JP2011022476A - Threshold management program for voice recognition, method of the same, and voice recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a technology for facilitating work related to setting of a threshold for voice recognition and suppressing working hours. <P>SOLUTION: There is provided a threshold management program for voice recognition for making a computer execute a determination value calculation step of calculating, based on reading information showing pronunciation for each word, a determination value for determining a threshold for determining whether to recognize an input voice based on a predetermined determination reference, and a threshold value setting step of determining the threshold based on the determination value and setting it for each word. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to setting of a threshold value used when recognizing input speech.

  A speech recognition device has been developed for recognizing speech, such as converting speech uttered by a speaker into a character string, translating content uttered by the speaker, and the like. The voice recognition device performs voice recognition as follows, for example. First, the speech recognition apparatus accepts speech input, and compares a feature pattern extracted from the input speech with a standard pattern representing the speech features of each word registered in advance. Next, when the degree of coincidence between the feature pattern and the standard pattern is greater than a predetermined threshold, the speech recognition apparatus recognizes the input speech as a word corresponding to the standard pattern. As described above, since the threshold value is a criterion for determining whether or not speech is recognized, the accuracy of speech recognition depends on how the threshold value is set.

  For example, the speech recognition apparatus sets a threshold as follows. First, the speech recognition apparatus accepts speech of a plurality of words from a speaker and creates a standard pattern for each word based on the accepted speech. Further, the similarity between the standard pattern of the focused word and the standard pattern of other words is analyzed, and the threshold value of the focused word is set based on the analysis result. That is, the threshold is set based on the position where the standard pattern of the focused word should be in the relationship with the standard patterns of other words.

  Even when a new word is registered in the speech recognition device, the speech recognition device accepts the speech of the new word from the speaker and sets a threshold value as described above.

Japanese Patent No. 2553173

  However, in order to set the threshold value by the above method, it is necessary to input the voice of each word one by one, which is a complicated operation. In addition, every time a new word is registered in the speech recognition apparatus, a new word voice must be input, which makes it more complicated. As described above, since the threshold value can be set only after the speech is input for each word, it takes time before the threshold value is set, and all the work related to the threshold value setting takes a long time.

  Therefore, an object is to provide a technique capable of suppressing the work time related to the setting of the threshold value.

In order to solve the above problems, a threshold management program including the following steps is provided.
A determination value calculation step of calculating a determination value for determining a threshold value for determining whether or not to recognize the input voice based on a predetermined determination criterion based on reading information indicating pronunciation for each word.
A threshold value setting step for obtaining the threshold value based on the determination value and setting the threshold value for each word.

  It is possible to provide a technique capable of suppressing the work time related to the setting of the threshold value for voice recognition.

It is a block diagram which shows an example of the hardware constitutions of the speech recognition apparatus which concerns on the example of 1st Embodiment. It is a block diagram which shows an example of a function structure of the speech recognition apparatus which concerns on the example of 1st Embodiment. It is an example of the word dictionary DB 30 according to the first embodiment. It is a flowchart which shows an example of the flow of the threshold value management program which concerns on the example of 1st Embodiment. It is a block diagram which shows an example of a function structure of the speech recognition apparatus which concerns on the modification 1. FIG. 10 is a table showing an example of a complexity table according to Modification 1. It is a block diagram which shows an example of a function structure of the speech recognition apparatus which concerns on the example of 2nd Embodiment. It is an example of word dictionary DB30 concerning the example of a 2nd embodiment. It is a flowchart which shows an example of the flow of the threshold value management program which concerns on 2nd Embodiment. It is a block diagram which shows an example of a function structure of the speech recognition apparatus which concerns on the modification 1. FIG. 10 is a table showing an example of a similarity table according to Modification 1. It is a block diagram which shows an example of a function structure of the speech recognition apparatus which concerns on the example of 3rd Embodiment. It is an example (1) of word dictionary DB30 concerning the example of a 3rd embodiment. It is an example (2) of word dictionary DB30 concerning the example of a 3rd embodiment. It is a flowchart which shows an example of the flow of the threshold value management program which concerns on 3rd Embodiment. It is a block diagram which shows the relationship between a speech recognition apparatus, another memory | storage device, a recording medium, etc.

  Speech recognition is performed based on whether or not the degree of coincidence between the acoustic feature of the input speech and the acoustic feature of a certain word exceeds a predetermined threshold. That is, the threshold value is a determination reference value for determining whether or not the input speech is recognized as a certain word. Here, the threshold value is set based on a judgment value calculated from word reading information. The word reading information is information indicating the pronunciation of the word, for example, information representing a syllable or phoneme in a character string. For example, the reading information of the word “customer” is “Torihisaki”.

  According to this method, since the threshold value is set based on the reading information, it is not necessary to input the voice of each word to the computer prior to setting the threshold value. By eliminating the need for voice input work for setting the threshold in this way, the work time required for the previous stage of threshold setting can be suppressed, and thus the total work time related to the threshold setting can be suppressed.

  Judgment values for setting the threshold value are roughly classified into the following three patterns based on the calculation method.

  Pattern 1: The judgment value is calculated based on the reading complexity of the reading information.

  Pattern 2: The judgment value is calculated based on the similarity between the reading information.

  Pattern 3: The judgment value is calculated based on the complexity and / or similarity according to the reading information.

  These patterns 1 to 3 will be described below using first to third embodiments.

<First embodiment>
In the first embodiment, as shown in the pattern 1, the determination value is calculated based on the reading complexity of the reading information, and the threshold is set based on the complexity. The first embodiment will be described below.

(1) Hardware Configuration FIG. 1 is a block diagram showing an example of a hardware configuration of the speech recognition apparatus according to the first embodiment. In the speech recognition apparatus 1 shown in FIG. 1, a CPU (Central Processing Unit) 11, a RAM (Random Access Memory) 12, a ROM (Read Only Memory) 13, an HDD (Hard Disk Drive) 14, an input I / F (Inter Face). 15, an output I / F 17 and a communication I / F 19 are connected via the bus 10.

