certainty, and to understand the continuous sound as described above based on the value indicating the certainty.

For example, it is assumed that when the input speaker utters "Good morning, good weather", an audio signal as shown in FIG. 7A is output. With respect to such an audio signal, an output as shown in FIG. 3B is obtained in an audio signal portion of “good morning”, and an output as in FIG. 2C is obtained in an audio signal portion of “weather”. Can be FIG.
In (c), a numerical value such as 0.9 or 0.8 is a numerical value indicating a certainty (degree of approximation), and a high numerical value such as 0.9 or 0.8 indicates that the word exists in the input voice with high certainty. be able to. That is, the registered word "good morning" is, as shown in FIG.
The registered word "weather" exists in the w1 portion on the time axis of the input voice signal at the w2 portion on the time axis of the input voice signal as shown in FIG. It can be seen that it exists with a certainty of 0.8.

In this manner, a DRNN word model is created for each word to be recognized, and the DRNN word model is created.
Recognize input speech using a word model.

When a DRNN word model is created for a certain word, two words, that is, the word to be recognized and the other words are uttered in succession and learned.

For example, as shown in FIG. 8, a continuous speech signal of two words (the word to be recognized is word 1 and the other is word 2) is output, while the speech data of word 1 is output. The output is turned off for the voice data of the word 2 that follows, and the output is turned off for the voice data of the word 2 (not shown). Learning to start the output is performed for the audio data of the following word 1.

[0012]

In this way, the DRN
The N word model is learned, but the problem in the speech recognition processing using the DRNN word model created by learning in this way is that words similar to the recognition target word (similar words) are uttered. That is, even if the word is not a recognition target word, a DRNN output having certainty or more certainty is output.

[0013] When learning a DRNN word model, as described above, learning is performed by continuously providing speech data of two words, but at this time, learning is not performed using words having similar pronunciations. Is customary. The reason is, for example, when “what time” and “how many times” are considered as words with similar pronunciations (similar words), when creating a voice model for “Nanji”, the voice data of “Nanji” And "Nando" voice data are given in succession, the output is started for "Niji" voice data, and the learning that "Nando" with similar pronunciation is not started is learned by both. This is because learning is inconsistent in the part of the same phoneme sequence, “what”.

Therefore, in the case of the DRNN word model trained on “Nanji” as a recognition target word, when the input speaker utters “Nando”, a DRNN output equivalent to the case of uttering “Nanji” is output. In many cases, in this case, it is recognized as "what."

[0015] In addition, for example, in addition to the word "Nanji" which has been learned in advance and registered as a recognition target word, the user requests that the word "Nando" be recognizable. There are cases. As described above, when the user requests that the similar words can be surely recognized, it is necessary to be able to respond by simple processing.

Therefore, the present invention utilizes a current speech model without changing the current learning method of a DRNN speech model, and enables simple and highly accurate recognition of similar words. An object is to enable similar words to be identified.

[0017]

In order to achieve the above-mentioned object, according to the first aspect of the present invention, there is provided a similar word identification method according to the present invention, wherein a predetermined output representing certainty in response to voice data of a certain word is provided. It has a speech model that has been learned so that it can be obtained.When the output for an input word is extracted using this speech model and recognition processing is performed based on the output, identification of a similar word that may be erroneously recognized is performed. In the similar word identification method to be performed, when a voice of a certain word is input, when an output representing certainty or more certainty is output by a voice model that reacts to voice data of the input word, a characteristic of the input word is included in the output. A predetermined section including a portion is set, the characteristics of the voice data of the input word in the predetermined section are examined, and based on the result, the input word and a unit similar to the input word are determined. It is characterized by performing the identification of the.

Further, the invention according to claim 2 has a DRNN speech model that has been trained so as to obtain a predetermined output representing certainty in response to speech data of a certain word.
When a DRNN output for an input word is extracted using an RNN speech model and recognition processing is performed based on the output, in a similar word identification method for identifying a similar word that may be erroneously recognized, a speech of a certain word is When input, a DRNN output corresponding to the input word voice data is output using a DRNN voice model, and the input word voice data is converted into code data using a codebook,
DR indicating certainty or more certainty for the input word
When an NN output is issued, a predetermined section including the characteristic portion of the input word is set in the DRNN output, the code data is checked in the set predetermined section, and based on the result, the input word and its It is characterized in that it is distinguished from words similar to the input word.

According to a third aspect of the present invention, in the second aspect of the present invention, code data corresponding to a vowel is examined among the code data in the set predetermined section, and an input word and a vowel are determined according to which vowel is. A word similar to the input word is identified.

According to a fourth aspect of the present invention, in the second or third aspect, the codebook is a codebook generated from five vowels.

[0021] Further, the invention according to claim 5 is the invention according to claim 2,
In the invention of 3 or 4, the DRNN speech model is made to correspond to similar word groups grouped for each type of similar word, and in each similar word group, a certain level of certainty is set for all the words in the group. It is characterized by using a speech model that has been trained so as to obtain a DRNN output representing likeness.

As described above, according to the present invention, as one method of identifying similar words, input speech data is vector-quantized using a codebook, code data is obtained, and DRN is performed.
By examining what vowels exist in a predetermined section of N outputs, an input word is identified. This allows
For similar words that cannot be identified only by DRNN output,
Similar words can be identified with high accuracy by using the existing DRNN model without changing the method of learning the DRNN speech model. Further, by setting the codebook to a codebook generated from five vowels, the processing can be greatly simplified.

The invention according to claim 6 has a DRNN speech model that has been trained to obtain a predetermined output representing certainty in response to speech data of a certain word.
When a DRNN output for an input word is extracted using an RNN speech model and recognition processing is performed based on the output, a similar word identification method for identifying a similar word that may be erroneously recognized may cause a false recognition. For each similar word having sex, speech data obtained by a plurality of speakers speaking for each word is coded using a predetermined codebook, and each coded word is used for each word using the coded code data. , Generate a code histogram in a predetermined section including the characteristic portion of the word, store the histogram data for each word as standard histogram data, and, when speech of a certain word is input, DRNN
A DRNN output corresponding to the input word voice data is output using a voice model, the input word voice data is converted into code data using a predetermined codebook, and a DRNN representing certainty or more certainty for the input word is provided.
When an output is issued, a predetermined section including the characteristic portion of the input word is set in the DRNN output, a code histogram in the predetermined section is generated using the code data, and the histogram data and the standard The method is characterized in that the input word and words similar to the input word are identified by comparing the input word with histogram data.