  The CPU 11 is a central processing unit that performs various controls and arithmetic processing, and realizes functions such as voice recognition and threshold setting based on various programs stored in a RAM 12 and a ROM 13 described later.

  The ROM 13 is a read-only storage device, and stores a voice recognition program for performing voice recognition, a threshold management program for setting a threshold described later, and the like.

  The RAM 12 is a main storage device and temporarily stores a voice recognition program, a threshold management program, and the like stored in the ROM 13. In addition, a word dictionary and a speech recognition model used for speech recognition are stored.

  The HDD 14 is an auxiliary storage device and stores various control programs and data.

  The input I / F 15 is connected to an input device 16 such as a microphone, a keyboard, and a mouse. For example, the input I / F 15 supplies the voice of the speaker received from the input device 16 to the CPU 11 and the RAM 12 via the bus 10.

  The output I / F 17 is connected to an output device 18 such as a display and a speaker, and outputs an output from the CPU 11 and the RAM 12 to the output device 18 via the bus 10.

  The communication I / F 19 transmits and receives data between the external network and the voice recognition device 1. For example, a new word dictionary and a speech recognition model are received from resources on the external network, or data processed by the speech recognition apparatus 1 is transmitted to the resources on the network.

  The bus 10 includes, for example, a PCI (Peripheral Component Interconnect) bus, an ISA (Industrial Standard Architecture) bus, and the like, and connects the above configurations to each other.

(2) Functional Configuration FIG. 2 is a block diagram illustrating an example of a functional configuration of the speech recognition apparatus according to the first embodiment. FIG. 2 shows only a functional configuration for performing functions such as voice recognition and threshold setting.

(2-1) RAM
The RAM 12 includes a word dictionary DB (Data Base) 30 and a voice recognition model DB 31.

(A) Word dictionary DB
The word dictionary DB 30 stores a word and the reading information of each word in association with each other. FIG. 3 is an example of the word dictionary DB 30 according to the first embodiment. In the word dictionary DB 30, at least words and reading information are stored in association with each other. When the complexity and the threshold are calculated by executing a threshold management program described later, the complexity and the threshold are stored in association with the word and the reading information.

  Here, the word reading information is information indicating pronunciation of the word, for example, information representing syllables and phonemes as character strings. As shown in FIG. 3, for example, the reading information of the word “customer” is “Torihisaki”. Complexity is an index indicating how complex the acoustic features of the pronunciation of a word are.

  The words and the reading information in the word dictionary DB 30 can be updated at any time by manual input via the input device 16 such as a keyboard. Alternatively, it is also possible to download words and reading information from resources of an external network via the communication I / F 19 or the like. The reading information may be generated by the voice recognition device 1 when a word is added to the word dictionary DB 30.

(B) Speech recognition model DB
The voice recognition model DB 31 stores a voice recognition model. The speech recognition model is, for example, a syllable or phoneme model including an acoustic feature pattern for each syllable or phoneme. Speech recognition is performed based on a speech recognition model including the syllable or phoneme model.

(2-2) CPU
The CPU 11 includes a voice input unit 20, a voice recognition unit 21, a complexity calculation unit 22, and a threshold setting unit 23.
(A) Audio Input Unit The audio input unit 20 generates an input audio signal by A / D converting the audio input by the input device 16 such as a microphone.
(B) Speech Recognition Unit The speech recognition unit 21 receives an input speech signal from the speech input unit 20 and performs speech recognition based on the word dictionary DB 30 and the speech recognition model DB 31. For example, the voice recognition unit 21 extracts an acoustic feature represented by the input voice signal and generates a feature pattern. The speech recognition unit 21 combines the syllables or phoneme models in the speech recognition model DB 31 to generate a syllable or phoneme model sequence (hereinafter referred to as a speech recognition model sequence). Next, it is determined whether or not the degree of coincidence between the feature pattern extracted from the input speech signal and the generated speech recognition model sequence exceeds the threshold of the word dictionary DB 30. If the degree of coincidence exceeds the threshold value in this determination, the voice recognition unit 21 recognizes the input voice as reading information of the corresponding threshold value, that is, a word.

  For example, a case where a speaker utters “Torihisaki” on the input device 16 will be described as an example. When the degree of coincidence between the feature pattern of the input voice signal “Torihisaki” and the speech recognition model sequence that is uttered exceeds the threshold 94 of “Torihisaki” shown in FIG. 3, the speech recognition unit 21 Recognize the speaker's utterance as "Torihisaki".

The voice recognition unit 21 outputs the recognition result to the output device 18 via the bus 10, for example.
(C) Complexity Calculation Unit The complexity calculation unit 22 calculates the reading complexity of reading information for each word.

  Here, as the number of phonemes and syllables of a word increases, it becomes difficult to utter and the acoustic features tend to be complicated. For example, “ka” consisting of two syllables of a plosive and a vowel is more complex than “a” consisting of only one syllable, and “cha” consisting of three syllables of a plosive, friction and vowel. It becomes complicated. When multiple syllables are combined, “Aka” consists of three syllables of vowels, plosives and vowels, and “Akacha” consists of six syllables of vowels, plosives, vowels, plosives, friction sounds and vowels, These features are further complicated. Therefore, in this embodiment, the complexity is calculated based on the number of elements constituting the pronunciation of the reading information, such as the number of syllables or phonemes obtained from the reading information. In addition, the element includes a unit of a speech recognition model such as a syllable model and a phoneme model used for speech recognition.

  Specifically, the complexity calculation unit 22 extracts the reading information of each word from the word dictionary DB 30, decomposes the reading information into syllables or phonemes, and calculates the syllable number or phoneme number as the complexity. For example, “Torihi Saki” shown in FIG. 3 is composed of six syllables “to” “ri” “hi” “ki” “sa” “ki”, and therefore the complexity is calculated as “6”. Since “sai” is composed of four syllables “sa”, “i”, “shi”, and “n”, the complexity is calculated as “4”.

Note that the complexity is not defined by the number of syllables or phonemes, but is defined by the function of the number of syllables or phonemes, for example, the complexity increases exponentially as the number of syllables or phonemes increases. May be.
(D) Threshold setting unit The threshold setting unit 23 sets a threshold for the reading information based on the complexity calculated by the complexity calculating unit 22. It is preferable to set the threshold value to be smaller as the complexity increases. The reason for this will be described next.

  The greater the complexity of reading a word, the more complex the acoustic features of the word. For example, when the complexity is defined by the number of syllables, the two-syllable “Yama” and the six-syllable “Torihisaki” have more syllables and more complicated acoustic features. Becomes larger.

  Here, when recognizing a voice, for example, the degree of coincidence between a voice recognition model sequence representing the acoustic characteristics of each word and the voice that is uttered is compared based on a threshold value, and the uttered word is recognized. If the complexity of the word is large, the acoustic characteristics of the word become complicated as described above, so that the voice of the spoken word is difficult to match the speech recognition model sequence of the word, but the speech recognition of another word It becomes difficult to match the model column. Therefore, when the complexity is large, the threshold value is set small enough to recognize the spoken word and not to misrecognize it. As a result, the level of coincidence between the voice of the spoken word and the voice recognition model sequence is lowered, and erroneous recognition is suppressed while making it easier to recognize the spoken word.

  Conversely, if the complexity is low, the acoustic characteristics of a word become simple, so that the voice of a spoken word easily matches the speech recognition model string of that word. It becomes easy to match. Therefore, when the complexity is small, a large threshold is set. As a result, the level of coincidence between the voice of the uttered word and the voice recognition model sequence is increased, and erroneous recognition is suppressed while the uttered word is recognized correctly.

  Specifically, the threshold setting unit 23 sets the threshold based on, for example, the following equation (1) so that the threshold decreases as the complexity increases.

Threshold = predetermined value-complexity (1)
For example, when the degree of coincidence between the speech recognition model sequence and the speaker's speech is calculated in the range of 0 to 100, the predetermined value is 100 as an example. The relational expression between the threshold value and the complexity is not limited to the above formula (1), and the complexity may be multiplied by a coefficient.

  According to the above formula (1), the threshold value is “94” because the complexity of “Torihi Saki” is “6”, and the threshold value is “96” because the complexity of “Saishin” is “4”. .

  The threshold value setting unit 23 outputs the threshold value set in this way to the word dictionary DB 30, and stores the threshold value as shown in FIG.

(3) Process Flow Next, a process flow for setting the threshold will be described. FIG. 4 is a flowchart showing an example of the flow of the threshold management program according to the first embodiment. The threshold management program is executed by the complexity calculation unit 22 and the threshold setting unit 23.

  Step S1: The complexity calculator 22 determines whether there is a word for which no threshold is set in the word dictionary DB 30. If there is a word for which no threshold is set, the process proceeds to step S2. On the other hand, if thresholds are set for all words, the threshold management program is terminated.

  Step S2: The complexity calculator 22 selects a word for which no threshold is set from the word dictionary DB 30. That is, one record with no threshold set is selected in the word dictionary DB 30 of FIG.

  Steps S3 and S4: The complexity calculator 22 decomposes the reading information of the selected word into syllables or phonemes (S3), and calculates the syllable number or phoneme number as complexity (S4).

  Step S5: The threshold setting unit 23 receives the complexity from the complexity calculation unit 22, and sets a threshold based on the complexity. For example, the threshold value is set so as to decrease as the complexity increases, for example, based on the above formula (1).

  Step S6: The threshold value setting unit 23 outputs the threshold value to the word dictionary DB 30 and updates the word dictionary DB 30.

  In the above processing, the complexity is calculated based on the number of syllables or phonemes obtained from the reading information. Since the threshold is set based on this complexity, it is not necessary to input the voice of each word to the computer prior to setting the threshold. By eliminating the need for voice input work for setting the threshold in this way, the work time required for the previous stage of threshold setting can be suppressed, and thus the total work time related to the threshold setting can be suppressed.

(4) Experimental Results (a) Experimental Example An experiment was performed to verify the accuracy of speech recognition using the speech recognition apparatus 1 in which the threshold value was set by the above method. The word dictionary DB 30 stores 358 words including common nouns such as “confirmation”, “acceptance”, “setting”, and proper nouns such as “XX Inc.” and reading information. Moreover, the threshold value was set by the above method for each word. To the voice input unit 20, voice data for 70 minutes (44 calls for the call center) was input to perform voice recognition. As a result, as shown in Table 1 below, the number of voice recognition failures was 224. Here, the number of failures is expressed as the sum of the number of times the word that should have been recognized could not be recognized and the number of times of erroneous recognition.
(B) Comparative Examples 1-3
In the speech recognition apparatus 1 used in the above experimental example, the optimum value was 92 when the threshold value was set to a constant value. Therefore, in the comparative example, the threshold value was swung up and down with respect to the optimum value and set to 90, 92, and 94 to be comparative examples 1 to 3, and voice recognition was performed in each case. The speech data to be recognized is the same as the above experimental example. Except for the point that the threshold value is constant, the 358 words stored in the word dictionary DB 30 are the same as described above. The number of voice recognition failures at each threshold is shown in Table 1 below.

使用した音声認識装置１では、実験例の失敗回数は、最適値９２を閾値として設定した比較例２の場合よりも少なくなった。また、実験例の失敗回数は、比較例１及び比較例３の場合よりも少ない。よって、単語の読み情報に基づいて閾値を設定した場合であっても、閾値として適度な値が設定されていることが分かった。

In the used speech recognition apparatus 1, the number of failures in the experimental example is smaller than that in the comparative example 2 in which the optimum value 92 is set as a threshold value. In addition, the number of failures in the experimental example is smaller than in the case of Comparative Example 1 and Comparative Example 3. Therefore, it was found that even when a threshold value was set based on word reading information, an appropriate value was set as the threshold value.

(5) Modification (5-1) Modification 1
Next, a modification of the first embodiment will be described. In the above description, the complexity is calculated based on the number of syllables or phonemes of a word. In this modification, the complexity of a word is calculated based on the complexity defined for each syllable.