According to a seventh aspect of the present invention, in the sixth aspect, after each histogram is normalized, a difference between the two is obtained, and an input word and a word similar to the input word are determined based on the magnitude of the difference. It is characterized by performing identification.

As described above, the method of identifying similar words from the histogram data created based on the input word data and the standard histogram data also provides the same method as described above.
DRN for similar words that cannot be identified only with N output
Existing DR without changing the learning method of N voice model
Using the NN model as it is, it is possible to identify similar words with high accuracy. Furthermore, in the present invention, the frequency distribution of the code data in the predetermined section including the characteristic portion of the word is compared with the frequency distribution of the standard speaker, so that it is possible to perform the classification with higher accuracy, and it is extremely easy to be erroneously recognized. Similar words can be identified with high accuracy.

The invention of claim 8 is the invention according to claim 6 or 7, wherein the DRNN speech model is made to correspond to similar word groups grouped for each type of similar word, and in each similar word group, , A speech model trained so as to obtain a DRNN output representing certainty or more certainty for all the words in the group.

Further, the invention of claim 9 has a DRNN speech model that has been trained so as to obtain a predetermined output representing certainty in response to speech data of a certain word, and uses this DRNN speech model. DRNN for input word
When extracting output and performing recognition processing based on the output,
In a similar word identification method for identifying a similar word that is likely to be erroneously recognized, a DRNN output indicating certainty or more certainty is obtained for each characteristic portion of a similar word that may be erroneously recognized. DRN learned as
If an N-sub-speech model is created and a voice of a certain word is input, and a DRNN output indicating certainty or more certainty is output for the input word, the DRNN is output.
A predetermined section including the characteristic portion of the word is set in the output, the DRNN output state by the DRNN sub-speech model is checked in the predetermined section, and based on the result, an input word and a word similar to the input word are identified. It is characterized by performing.

According to a tenth aspect of the present invention, in the ninth aspect of the present invention, an input word and a word similar to the input word are determined based on which DRNN sub-speech model has a value indicating certainty or more in the DRNN output. Is characterized.

As described above, even if a similar word is identified by using the DRNN sub-speech model, or if a similar word that cannot be identified only by the DRNN output is used, the existing DR is used.
Similar words can be reliably identified without changing the learning method of the NN speech model. In addition, the present invention
Since the determination is made from the DRNN output using the DRNN voice model of only the characteristic portion between the similar words, the identification can be performed with higher accuracy, and the similar words that are easily misrecognized can be identified with high accuracy.

[0030] Further, according to the invention of claim 11, in the invention of claim 9 or 10, the DRNN speech model corresponds to a similar word group grouped for each type of similar word, and In the group, the speech model is characterized in that it is a speech model that has been trained so as to obtain a DRNN output representing certainty or more certainty for all the words in the group.

As described above, three methods have been roughly described here, but all of these methods are easy to process, and similar words can be reliably identified without changing the existing DRNN speech model learning method. can do. Further, the DRNN speech model corresponds to a similar word group grouped for each type of similar word, and in each similar word group, a DRNN output representing a certain degree or more certainty of all the words in the group is output. When the similar words are used as the recognition target words, there is no need to create a speech model for each similar word, and the cost is excellent. It will be. Further, among the similar word identification devices of the present invention, the invention of claim 12 has a speech model that has been learned so as to obtain a predetermined output representing certainty in response to speech data of a certain word. When an output for an input word is extracted using a speech model and subjected to recognition processing based on the output, a similar word identification device that identifies a similar word that may be erroneously recognized receives a speech of a certain word. A word detection signal output means for outputting an output representing certainty or more by a voice model responding to the voice data of the input word, and an output representing certainty or more certainty from the word detection signal output means. In this case, a predetermined section including a characteristic portion of the input word is set in the output, the characteristics of the voice data of the input word in the predetermined section are checked, and based on the result,
It is characterized by having recognition processing means for identifying an input word and a word similar to the input word.

The invention according to claim 13 has a DRNN speech model that has been trained so as to obtain a predetermined output representing certainty in response to speech data of a certain word, and uses this DRNN speech model. DRNN for input word
When extracting output and performing recognition processing based on the output,
In a similar word identification device that identifies a similar word that may be erroneously recognized, when a voice of a certain word is input, a word that outputs a DRNN output corresponding to the input word voice data using a DRNN voice model Detection signal output means, coding means for converting the input word voice data into code data using a codebook, and when a DRNN output indicating certainty or more certainty is output from the word detection signal output means, In the output, a predetermined section including the characteristic portion of the input word is set, and in the set predetermined section, code data is coded by the coding means, and code data is checked. And a recognition processing means for identifying a word to be processed.

According to a fourteenth aspect of the present invention, in the thirteenth aspect, code data corresponding to a vowel is examined from among the code data in the set predetermined section, and an input word and an input word are determined according to which vowel. It is characterized by identifying a similar word.

According to a fifteenth aspect, in the thirteenth or fourteenth aspect, the codebook is a codebook generated from five vowels.

The invention according to claim 16 is based on claim 13,
In the invention of the fourteenth or fifteenth aspect, the DRNN speech model corresponds to a similar word group grouped for each type of the similar word, and in each similar word group, a certain level of certainty for all the words in the group is determined. It is a speech model that has been trained so as to obtain a DRNN output representing likeness.

As described above, the similar word identification apparatus of the present invention uses the codebook to vector quantize the input speech data, obtains the code data, and determines what vowels exist in a predetermined section of the DRNN output. It examines and identifies the input words. As a result, similar words that cannot be identified only by the DRNN output can be identified with high accuracy by using the existing DRNN model without changing the learning method of the DRNN speech model. Further, by setting the codebook to a codebook generated from five vowels, the processing can be greatly simplified.

Further, the invention of claim 17 has a DRNN speech model that has been trained so as to obtain a predetermined output representing certainty in response to speech data of a certain word, and uses this DRNN speech model. DRNN for input word
When extracting output and performing recognition processing based on the output,
In a similar word identification device for identifying a similar word that may be misrecognized, speech data obtained by a plurality of speakers uttering each of the similar words that may be misrecognized is encoded. A standard histogram data storage unit that stores a code histogram created using code data in a predetermined section including a characteristic portion of the word in the code data as standard histogram data, and when a voice of a certain word is input, DRN corresponding to the input word voice data using the DRNN voice model
A word detection signal output unit that outputs N outputs, a coding unit that converts the input word voice data into code data using a predetermined codebook, and a DRNN output that indicates certainty or more certainty from the word detection signal output unit. When the code is output, a predetermined section including the characteristic portion of the input word is set in the DRNN output, and a code histogram in the predetermined section is generated using the code data coded by the coding means. It is characterized by having recognition processing means for comparing the histogram data with the standard histogram data to identify the input word and a word similar to the input word.