  FIG. 5 is a block diagram illustrating an example of a functional configuration of the speech recognition apparatus according to the first modification. The difference from FIG. 2 is that the RAM 12 has a complexity table 32. Since other configurations and processing flows are the same as those in the above embodiment, description thereof is omitted.

  FIG. 6 is a table showing an example of the complexity table according to the first modification. Complexity is defined for each syllable such as “A” and “I”. Thus, by setting the complexity for each syllable, the complexity of the word can be calculated according to the acoustic features of each syllable.

  The complexity calculator 22 selects a word for which no threshold is set from the word dictionary DB 30 and breaks it down into syllables. Further, the complexity calculation unit 22 calculates the sum of the complexity of each syllable using the complexity table 32, and calculates the complexity of the word.

  Specifically, the reading information of “Torihi Saki” and “Saishin” will be described as an example. When "Torihisaki" is broken down into syllables, it becomes "to", "ri", "hi", "ki", "sa" and "ki". Since the complexity of each syllable is 2, the complexity of “Torihisaki” is calculated as 12 from the sum. Also, when “saishin” is broken down into syllables, “sa”, “i”, “shi” and “n”. Since the complexity of “sa” and “shi” is 2, and the complexity of “i” and “n” is 1, the complexity of “sai” is calculated as 6 from the sum.

  Although the complexity table 32 defines the complexity for each syllable, the complexity may be defined for each element that constitutes the pronunciation of reading information such as phonemes. In addition, the elements include units of a speech recognition model such as a syllable model and a phoneme model. The complexity calculator 22 decomposes the word reading information based on the elements of the complexity table 32. That is, in the complexity table 32, when the complexity is defined for each phoneme, the reading information is decomposed into phonemes and the complexity of the word is calculated.

(5-2) Modification 2
In the word dictionary DB 30 of the above embodiment, only one reading information is associated with one word, but a plurality of reading information may be associated with one word. For example, there is a case where the reading information of “10” and “to” is associated with the word “ten”. Further, as described later in the second embodiment, the first reading information indicating the original pronunciation of the word and the second reading information modified from the first reading information are associated with one word. There are cases where it is.

  When a plurality of pieces of reading information are associated with one word, the complexity calculating unit 22 calculates the complexity for each piece of reading information. The threshold setting unit 23 calculates a threshold for each reading information based on the complexity for each reading information.

  For example, a case where one word “Okinawa” has first reading information “Okinawa” and second reading information “Kinaa” will be described. When calculating the complexity based on the number of syllables, the complexity calculating unit 22 calculates the first complexity as “4” because the first reading information “Okinawa” consists of four syllables. Further, the complexity calculation unit 22 calculates the second complexity as “3” because the second reading information “Kinaa” includes three syllables. The first complexity is the reading complexity of the first reading information, and the second complexity is the reading complexity of the second reading information.

  The threshold value setting unit 23 sets a first threshold value for the first reading information based on the first complexity, and sets a second threshold value for the second reading information based on the second complexity. Therefore, the threshold value can be finely set according to the difference in complexity of each reading information.

(5-3) Modification 3
The complexity calculation method is not limited to the methods of the first embodiment, the first modification, and the second modification. For example, even if the complexity of the reading information of interest is calculated based on the degree of similarity between the speech recognition model sequence of the reading information of interest and the average model of all the speech recognition model sequences good. The speech recognition model sequence includes various parameters, templates, and the like, but the complexity may be calculated based on the number of parameters and the number of templates of the speech recognition model sequence constituting the reading information.

<Second Embodiment>
In the second embodiment, a plurality of reading information is associated with one word, and the determination value is calculated based on the similarity between the reading information as shown in the pattern 2 above. The threshold is set based on the similarity. The second embodiment will be described below.

  Since the hardware configuration is the same as that of the first embodiment, description thereof is omitted.

(1) Functional Configuration FIG. 7 is a block diagram showing an example of a functional configuration of the speech recognition apparatus according to the second embodiment.

(1-1) RAM
The RAM 12 includes a word dictionary DB 30 and a speech recognition model DB 31. Since the voice recognition model DB 31 is the same as that of the first embodiment, description thereof is omitted.

  The word dictionary DB 30 stores a word and the reading information of each word in association with each other. FIG. 8 shows an example of the word dictionary DB 30 according to the second embodiment. In the word dictionary DB 30, a plurality of pieces of reading information are associated with one word. Specifically, the first reading information and the second reading information are associated with one word. Here, the first reading information is information indicating the original pronunciation of the word, and the second reading information is information indicating the pronunciation modified from the first reading information. In the case of the word “Okinawa”, as will be described later, the first reading information is “Okinawa”, and the second reading information is, for example, “Okinawa” and “Kinaa”.

  Since the first reading information is a standard pronunciation of a word, it can be said that the speaker generally tries to utter when a certain word is uttered. However, even if the speaker is trying to pronounce based on the first reading information, the speaker may not always be able to pronounce based on the first reading information due to the tone of the speaker or the relationship with the utterance contents before and after. . Therefore, not only the first reading information that is the original pronunciation of the word but also the second reading information is associated with one word.

  For example, a case where a speaker is going to utter the word “Okinawa” will be described as an example. Even if the speaker tries to utter “Okinawa”, which is the original pronunciation of the word, the speaker may utter “Okina”, “Kina”, etc. depending on the tone of the speaker. Therefore, the word “Okinawa” is associated with “Okinawa” and “Kinaa” as the second reading information in addition to “Okinawa” as the first reading information.

  In the word dictionary DB 30 of FIG. 8, only an example in which a plurality of reading information is associated with each word is shown, but a word in which a plurality of reading information is associated with a word in which only one reading information is associated. And may be mixed.

  The words and reading information in the word dictionary DB 30 can be updated as needed by manual input via the input device 16 or download from external resources via the communication I / F 19 or the like. The reading information may be automatically generated by the voice recognition device 1 when a word is added to the word dictionary DB 30.