The invention of claim 18 is the invention according to claim 17 in which, after normalizing each histogram, a difference between the two is obtained, and the difference between the input word and the word similar to the input word is determined based on the magnitude of the difference. It is characterized by performing identification.

As described above, the similar word identifying apparatus for identifying similar words from the histogram data created based on the input word data and the standard histogram data has the same structure as that of the above.
For similar words that cannot be identified only by RNN output,
Without changing the learning method of the RNN speech model, similar words can be identified with high accuracy using the existing DRNN model as it is. Furthermore, in the present invention, the frequency distribution of the code data in the predetermined section including the characteristic portion of the word is compared with the frequency distribution of the standard speaker, so that it is possible to perform the classification with higher accuracy, and it is extremely easy to be erroneously recognized. Similar words can be identified with high accuracy.

The invention of claim 19 is the invention of claim 17 or 1
In the invention of the eighth aspect, the DRNN speech model is made to correspond to similar word groups grouped for each type of similar words, and in each similar word group, a certain degree of certainty or more for all the words in the group is set. It is characterized by using a speech model that has been trained so as to obtain a DRNN output.

Further, the invention of claim 20 has a DRNN speech model that has been trained so as to obtain a predetermined output representing certainty in response to speech data of a certain word, and using this DRNN speech model. DRNN for input word
When extracting output and performing recognition processing based on the output,
In a similar word identification device that identifies a similar word that may be erroneously recognized, a DRNN output that indicates a certain degree of certainty or more for each characteristic portion of a similar word that may be erroneously recognized is obtained. DRN learned as
DRNN sub-speech model storage means for storing N sub-speech models;
A DRNN output corresponding to the input word data is output using the NN speech model, and a DRN output corresponding to the characteristic portion of the input word using the DRNN sub-speech model.
A word detection signal output unit that outputs N outputs, and when a DRNN output indicating certainty or more certainty is output from the word detection signal output unit using the DRNN voice model, the characteristic portion of the word is output to the DRNN output. A recognition process for setting a predetermined section including the input word, examining a DRNN output state of the input word by the DRNN sub-speech model with respect to the input word, and discriminating the input word and a word similar to the input word based on the result. And a part.

According to a twenty-first aspect of the present invention, in the twentieth aspect, an input word and a word similar to the input word are determined based on which DRNN sub-speech model has a value representing certainty or more in the DRNN output. Is characterized.

As described above, according to the similar word identification device for identifying similar words using the DRNN sub-speech model, similar methods cannot be used for the similar words that cannot be identified only by the DRNN output, without changing the DRNN speech model learning method. Using the existing DRNN model as it is, it becomes possible to identify similar words with high accuracy. Further, since the present invention uses a DRNN speech model of only the characteristic portion between similar words to make a judgment from the DRNN output, it is possible to perform highly accurate discrimination, and to perform highly accurate similar words that are easily misrecognized. Can be identified.

The invention of claim 22 is the invention of claim 20 or 21, wherein the DRNN voice model is
Corresponding to similar word groups grouped for each type of similar word, and in each similar word group,
It is a speech model that has been trained so as to obtain a DRNN output representing certainty or more certainty for all the words in the group.

As described above, here, three similar word identification devices are roughly divided and described. However, in any of these similar word identification devices, the configuration of the device is not significantly complicated, and existing similar word identification devices are not complicated. Similar words can be reliably identified by a simple process without changing the DRNN speech model learning method. Also, the DRNN speech model is:
Corresponding to similar word groups grouped for each type of similar word, and in each similar word group,
By using a speech model that has been trained so that a DRNN output representing a certain degree of certainty is obtained for all the words in the group, similar words are recognized one by one when similar words are used as recognition target words. On the other hand, there is no need to create a voice model, and the cost is excellent.

[0046]

Embodiments of the present invention will be described below with reference to the drawings. In the embodiment described below, an example will be described in which “what” and “how” are used as similar words that are likely to be erroneously recognized, and each of these words is recognized.

(First Embodiment) FIG. 1 is a block diagram showing a word identification device for realizing the first embodiment. A microphone 1 as an audio input unit, an A / D conversion unit 2, Voice analysis unit 3, word detection signal output unit 4, DRNN
It comprises an output information storage unit 5, a vector quantization unit 6, a code book 7, a code data storage unit 8, a recognition processing unit 9, and the like.

The voice input from the microphone 1 is A / D-converted by the A / D converter 2,
Is converted into a feature vector sequence represented by, for example, a 10-dimensional LPC cepstrum coefficient.

The word detection signal output section 4 comprises a DRNN word Dell storage section 41 and a word spotting section 42. The DRNN word model storage unit 41 stores the DRNN word model data for each recognition target word. As the recognition target words, for example, “Good morning”,
It is assumed that the word is "good night" or "Nanji".

When the speech data from the speech analysis unit 3 is input, the word spotting unit 42 uses the contents of the DRNN word model storage unit 41 to output a DRNN for the keyword (recognition target word) (start time of the DRNN output). , End time, output value indicating certainty)
Get. Each of these data is stored in the DRNN output information storage unit 5.

In this embodiment, as described above, "Nanji" and "Nando" are exemplified as similar words. Then, it has a word model learned for the word "Nanji", and an output with high certainty comes to the voice data "Nanji" spoken by the input speaker. I have. A similar output is also output for voice data "Nando" spoken by the input speaker.

The code book 7 is a code book having a code size of 5 created from five vowels "A / I / U / E / O".

The vector quantizing unit 6 quantizes the feature vector sequence of the input speech subjected to the speech analysis using the codebook 7 to generate code data.
The code data is stored in the code data storage unit 8.

Specific processing in such a configuration will be described.

As described above, the DRNN word model storage section 41 stores a DRNN word model for each recognition target word. Therefore, if the input speaker speaks “Nanji”, the word spotting section 42 outputs a DRNN output having a certain degree of certainty (DRNN output based on the voice model of “Nanji”), and outputs the DRNN output. Output values representing start time, end time, and certainty are detected. These data are stored in the DRNN output information storage unit 5.

In the description of the first embodiment and the subsequent second and third embodiments, when there is an expression such as a DRNN output from the word spotting section 42, "Nanji. DR based on voice model
It means NN output. On the other hand, when the input speaker utters “what”, similarly, a DRNN output having a certain degree of certainty is output from the word spotting unit 42, and the start time, the end time, and the certainty of the DRNN output are similarly determined. Then, data such as a start time, an end time, and an output value indicating certainty of the DRNN output are stored in the DRNN output information storage unit 5.

FIG. 2 (a) shows the DRNN output when the input speaker utters "I'm doing now" as utterance 1, and FIG. This shows the DRNN output when "Kana" is spoken. In this way, a DRNN output having a certain degree of certainty in both the “what” and “what” portions is output.

Then, when the input speaker utters "Now, what is it?", The voice data (feature vector sequence)
It is also given to the vector quantization unit 6. The vector quantization unit 6 uses the codebook 7 created from the five vowels to perform vector quantization on the voice data of the input speaker “now what it is”.

That is, the codebook 7 contains five code vectors corresponding to the respective vowels of “A / I / U / E / O”, ie, c0 and “I” for “A”. C1 for "U", c3 for "E", and c4 for "O". These code vectors "c0, c1,..., C4" Then, a distance between each of the feature vectors constituting the feature vector sequence obtained by the utterance of the input speaker is calculated, and is correlated with a code vector of the shortest distance to code and obtain code data. FIG. 2C shows an example of the coded code string. As can be seen from FIG. 2 (c), for example, in the phoneme sequence of "I'm now alive", "I" is the vowel "I" phoneme itself, and "Ji" Since there is a vowel "i" phoneme, its chord data has a large amount of c1, and "ma", "na", "ka",
Since "na" has a vowel "a" phoneme, its code data is data having a large c0.

The code data thus coded is stored in the code data storage unit 8. Based on the code data stored in the code data storage unit 8 and the DRNN output data stored in the DRNN output information storage unit 5, the recognition processing unit 9 determines the word spoken by the input speaker as “what Judgment ”or“ what ”. This determination processing will be described below.

As described above, the data stored in the DRNN output information storage unit 5 is the output time indicating the start time, end time, and certainty of the DRNN output. Of the DRNN output (see FIG. 2 (a)) corresponding to "."
In this case, since the section t1 is a discrimination between “Nanji” and “Nando”, an interval that sufficiently includes the DRNN output of the phoneme part of “Ji” or “Doh”, which seems to be the most different between the two. Set. That is, a section including a DRNN output for a characteristic portion (in this case, a portion of “ji” or “do”) of an input similar word is set.

Then, what kind of code vector mainly constitutes the code data corresponding to the section t1 in the code data string shown in FIG. 2C is examined. In this case, the code vector in the section t1 is c2, c
1, c1 and c1 exist.

As described above, since the code vector of "i" exists in the section t1, the code vector D shown in FIG.
It is determined that the RNN output is a DRNN output for “what”.

If the input speaker utters "Now, what is it?", The DRNN output is almost the same as "Nanji". Is encoded by the vector quantization unit 6 using the codebook 7, and a code sequence as shown in FIG. 2D is obtained. As can be seen from FIG. 2D, in this case, the code vectors in the section t1 are c2 and c
4, c4, and c4.

As described above, since the code vector of “O” exists in the section t1, the code vector D shown in FIG.
It is determined that the RNN output is a DRNN output for “what”.

As described above, in the first embodiment,
As a method of identifying similar words, input speech data is vector-quantized using a codebook created from five vowels, code data is obtained, and what vowels are present in a predetermined section of the DRNN output is examined. , Input words are identified. As a result, words that cannot be identified only by the DRNN output can be identified with high accuracy.

In the first embodiment, an example has been described in which the code book 7 uses five vowels and has five codes. However, the code book 7 has five codes. The present invention is not limited to this, and for example, a code book having a larger number of codes created from voice including consonants may be used. For example, if you use a codebook that includes all phoneme features,
Taking “Nanji” and “Nando” as an example, the voice data of “Ji” or “Do” obtained by uttering the input speaker is vector-quantized using a codebook, and “Nanji” Alternatively, by determining whether or not there is code data corresponding to “ji” or “do” in the set section t1 of the DRNN output of “what”, the word spoken by the input speaker is determined as
It is possible to determine whether it is "what" or "what". However, in the first embodiment,
Using a codebook created from five vowels is more advantageous in terms of processing amount.

By using the method described in the first embodiment described above, "Nando" and "Nanji" are actually uttered with a total of about two hundred and dozens of adult men and adult women. As a result of performing an experiment for obtaining a recognition rate, a recognition rate close to 95% was obtained. The codebook used here is a gender-specific codebook (a codebook of size = 5 created from discrete vowel utterance data). This experimental result also indicates that similar words can be identified with extremely high accuracy.

(Second Embodiment) Next, a second embodiment of the present invention will be described. FIG. 3 is a configuration diagram of a similar word identification device for realizing the second embodiment. The difference from the configuration of the first embodiment is that a standard histogram storage unit 11 and a histogram generation unit 12 are provided. On the point. The other components are almost the same as those in FIG. 1, and the same portions are denoted by the same reference numerals. However, this second
As the codebook 7 used in the embodiment, a codebook including characteristics of all phonemes is used. In the second embodiment, an example using a codebook having 64 codes will be described.

The standard histogram storage section 11 stores standard histogram data. This standard histogram data is obtained, for example, by performing vector quantization on voice data obtained by uttering each of several hundred speakers for the word “Nanji” using a codebook of 64 sizes, This is histogram data indicating which code vector of the code book has appeared and how many times. The standard histogram data is created in advance for each similar word that is easily misrecognized.

FIG. 4 (a) shows the results obtained when one of the hundreds of speakers uttered “Nanji” and obtained vector data using a 64-size codebook. A code data sequence is obtained for each of the code vectors, and the number of appearances is accumulated for each of the code vectors c0 to c63 to create a standard histogram.

Then, as described in the first embodiment, in the DRNN output (see FIG. 2A) output in response to the voice data of "what", a certain section Set t1. Then, among the created standard histograms, the standard histogram data in the section t1 is stored in the standard histogram storage unit 11. FIG. 4B shows an example of a standard histogram stored in the standard histogram storage unit 11.

That is, the standard histogram shown in FIG. 4B is the cumulative number of code vectors obtained from hundreds of speakers in the phoneme portion corresponding to the section t1 of the DRNN output for “Nanji”. Is represented.