(1-2) CPU
The CPU 11 includes a voice input unit 20, a voice recognition unit 21, a threshold setting unit 23, and a similarity calculation unit 24. Since the voice input unit 20 and the voice recognition unit 21 are the same as those in the first embodiment, description thereof will be omitted.
(A) Similarity calculation unit The similarity calculation unit 24 calculates the degree of similarity of each reading information with the first reading information indicating the standard pronunciation of the word. For example, the similarity is expressed based on a distance indicating how far each reading information is apart from the first reading information. This distance is calculated, for example, by the sum of the number of syllable substitutions, dropouts, and insertions from the first reading information to the second reading information. In addition, the relationship between similarity and distance shall be represented, for example by following Formula (2).

Similarity = 0-distance (2)
A method of calculating the similarity will be described by taking “Okinawa” in the word dictionary DB 30 of FIG. 8 as an example. The similarity calculating unit 24 compares the first reading information “Okinawa” with any of the second reading information. In calculating the similarity, the similarity calculation unit 24 first decomposes each reading information into syllables. Here, when “wa” is replaced with “a” in the first reading information “Okinawa”, the second reading information “Okinawa” is obtained. Therefore, the distance between the first reading information “Okinawa” and the second reading information “Okinawa” is represented by the number of substitutions “1”, and the similarity is represented by “−1”. Similarly, when “o” is dropped in the first reading information “Okinawa” and “wa” is replaced with “a”, the second reading information “kina” is obtained. Therefore, the distance is represented by the sum “2” of the dropout number “1” and the replacement number “1”, and the similarity is represented by “−2”. For the first standard reading “Okinawa”, the number of substitutions, the number of omissions and the number of insertions are “0”, and the similarity is represented by “0”.

In the above, the similarity is calculated based on the number of syllable replacements, the number of omissions, and the number of insertions, but the method of calculating the similarity is not limited to this. For example, the first reading information may be compared with the second reading information, and the similarity may be calculated based on the number of replacements, dropouts, and insertions of phonemes, syllable models, and phoneme models.
(B) Threshold setting unit The threshold setting unit 23 sets a threshold based on the similarity. It is preferable that the threshold value is set to increase as the degree of similarity decreases, that is, as the distance increases.

  Here, as described above, the first reading information is information indicating the standard pronunciation of the original word, and the second reading information is information indicating the pronunciation modified from the first reading information. Therefore, it can be said that the second reading information having a smaller degree of similarity with the first reading information, that is, a larger distance from the first reading information, is far from the standard pronunciation of the word. Therefore, the second reading information having a smaller degree of similarity with the first reading information is set to a larger threshold value to suppress erroneous recognition. The reason for setting the threshold in this way will be described next.

  When speech recognition is performed based on the second reading information, depending on the size of the threshold, not only the speech based on the second reading information but also the speech similar to the second reading information is recognized as the second reading information. However, even if the voice is similar to the second reading information, it may not be the voice of the word meant by the first reading information. For this reason, there is a possibility that misrecognition may occur because the word meant by the spoken voice is different from the word recognized by the voice recognition device 1.

  For example, the first reading information of “Tokyo” is “Tokyo”, and the second reading information is “Toyocho”. When the speaker utters “Toyocho”, the speech recognition apparatus 1 recognizes that “Tokyo” is uttered based on the second reading information “Toyocho”. On the other hand, it is assumed that the speaker utters “Tokkyo” similar to the second reading information “Toyocho”, for example. Here, the reading information “Tokkyo” means the word “patent”. Depending on the size of the threshold, the speech recognition apparatus 1 recognizes the uttered “Tokkyo” as the second reading information “Toyocho” and misrecognizes that “Tokyo” is uttered. It can be said that such misrecognition tends to become more prominent as the similarity between the first reading information and the second reading information is smaller. Therefore, in consideration of the similarity between the first reading information and the second reading information, the second reading information having a smaller similarity with the first reading information is set to a larger threshold value, and the recognition accuracy is suppressed by suppressing erroneous recognition. To increase.

  The threshold setting unit 23 sets the threshold based on, for example, the following equation (3) so that the threshold increases as the similarity decreases, that is, the threshold increases as the distance increases.

Threshold = predetermined value−similarity (3)
For example, when the degree of coincidence between the speech recognition model sequence and the speaker's speech is calculated in the range of 0 to 100, the predetermined value is 80 as an example.

  According to the above formula (3), since the similarity of the second reading information “OKINAA” is “−1” as described above, the threshold value is “81”, and the similarity of the second reading information “KINAA” is described above. Since the threshold is “−2”, the threshold value is “82”. The first reading information “Okinawa” has a threshold value of “80” because the similarity is “0”. That is, a default value, that is, a predetermined value of Expression (3) is set as the threshold value of the first reading information. The threshold value setting unit 23 outputs the threshold value set in this way to the word dictionary DB 30, and stores the threshold value as shown in FIG.

(2) Process Flow Next, a process flow for setting a threshold will be described. FIG. 9 is a flowchart showing an exemplary flow of a threshold management program according to the second embodiment. The threshold management program is executed by the similarity calculation unit 24 and the threshold setting unit 23.

  Step S11: The similarity calculation unit 24 determines whether there is reading information for which no threshold is set in the word dictionary DB 30. If there is reading information for which no threshold is set, the process proceeds to step S12. On the other hand, if threshold values are set for all reading information, the threshold management program is terminated.

  Step S12: The similarity calculation unit 24 selects reading information for which no threshold is set from the word dictionary DB 30, and determines whether the reading information is first reading information. If it is the second reading information, the process proceeds to step S14, and if it is the first reading information, the process proceeds to step S13.

  Step S13: When the selected reading information is the first reading information, the similarity with the first reading information is “0”. Therefore, the threshold value setting unit 23 sets a default value, that is, a predetermined value of the above formula (3) as the threshold value of the first reading information.

  Step S14: When the selected reading information is the second reading information, the similarity calculating unit 24 decomposes each reading information into syllables and the like, and calculates the similarity between the first reading information and the second reading information. .