Similarly, a standard histogram of code vectors obtained from hundreds of speakers in the phoneme portion corresponding to the interval t1 is created for the word “Nando” similar to “Nanji”. Keep it.

As described above, standard histograms of similar words (here, “Nanji” and “Nando”) in the predetermined section t1 of the DRNN output are created in advance, and stored in the standard histogram storage unit 11. Remember.

Then, when the user utters “Now, what is it?”, The voice data (feature vector sequence) is supplied to the word spotting section 42 and also to the vector quantization section 6. This vector quantization unit 6
Now, using a codebook 7 with a code size of 64,
The code data is obtained by vector-quantizing the voice data of the user "now what it is". This code data is stored in the code data storage unit 8.

When the recognition processing section 9 determines that the histogram generation processing is necessary, it generates a histogram of the input speaker in the section t1 obtained from the DRNN output information. The determination as to whether or not the histogram generation processing is necessary is performed by checking the contents of the DRNN output information storage unit 5 and outputting a value indicating a certain degree of certainty with respect to “Nanji” spoken by the user. In this case, it is determined that histogram generation processing is necessary.

The process of generating the histogram of the input speaker in the section t1 is performed in the section t1 of the DRNN output of the code data corresponding to the user's “now okana” voice data stored in the code data storage section 8. Is generated. FIG. 4 shows an example of the generated histogram.
It is shown in (c). Then, the distance between the histogram of the input speaker and the standard histogram is obtained. The standard histogram is a histogram obtained from several hundred people, and the histogram of the input speaker is a histogram obtained from one voice data. Are normalized to obtain a distance. This normalization processing is not particularly limited.

A difference histogram between the normalized input speaker histogram and the standard histogram for “Nanji” is obtained, and the input speaker histogram and the “Nando” are calculated.
Of the standard histogram with respect to. FIG. 4D shows a difference histogram (absolute value) between the input speaker histogram and the standard histogram for “Nanji”.

The sum of the cumulative frequencies in the difference histogram (absolute value) thus obtained is obtained.

The above processing is also performed on the input speaker histogram and the standard histogram for "Nando", a difference histogram between the two is obtained, and the cumulative frequency of the difference histogram is added to obtain the sum.

Then, the respective total values are compared and the smaller total value is selected. For example, the sum of the cumulative frequencies in the difference histogram (absolute value) obtained from the input speaker histogram and the standard histogram for “Nanji” was obtained from the input speaker histogram and the standard histogram for “Nanji”. If the sum is smaller than the total value in the difference histogram (absolute value), it is determined that the word spoken by the input speaker is “Nanji”.

As described above, in the second embodiment, each voice data obtained by uttering hundreds of people about similar words is coded, and the similar word described above is coded based on the code data. A difference histogram between the standard histogram and the input speaker histogram is obtained in advance, and an input word is identified from the difference histogram. According to the second embodiment, as in the first embodiment, the DRNN is used.
Existing DR for words that cannot be identified by output alone
Similar words can be reliably identified without changing the learning method of the NN speech model.

In the method described in the second embodiment, the frequency distribution of code data in a predetermined section including a characteristic portion of a word is compared with the frequency distribution of a standard speaker. Accurate identification is possible, and it becomes possible to identify highly similar words that are easily misrecognized with high accuracy.

Using the method described in the second embodiment described above, an experiment was conducted on "Nando" and "Nanji" with a total of more than 200 dozen speakers of adult men and adult women. , A recognition rate close to 100% was obtained. The codebook used here is a male codebook with a code size of 256, and standard histograms were created for male, female, and unisex. Similar high recognition rates were obtained.

(Third Embodiment) Next, a third embodiment of the present invention will be described. FIG. 5 is a configuration diagram of a similar identification device for realizing the third embodiment, and includes a microphone 1 as an audio input unit, an A / D conversion unit 2, an audio analysis unit 3, a word detection signal output unit 4, It comprises a DRNN output information storage unit 5, a recognition processing unit 9, a sub-word DRNN output information storage unit 13, and the like.

The word detection signal output unit 4 according to the third embodiment uses the DR used in the first and second embodiments.
NN word model storage unit 41, word spotting unit 4
2, a DRNN sub-word model storage unit 43 is provided. The DRNN sub-word model storage unit 43 stores the DRNN word model data of the characteristic parts “ji” and “do” in “nanji” and “nando” as similar words.

The sub-word DRNN output information storage unit 1
3 is a DRN output when word spotting processing is performed using the DRNN sub-word model storage unit 43.
It stores the start time and end time of N output, data representing certainty, and the like. In this embodiment, the DRNN output information storage unit 5 stores the sub-word DRN
DR of input word itself to N output information storage unit 13
Since the NN output information is stored, the input word DRNN output information storage unit 5 will be described below in order to distinguish between them.
That.

The processing of the third embodiment will be described below. Also in the third embodiment, similar words “Nanji” and “Nando” will be described as examples.

Now, assuming that the input speaker has uttered, "Now what is it?", The voice is analyzed by the voice analysis unit 3, output as a feature vector sequence, and given to the word spotting unit 42. As a result, word spotting processing is performed using the contents of the DRNN word model storage unit 41, and the DRNN output from the word spotting unit 42 having a certain degree of certainty or more as shown in FIG. Is output, and data such as a start time, an end time, and an output value representing certainty of the DRNN output are stored in the input word DRNN output information storage unit 5.

As described above, the user is asked "what is it now?"
DRNN is output in the part of “Noji”, but with this, the DRNN is output in the phoneme part of “Jiji”.
The word spotting process is performed using the contents of the sub-word model storage unit 43, and the word spotting unit 42 outputs a DRNN output having a certain degree of certainty or more as shown in FIG. . Then, the output value indicating the start time, the end time, and the certainty of the DRNN output in the phoneme portion of the “ji” is stored in the sub-word DRNN output information storage unit 13.

Then, when there is a DRNN output having a certain degree of certainty with respect to "Nanji", the recognition processing unit 9, as described in the first and second embodiments, outputs the input word DRNN. Based on the time information stored in the output information storage unit 5, a section t1 is set, the sub-word DRNN output in the section t1 is checked, and based on the result, which word is the input word is determined. That is, the recognition processing unit 9 recognizes the input word based on the contents of the input word DRNN output information storage unit 5 and the sub-word DRNN output information storage unit 13. Specific processing is performed as follows.

The sub audio DRNN output information storage unit 13
In this case, data such as a start time and an end time of the DRNN output for the “ji” and an output value indicating the likelihood are stored.

Therefore, the recognition processing section 9 sets the “Nanji”
If there is a DRNN output that has a certain degree of certainty with respect to and a subword DRNN output that has a certain degree of certainty exists in the section t1, it is determined that the input speech is "what."

On the other hand, when the input speaker utters "now", the word spotting section 42 outputs a DRNN output having a certain degree of certainty or more as shown in FIG. . And that DRN
Output values representing the start time, end time, and certainty of the N output are detected, and each data is stored in the input word DRNN output information storage unit 5.

As described above, the user is asked "what is it now?"
DRNN is output at the part of "do", but the word phoneting processing is performed on the phoneme part of "do" using the DRNN sub-word model, and the part of "do" is As a result, a DRNN output having a certain degree of certainty as shown in FIG. Then, the DRNN in this "do" phoneme part
Output values indicating the output start time, end time, and certainty are detected, and each detected data is stored in the sub-word DRNN output information storage unit 13.

Therefore, the recognition processing section 9 sets “Nando”
If there is a DRNN output having a certain degree of certainty with respect to and a sub-word DRNN output having a certain degree of certainty exists in the section t1, it is determined that the input speech is "what".

By the way, in the third embodiment,
It has a DRNN word model of one phoneme such as “ji” or “do”. In actual recognition, one phonological DRNN word model often outputs various voices other than “ji” and “do”, and the output is often 1
The DRNN word model for recognizing one phoneme itself has a problem at this stage. However, if it is used as a sub-word model as in the third embodiment, practically sufficient data can be obtained.

Also, since “Nanji” and “Nando” have been described here, the DRNN sub-word model is “Ji”.
And a “Don” corresponding to a single phoneme DRNN sub-word model.
There may be a case where a sub-word model having two or more phonemes is used instead of the N sub-word model.

For example, assuming that “Nanji” and “Nanichi (what day)” are similar words, in this case, the DNN of “Ji” is used as a DRNN sub-word model for “Nanji”.
An RNN sub-word model is prepared. For “what”, for example, a DRNN sub-word model of “ni” is prepared. In this way, various similar words can be handled.

According to the third embodiment described above, similarly to the first and second embodiments, the existing DRNN word model learning method is changed for words that cannot be identified only by the DRNN output. Similar words can be reliably identified without the need for such a word. Further, in the third embodiment, similar words are identified by the DRNN output for the phonemes themselves of different parts, so that further similarity is achieved.
Highly accurate identification becomes possible.

As described above, in speech recognition using the DRNN word model, when similar words such as “Nanji” and “Nando” are input, a DRNN output is output for both voices. However, according to the first to third embodiments, simple processing is added to such similar words without changing the learning method of the DRNN word model possessed by the recognition device. Thus, similar words can be identified with high accuracy.

In the present invention, similar words can be identified other than in the first to third embodiments. For example, start / end free D that allows keyword spotting
Identification can also be performed using a general speech recognition technique such as a P matching method, an HMM (Hidden Markov) method, or a neural network method. A brief description will be given below of the case where the start-end free DP matching method is used. Also here, "Nanji" and "Nando" shall be identified.

The standard speaker characteristic data of “Nanji” and “Nando” are prepared, and a portion corresponding to “Nanji” output by, for example, “Now now kana” spoken by the input speaker. T1 to the DRNN output of
Is set, and in the section t1, a distance is obtained by performing DP matching between the input voice characteristic data and the standard speaker characteristic data, and an input word is identified from the distance. This identification method requires a larger amount of processing than the first to third embodiments described above, but it is sufficiently possible to realize the present invention also by this method.

Further, in each of the above-described embodiments, “Nanji” and “Nando” have been described as examples of similar words that are easily misrecognized. However, similar words are not limited to these. The present invention can also identify other similar words.

Also, for example, assuming “Nanji” and “Nando” as a set of similar words that are easily misrecognized,
Assuming that the word model originally possessed by the speech recognition device is the word model of “Nanji”, if “Nado” reacts to this word model and a DRNN output exceeding a certain level is output, the above-mentioned respective implementations are performed. By performing the processing described in the form, the original recognition target word "Nanji"
Can be identified with high accuracy,
It is also possible to add “Nando” as a new recognition target word to this speech recognition device. The same can be said for other sets of similar words.

Further, when implementing the present invention, the word model possessed by the recognition device may be learned from the beginning so that a DRNN output equal to or greater than a certain value is output for all similar words.

For example, if “Nanji” and “Nando” are taken as examples of similar words, a speech model that produces sufficient DRNN output for “Nanji” and “Nando” is created in advance. The DRNN output is actively output regardless of which of these similar words is input, and thereafter, the identification processing described in each of the above embodiments is performed to determine which word is input. You may.

In this way, when there are a plurality of similar words that are likely to be erroneously recognized among the words to be recognized, it is not necessary to have a speech model for each of the similar words, which is advantageous in terms of cost.

Further, in this connection, a word model in which a group of similar words is created and a DRNN output having a certain degree of certainty is output for any word in the group is generated for each group. You may make it create.

For example, the similar word A group includes “Nanji” and “Nando”, and the similar word B group includes “Dengon (message)”, “Denwa (telephone)”, and “Denki (electric)”. Words are grouped by words that are likely to be misrecognized. Then, a DRNN word model A trained so that an output having a certain degree of certainty is output for any word in the similar word A group is created,
Further, D is learned so that an output having a certain degree of certainty is output for any word in the similar word B group.
A word model corresponding to each group is created, such as creating an RNN word model B. Similarly, a word model corresponding to each group is created for other similar word groups.

In this way, a group of similar words is created, and a certain or more D
By having a word model corresponding to each group so that an RNN output is output, for example, when an input "Nanji" is input, the corresponding word model is activated to provide a DRNN output having a certain degree of certainty or more. After that, if the processing described in each of the above-described embodiments is performed, it can be determined that the input voice is “what”.

In this way, even if there are many similar words which are likely to be erroneously recognized among the words to be recognized, it is not necessary to have a word model for each similar word, which is advantageous in terms of cost.

Further, in the present invention, the similar word identification processing described in each of the above embodiments may be used in combination.

The processing program for performing the processing of the present invention can be stored in a storage medium such as a floppy disk, an optical disk, or a hard disk. The present invention also includes such a storage medium. Alternatively, data may be obtained from a network.