  Step S15: The threshold setting unit 23 receives the similarity from the similarity calculation unit 24, and sets a threshold based on the similarity.

  Step S16: The threshold value setting unit 23 outputs a threshold value to the word dictionary DB 30, and updates the word dictionary DB 30.

  In the above processing, since the threshold is set based on the similarity calculated from the reading information, it is not necessary to input the voice of each word to the computer prior to setting the threshold. By eliminating the need for voice input work for setting the threshold in this way, the work time required for the previous stage of threshold setting can be suppressed, and thus the total work time related to the threshold setting can be suppressed.

(3) Modification (3-1) Modification 1
Next, a modified example of the second embodiment will be described. In the above description, the similarity is calculated based on the number of syllable replacements, dropouts, and insertions from the first reading information to the second reading information. However, in this modification, the similarity between syllables is defined in advance in a similarity table, and the similarity of reading information is calculated based on this similarity table.

  FIG. 10 is a block diagram illustrating an example of a functional configuration of the speech recognition apparatus according to the first modification. The difference from FIG. 7 is that the RAM 12 has a similarity table 33. Since other configurations and processing flows are the same as those in the above embodiment, description thereof is omitted.

  FIG. 11 is a table showing an example of a similarity table according to the first modification. For each of substitution, insertion, and omission, a distance that is an index of similarity is defined for each syllable. By setting the distance for each syllable in this way, it is possible to finely calculate the similarity between reading information according to the acoustic characteristics of each syllable.

  The similarity calculation unit 24 uses the similarity table 33 to calculate the similarity between the first reading information and the second reading information. For example, the similarity between the first reading information “Okinawa” and the second reading information “Okina” is “4” because the distance of replacement from “wa” to “a” is “4”. Calculated as “−4”. Also, the similarity between the first reading information “Okinawa” and the second reading information “Kinaa” is the dropout distance “5” of “O” and the replacement distance “4” from “wa” to “a”. Therefore, “−9” is calculated from the above equation (2).

  Although the distance is defined for each syllable in the similarity table 33 described above, the distance may be defined for each element constituting the pronunciation of reading information such as phonemes. In addition, the elements include units of a speech recognition model such as a syllable model and a phoneme model. The similarity calculation unit 24 decomposes the word reading information based on the elements of the similarity table 33. That is, when the distance is defined for each phoneme in the similarity table 33, the reading information is decomposed into phonemes and the similarity is calculated.

(3-2) Modification 2
In consideration of the number of reading information for one word, a threshold may be set as in the following equation (4).

Threshold = predetermined value−similarity + number of reading information (4)
The greater the number of reading information, the wider the range of pronunciation similar to the reading information and the higher the possibility of erroneous recognition. Therefore, the threshold is set so as to increase as the number of reading information increases.

  For example, the word “Okinawa” shown in FIG. 8 has a total of three pieces of reading information of “Okinawa”, “Okinawa”, and “Kinaa”. Therefore, in the case of the second reading information “OKINAA”, the threshold value is calculated as “84” by the predetermined value “80” −similarity “−1” + number of reading information “3” based on the equation (4). The

<Third Embodiment>
In the third embodiment, similarly to the second embodiment, the first reading information and the second reading information are associated with one word. In the third embodiment, the determination value is calculated based on the complexity and / or similarity according to the reading information as shown in the pattern 3 above. The third embodiment will be described below.

  Since the hardware configuration is the same as that of the first embodiment, description thereof is omitted.

(1) Functional Configuration FIG. 12 is a block diagram showing an example of a functional configuration of a speech recognition apparatus according to the third embodiment. 13 and 14 show an example of the word dictionary DB 30 according to the third embodiment.

  The CPU 11 includes a voice input unit 20, a voice recognition unit 21, a complexity calculation unit 22, a threshold setting unit 23, and a similarity calculation unit 24. The RAM 12 includes a word dictionary DB 30 and a speech recognition model DB 31. Since the voice input unit 20, the voice recognition unit 21, and the voice recognition model DB 31 are the same as those in the first embodiment, description thereof is omitted.

(1-1) Calculation method A
First, the first threshold value of the first reading information is calculated based on the first complexity, and the second threshold value of the second reading information is calculated based on the second complexity and the similarity (calculation method A). explain. Here, the first threshold is a threshold for determining whether or not the voice is uttered by the first reading information, and the second threshold is a threshold for determining whether or not the voice is uttered by the second reading information. . The first complexity is the reading complexity of the first reading information, and the second complexity is the reading complexity of the second reading information. The similarity is a distance indicating how far the second reading information is separated from the first reading information.

  The complexity calculator 22 calculates the first complexity and the second complexity for each of the first reading information and the second reading information. Since the complexity calculation method is the same as that in the first embodiment, description thereof is omitted. For example, when the complexity is expressed by the number of syllables, the complexity of each reading information is calculated as shown in FIG. FIG. 13 is an example of the word dictionary DB 30.

  The similarity calculation unit 24 calculates the similarity between each reading information and the first reading information. Since the similarity calculation method is the same as that in the second embodiment, description thereof is omitted. For example, when the similarity is represented by the number of syllable replacements, the number of omissions, and the number of insertions, the similarity of each piece of reading information is calculated as shown in FIG.

  The threshold setting unit 23 sets the threshold based on, for example, the following equation (5) so that the threshold decreases as the complexity increases and so that the threshold increases as the similarity decreases.

Threshold = predetermined value−complexity−similarity (5)
For example, when the degree of coincidence between the speech recognition model sequence and the speaker's speech is calculated in the range of 0 to 100, the predetermined value is 90 as an example.

  According to the above equation (5), the first reading information “Okinawa” has a first complexity of “4” and a similarity of “0”, so the first threshold is “86”. Further, the second reading information “OKINAA” has a second complexity of “4” and a similarity of “−1”, so the second threshold is “87”. The first threshold and the second threshold are stored in the word dictionary DB 30 as shown in FIG.