[0116]

As described above, according to the present invention, D
For similar words that cannot be identified only by RNN output,
Without changing the learning method of the RNN word model, similar words can be identified with high accuracy using the existing DRNN word model as it is. As one method for realizing this, the input speech data is vector-quantized using a codebook, code data is obtained, and what vowels are present in a predetermined section of the DRNN output is checked. , It becomes possible to identify similar words with high accuracy by using the existing DRNN word model as it is without changing the method of learning the DRNN word model for similar words that cannot be identified only by the DRNN output. . Further, by making the codebook a codebook generated from five vowels, similar words can be identified with an extremely small amount of processing.

Also, by the method of identifying a similar word from the histogram data created based on the input word data and the standard histogram data, the learning of the DRNN word model is performed for the similar word which cannot be identified only by the DRNN output. Without changing the method, similar words can be identified with high accuracy using the existing DRNN word model as it is. In this way, by comparing the frequency distribution of the code data of the predetermined section including the characteristic part of the word with the frequency distribution of the standard speaker and identifying the similar words, it is possible to perform the identification with higher accuracy, and the Similar words that are easily misrecognized can also be identified with high accuracy.

Further, similar to the above, the DRN sub-word model is used to identify similar words.
Even for similar words that cannot be identified only by N outputs, similar words can be reliably identified without changing the learning method of the existing DRNN word model. In this way, by using the DRNN word model of only the characteristic part between similar words and identifying the different parts of the similar words by the DRNN output for the phonemes themselves,
Highly accurate identification becomes possible.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a similar word identification device according to a first embodiment of the present invention.

FIG. 2 is a diagram illustrating an identification process according to the first embodiment.

FIG. 3 is a configuration diagram of a similar word identification device according to a second embodiment of the present invention.

FIG. 4 is a diagram illustrating an identification process according to a second embodiment.

FIG. 5 is a configuration diagram of a similar word identification device according to a third embodiment of the present invention.

FIG. 6 is a diagram illustrating an identification process according to a third embodiment.

FIG. 7 is a view for explaining DRNN output when performing word spotting processing using a DRNN word model.

FIG. 8 is a view for explaining a process of learning two words in succession when learning a DRNN word model.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 microphone 2 A / D conversion unit 3 voice analysis unit 4 word detection signal analysis unit 5 DRNN output information storage unit 6 vector quantization unit 7 codebook 8 code data storage unit 9 recognition processing unit 11 standard histogram data storage unit 12 histogram generation Unit 13 Sub-word DRNN output information storage unit 41 DRNN word model storage unit 42 Word spotting unit 43 DRNN sub-word model storage unit

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Hiroo Hasegawa 3-5-5 Yamato, Suwa City, Nagano Prefecture Seiko Epson Corporation

Claims (22)

[Claims] 1. A speech model which has been trained so as to obtain a predetermined output representing certainty in response to speech data of a certain word, and uses the speech model to extract an output for an input word and output the output. In a similar word identification method for identifying a similar word that may be erroneously recognized when performing recognition processing based on a speech model, a speech model reacting to speech data of the input word when a speech of a certain word is input When an output representing certainty or more certainty is output, a predetermined section including a characteristic part of the input word is set in the output, the characteristics of the voice data of the input word in the predetermined section are checked, and based on the result, And identifying the input word and words similar to the input word. 2. A DR trained so as to obtain a predetermined output representing certainty in response to voice data of a certain word.
Similar word identification that has an NN speech model and uses this DRNN speech model to extract a DRNN output for an input word and perform recognition processing based on the output to identify a similar word that may be erroneously recognized In the method, when a voice of a certain word is input, a DRNN output corresponding to the input word voice data is output using a DRNN voice model, and the input word voice data is converted into code data using a codebook. DR that indicates certainty or more certainty for a word
When an NN output is issued, a predetermined section including the characteristic portion of the input word is set in the DRNN output, the code data is checked in the set predetermined section, and based on the result, the input word and its A similar word identification method characterized by identifying a word similar to an input word. 3. A process of examining the code data in the set predetermined section and identifying an input word and a word similar to the input word based on a result of the code data in the set predetermined section Of which
3. The similar word identification method according to claim 2, wherein code data corresponding to the vowel is examined, and an input word and a word similar to the input word are identified based on which vowel is used. 4. The code book according to claim 2, wherein the code book is a code book generated from five vowels.
Or the similar word identification method described in 3. 5. The DRNN speech model is adapted to correspond to similar word groups grouped for each type of similar word, and in each similar word group, a certain degree of certainty or more for all the words in the group is set. 5. The similar word identification method according to claim 2, wherein the speech model is a speech model learned so as to obtain a DRNN output. 6. A DR trained so as to obtain a predetermined output representing certainty in response to voice data of a certain word.
Similar word identification that has an NN speech model and uses this DRNN speech model to extract a DRNN output for an input word and perform recognition processing based on the output to identify a similar word that may be erroneously recognized In the method, for each similar word that may be misrecognized, a plurality of speakers utter the speech data for each word, and code the obtained speech data using a predetermined code book. Using the code data, a code histogram is created for each word in a predetermined section including the characteristic portion of the word, and the histogram data for each word is stored as standard histogram data. A DRNN output corresponding to the input word voice data is output using a voice model, and The audio data encoding data by using a predetermined codebook, DR representing at least a certain likelihood for the input word
When the NN output is issued, a predetermined section including the characteristic portion of the input word is set in the DRNN output, and a code histogram in the predetermined section is created using the code data. A similar word identification method comprising comparing the input word and a word similar to the input word by comparing the input word with standard histogram data. 7. A process of comparing the histogram data created from the input word and the standard histogram data to identify the input word and a word similar to the input word, includes: 7. The similar word identification method according to claim 6, wherein a difference between the two words is obtained, and an input word and a word similar to the input word are identified based on the magnitude of the difference. , 8. The DRNN speech model associates similar word groups grouped for each type of similar word, and in each similar word group, a certain degree of certainty or more for all the words in the group. 8. The similar word identification method according to claim 6, wherein the speech model is a speech model trained so as to obtain a DRNN output. 9. A DR trained so as to obtain a predetermined output representing certainty in response to voice data of a certain word.
Similar word identification that has an NN speech model and uses this DRNN speech model to extract a DRNN output for an input word and perform recognition processing based on the output to identify a similar word that may be erroneously recognized In the method, a DRNN sub-speech model trained so as to obtain a DRNN output representing certainty or more certainty is obtained for each characteristic portion of similar words that may be erroneously recognized, and a certain word When a voice is input, when a DRNN output indicating certainty or more certainty is output for the input word, a predetermined section including the characteristic portion of the word is set in the DRNN output, and within the predetermined section, The DRNN
A similar word identification method characterized by examining a DRNN output state by a sub-speech model and identifying an input word and a word similar to the input word based on the result. 10. The DRNN in the predetermined section.
The process of examining the DRNN output state by the sub-speech model and discriminating an input word and a word similar to the input word based on the result is a value indicating the certainty that the DRNN output by the DRNN sub-speech model is equal to or more than a certain level. 10. The similar word identification method according to claim 9, wherein an input word and a word similar to the input word are identified depending on whether the input word is set to or not. 11. The DRNN speech model is adapted to correspond to similar word groups grouped for each type of similar word, and in each similar word group, a certain level of certainty is determined for all the words in the group. The similar word identification method according to claim 9 or 10, wherein the speech model is a speech model that has been learned so as to obtain a DRNN output. 12. A speech model that has been trained so as to obtain a predetermined output representing certainty in response to speech data of a certain word, and uses the speech model to extract an output for an input word and output the output. In a similar word identification device that identifies a similar word that may be erroneously recognized when performing recognition processing based on a voice model, a voice model that reacts to voice data of the input word when a voice of a certain word is input A word detection signal output unit that outputs an output representing certainty or more certainty, and when an output indicating a certainty or more certainty is output from the word detection signal output unit, the output includes a characteristic portion of the input word. A section is set, the characteristics of the voice data of the input word in the predetermined section are examined, and based on the result, the input word and a word similar to the input word are identified. Similar word identification device, characterized in that it comprises a recognition processing unit for performing a. 13. A D which has been learned so as to obtain a predetermined output representing certainty in response to voice data of a certain word.
Similar word identification that has an RNN speech model and uses the DRNN speech model to extract a DRNN output for an input word and perform recognition processing based on the output to identify a similar word that may be erroneously recognized. In the apparatus, when a voice of a certain word is input, a word detection signal output unit that outputs a DRNN output corresponding to the input word voice data using a DRNN voice model, and codes the input word voice data using a codebook. When a DRNN output representing certainty or more certainty is output from the encoding means for converting to data and the word detection signal output means, a predetermined section including the characteristic portion of the input word is set in the DRNN output, In the set predetermined section, code data is coded by the coding means, and code data is checked. Input word Te and the recognition processing means for identification of the word that is similar to the input word, similar word identification device characterized in that it comprises a. 14. A process for examining the code data in the set predetermined section and identifying an input word and a word similar to the input word based on a result of the examination, the code data in the set predetermined section Of which
14. The similar word identification apparatus according to claim 13, wherein code data corresponding to the vowel is examined, and an input word and a word similar to the input word are identified based on which vowel is used. 15. The codebook according to claim 1, wherein the codebook is a codebook generated from five vowels.
15. The similar word identification device according to 3 or 14. 16. The DRNN speech model is adapted to correspond to similar word groups grouped for each type of similar word, and in each similar word group, certain degree of certainty is given to all the words in the group. 16. A speech model that has been trained to obtain a DRNN output.
Description similar word identification device. 17. A D which is learned so as to obtain a predetermined output representing certainty in response to voice data of a certain word.
Similar word identification that has an RNN speech model and uses the DRNN speech model to extract a DRNN output for an input word and perform recognition processing based on the output to identify a similar word that may be erroneously recognized. In the apparatus, of code data obtained by coding speech data obtained by a plurality of speakers uttering each of similar words that may be erroneously recognized, a code in a predetermined section including a characteristic portion of each similar word Standard histogram storage means for storing, as standard histogram data, a code histogram for each similar word created using the data, and when a voice of a certain word is input, corresponding to the input word voice data using a DRNN voice model A word detection signal output means for outputting a DRNN output to perform A DRNN output indicating a certain degree of certainty is output from the word detection signal output means, and a predetermined section including the characteristic portion of the input word in the DRNN output. Setting and generating a code histogram in the predetermined section using the code data coded by the coding means, comparing the histogram data with the standard histogram data, and A recognition processing means for identifying a word similar to a word; and a similar word identification device, comprising: 18. A process of comparing the histogram data created from the input word and the standard histogram data to identify the input word and a word similar to the input word includes normalizing each histogram. 18. The similar word identification device according to claim 17, wherein after the conversion, the difference between the two is obtained, and the input word and a word similar to the input word are identified based on the magnitude of the difference. 19. The DRNN speech model is adapted to correspond to similar word groups grouped for each type of similar word, and in each similar word group, a certain degree of certainty or more for all the words in the group is set. 19. The similar word identification apparatus according to claim 17, wherein the speech model is a speech model learned so as to obtain a DRNN output. 20. D learned in response to voice data of a certain word to obtain a predetermined output representing certainty
Similar word identification that has an RNN speech model and uses the DRNN speech model to extract a DRNN output for an input word and perform recognition processing based on the output to identify a similar word that may be erroneously recognized. In the apparatus, a DRNN sub-speech model storage means for storing a DRNN sub-speech model learned so as to obtain a DRNN output representing certainty or more certainty with respect to a characteristic portion of each of similar words that may be erroneously recognized. When a voice of a certain word is input, a DRNN output corresponding to the input word data is output using the DRNN voice model, and a DRNN corresponding to the characteristic portion of the input word is output using the DRNN sub-voice model. A word detection signal output unit for outputting an output, and the DRNN voice model from the word detection signal output unit. If DRNN output representing at least a certain likelihood using is issued, the DRNN sets a predetermined interval including a characteristic part of the word to the output, DR by the DRNN sub speech model for said input word within the predetermined interval
A recognition processing unit for examining the NN output and identifying an input word and a word similar to the input word based on a result thereof; 21. A process for examining a DRNN output of the input word according to the DRNN sub-speech model in the predetermined section and discriminating an input word and a word similar to the input word based on a result thereof, 21. The similar word identification apparatus according to claim 20, wherein an input word and a word similar to the input word are identified based on whether the DRNN output by the sub-speech model has a value indicating certainty or more. 22. The DRNN speech model corresponds to similar word groups grouped for each type of similar word, and in each similar word group, a certain level of certainty or more is determined for all the words in the group. 22. The similar word identification apparatus according to claim 20, wherein the similar word recognition apparatus is a speech model that has been learned so as to obtain a DRNN output.