  Here, since the similarity of the first reading information is “0”, the first threshold is set for the first reading information only by the first complexity. On the other hand, in the second reading information, the second threshold is set in consideration of both the second complexity and the similarity. Therefore, the second threshold can be set more finely according to the acoustic characteristics of the second reading information.

(1-2) Calculation method B
Next, a case where the first threshold value of the first reading information is calculated based on the first complexity and the second threshold value of the second reading information is calculated based only on the similarity (calculation method B) will be described.

  The complexity calculator 22 calculates the first complexity for the first reading information. The similarity calculation unit 24 calculates the similarity between each reading information and the first reading information.

  The threshold setting unit 23 sets the first threshold of the first reading information based on the first complexity. For example, the first threshold is set by setting the similarity to “0” in the above equation (5). Since the first complexity of the first reading information “Okinawa” is “4”, the first threshold value is “86”.

  On the other hand, the threshold setting unit 23 sets the second threshold of the second reading information based only on the similarity. For example, in the above equation (5), the second threshold is set with the complexity set to “0”. Since the similarity of the second reading information “OKINAA” is “−1”, the second threshold value is “91”. The first threshold value, the second threshold value, and the like are stored in the word dictionary DB 30 as shown in FIG.

(2) Process Flow Next, a process flow for setting a threshold will be described. FIG. 15 is a flowchart showing an exemplary flow of a threshold management program according to the third embodiment. The threshold management program is executed by the complexity calculation unit 22, the threshold setting unit 23, and the similarity calculation unit 24.

  Step S21: The complexity calculation unit 22 determines whether there is a word for which no threshold is set in the word dictionary DB 30.

  Step S22: The complexity calculator 22 selects a word for which no threshold is set from the word dictionary DB 30.

  Step S23: The complexity calculator 22 calculates the complexity of the reading information.

  Here, in the case of the calculation method A, the complexity calculation unit 22 calculates all the complexity of the first reading information and the second reading information. On the other hand, in the case of the calculation method B, the complexity calculation unit 22 calculates only the first complexity of the first reading information.

  Step S24: The similarity calculation unit 24 calculates the similarity between the first reading information and the second reading information.

  Step S25: The threshold setting unit 23 sets the first threshold and the second threshold based on the complexity and the similarity.

  Step S26: The threshold value setting unit 23 outputs the first threshold value and the second threshold value to the word dictionary DB 30, and updates the word dictionary DB 30.

  Step S23 and step S24 are in no particular order. Further, the subject performing the processes of steps S21 and S22 does not have to be the complexity calculating unit 22, and may be, for example, the similarity calculating unit 24.

(3) Modification (3-1) Modification 1
In formula (5) of the above embodiment, the predetermined value is a constant “90”. However, when setting the second threshold, a predetermined value may be set as the first threshold. For example, when the first threshold value of the first reading information is “86” as shown in FIG. 13, the second threshold value is set based on the following equation (6).

Second threshold = 86−complexity−similarity (6)
When the second threshold of the second reading information is set in consideration of both the complexity and the similarity, the second threshold of the second reading information “Okinaa” is 86−complexity “4” −similarity “−. 1 "= 83 is set. Thereby, the second threshold value of the second reading information can be calculated in consideration of the first threshold value of the first reading information.

(3-2) Modification 2
Similarly to the second modification of the second embodiment, the threshold may be set as in the following equation (7) in consideration of the number of reading information.

Threshold = predetermined value−complexity−similarity + number of reading information (7)
<Other embodiment examples>
(A) In the above description, the threshold value is set by taking a Japanese word as an example. However, the threshold value setting method can be applied to any language.

(B)
In the above embodiment, various programs such as a voice recognition program and a threshold management program are stored in the ROM 13. In addition, the various programs may be stored in the RAM 12, the HDD 14, or the like.

  The various programs do not need to be stored in the voice recognition device 1 and may be stored in an external storage device 64 connected to the voice recognition device 1 as shown in FIG. FIG. 16 is a block diagram showing the relationship between the voice recognition device and other storage devices and recording media.

  A computer-readable recording medium 62 that records the program is included in the scope of the present invention. Here, examples of the computer-readable recording medium include a flexible disk, a hard disk, a CD-ROM, an MO, a DVD, a DVD-ROM, a DVD-RAM, a BD (Blu-ray Disc), a USB memory, and a semiconductor memory. Can be mentioned.

  The program may be transmitted from the storage device 61 via an electric communication line, a wireless or wired communication line, a network represented by the Internet, or the like.

  Regarding the above embodiment and other embodiments, the following additional notes are disclosed.

<Appendix>
(Appendix 1)
A determination value calculating step for calculating a determination value for determining a threshold value for determining whether or not to recognize the input voice based on a predetermined determination criterion based on reading information indicating pronunciation for each word;
A threshold setting step for obtaining the threshold based on the determination value and setting the threshold for each word;
Threshold management program for voice recognition for causing a computer to execute.

(Appendix 2)
The determination value calculating step includes a complexity calculation step of calculating a reading complexity of the reading information as the determination value,
The threshold management for speech recognition according to claim 1, wherein in the threshold setting step, the threshold is obtained so that the threshold decreases as the complexity increases, and is set for each word based on the complexity. program.

(Appendix 3)
The computer includes a word dictionary that associates the word with the reading information,
The threshold management program for speech recognition according to attachment 2, wherein in the threshold setting step, a threshold is set for every word in the word dictionary.

(Appendix 4)
The reading information includes first reading information indicating the original pronunciation of a word, and second reading information that is a pronunciation of the word different from the first reading information,
In the complexity calculation step, a first complexity is calculated based on the first reading information, a second complexity is calculated based on the second reading information,
In the threshold setting step, a first threshold for determining whether or not the word is uttered by the first reading information is set based on the first complexity, and the word is uttered by the second reading information. Setting a second threshold for determining whether or not the determination has been made based on the second complexity.
The threshold value management program for speech recognition according to appendix 2 or 3.

(Appendix 5)
The determination value calculating step further includes a similarity calculating step for calculating a similarity between the first reading information and the second reading information,
In the threshold setting step, the second threshold is obtained and set for each word based on the similarity in addition to the second complexity so that the second threshold increases as the similarity decreases. The threshold management program for speech recognition according to appendix 4.

(Appendix 6)
The reading information includes first reading information indicating the original pronunciation of a word, and second reading information that is a pronunciation of the word different from the first reading information,
In the complexity calculation step, a first complexity is calculated based on the first reading information,
The determination value calculating step further includes a similarity calculating step for calculating a similarity between the first reading information and the second reading information,
In the threshold setting step, a first threshold for determining whether or not the word is uttered by the first reading information is set based on the first complexity, and the word is uttered by the second reading information. The second threshold value for determining whether or not the determination has been made is determined for each word based on the similarity, so that the second threshold value increases as the similarity decreases. Threshold management program for voice recognition.

(Appendix 7)
The threshold management program for speech recognition according to any one of appendices 2 to 6, wherein in the complexity calculation step, the complexity is calculated based on the number of elements constituting the pronunciation of the word.

(Appendix 8)
The threshold for speech recognition according to any one of appendices 2 to 6, wherein in the complexity calculation step, the complexity is calculated based on a sum of element complexity defined for each element constituting the pronunciation of the word. Management program.

(Appendix 9)
The reading information includes first reading information indicating the original pronunciation of a word, and second reading information that is a pronunciation of the word different from the first reading information,
The determination value calculation step includes a similarity calculation step of calculating a similarity between the first reading information and the second reading information as the determination value,
In the threshold setting step, a second threshold for determining whether or not the word is uttered by the second reading information is based on the similarity, and the second threshold increases as the similarity decreases. The threshold management program for speech recognition according to appendix 1, wherein the second threshold is obtained as described above and set for each word.

(Appendix 10)
A threshold management method for speech recognition executed by a computer,
A determination value calculating step for calculating a determination value for determining a threshold value for determining whether or not to recognize the input voice based on a predetermined determination criterion based on reading information indicating pronunciation for each word;
A threshold setting step for obtaining the threshold based on the determination value and setting the threshold for each word;
Threshold management method for speech recognition including

(Appendix 11)
A determination value calculating means for calculating a determination value for determining a threshold value for determining whether or not to recognize an input voice based on a predetermined determination criterion, based on reading information indicating pronunciation for each word;
Threshold setting means for obtaining the threshold based on the judgment value and setting the threshold for each word;
Speech recognition means for determining whether to recognize the input speech as the word using the threshold;
A speech recognition device.

1: Speech recognition device 20: Speech input unit 21: Speech recognition unit 22: Complexity calculation unit 23: Threshold setting unit 24: Similarity calculation unit 30: Word dictionary DB
31: Voice recognition model DB
32: Complexity table 33: Similarity table

Claims (9)

A determination value calculating step for calculating a determination value for determining a threshold value for determining whether or not to recognize the input voice based on a predetermined determination criterion based on reading information indicating pronunciation for each word;
A threshold setting step for obtaining the threshold based on the determination value and setting the threshold for each word;
Threshold management program for voice recognition for causing a computer to execute. The determination value calculating step includes a complexity calculation step of calculating a reading complexity of the reading information as the determination value,
The threshold value for speech recognition according to claim 1, wherein in the threshold value setting step, the threshold value is determined for each word based on the complexity so that the threshold value decreases as the complexity increases. Management program. The computer includes a word dictionary that associates the word with the reading information,
The threshold management program for speech recognition according to claim 2, wherein in the threshold setting step, a threshold is set for every word in the word dictionary. The reading information includes first reading information indicating the original pronunciation of a word, and second reading information that is a pronunciation of the word different from the first reading information,
In the complexity calculation step, a first complexity is calculated based on the first reading information, a second complexity is calculated based on the second reading information,
In the threshold setting step, a first threshold for determining whether or not the word is uttered by the first reading information is set based on the first complexity, and the word is uttered by the second reading information. Setting a second threshold for determining whether or not the determination has been made based on the second complexity.
The threshold value management program for speech recognition according to claim 2 or 3. The determination value calculating step further includes a similarity calculating step for calculating a similarity between the first reading information and the second reading information,
In the threshold setting step, the second threshold is obtained and set for each word based on the similarity in addition to the second complexity so that the second threshold increases as the similarity decreases. The threshold management program for speech recognition according to claim 4. The reading information includes first reading information indicating the original pronunciation of a word, and second reading information that is a pronunciation of the word different from the first reading information,
In the complexity calculation step, a first complexity is calculated based on the first reading information,
The determination value calculating step further includes a similarity calculating step for calculating a similarity between the first reading information and the second reading information,
In the threshold setting step, a first threshold for determining whether or not the word is uttered by the first reading information is set based on the first complexity, and the word is uttered by the second reading information. The second threshold value for determining whether or not the determination has been made is determined for each word based on the similarity level so that the second threshold value increases as the similarity level decreases. The threshold management program for voice recognition as described. The reading information includes first reading information indicating the original pronunciation of a word, and second reading information that is a pronunciation of the word different from the first reading information,
The determination value calculation step includes a similarity calculation step of calculating a similarity between the first reading information and the second reading information as the determination value,
In the threshold setting step, a second threshold for determining whether or not the word is uttered by the second reading information is based on the similarity, and the second threshold increases as the similarity decreases. The threshold management program for speech recognition according to claim 1, wherein the second threshold is obtained and set for each word. A threshold management method for speech recognition executed by a computer,
A determination value calculating step for calculating a determination value for determining a threshold value for determining whether or not to recognize the input voice based on a predetermined determination criterion based on reading information indicating pronunciation for each word;
A threshold setting step for obtaining the threshold based on the determination value and setting the threshold for each word;
Threshold management method for speech recognition including A determination value calculating means for calculating a determination value for determining a threshold value for determining whether or not to recognize an input voice based on a predetermined determination criterion, based on reading information indicating pronunciation for each word;
Threshold setting means for obtaining the threshold based on the judgment value and setting the threshold for each word;
Speech recognition means for determining whether to recognize the input speech as the word using the threshold;
A speech recognition device.

Images