JP6451171B2 - Speech recognition apparatus, speech recognition method, and program - Google Patents

Description

  The present invention relates to a voice recognition device, a voice recognition method, and a program.

  In recent years, devices that recognize only specific utterance contents and specific commands have been developed and are beginning to be utilized in information systems. A speech recognition device is known as an example of such a device. The voice recognition device is a device that recognizes a user's utterance content and receives it as an input to the information providing system.

  Such a speech recognition apparatus is used in, for example, a call center system. In this case, for example, the speech recognition apparatus extracts the call contents including a keyword (recognition word) registered in advance from a large amount of call contents and classifies or analyzes the call contents to include the keywords. It is used to give an index accordingly.

  In order to improve the accuracy of the voice search function and information providing function, it is necessary to improve the accuracy of the response to the user's utterance content. That is, it is essential to improve the accuracy of the speech recognition apparatus. That is, in the speech recognition apparatus, there is a demand for a technique for reducing erroneous detection and detecting only an appropriate recognition result with high accuracy, that is, a technique capable of obtaining a high reproduction rate and a high matching rate.

  Here, the recall is an index relating to completeness. In speech recognition, the number of correctly recognized words is divided by the number of words to be recognized. In addition, the relevance rate is an index related to accuracy, and in speech recognition, the number of correctly recognized words is divided by the number of recognized words.

  For example, Patent Document 1 (hereinafter also referred to as Prior Art 1) and Patent Document 2 (hereinafter also referred to as Prior Art 2) have proposed techniques aimed at accurately detecting appropriate recognition words. Yes.

  The speech recognition device proposed in Patent Document 1 (hereinafter referred to as the speech recognition device of Prior Art 1) uses a noise model adapted to the environment (hereinafter referred to as the environment adaptive noise model) separately from the recognition word to be recognized. The feature amount of the input speech is stored and compared with the recognized word and the environment adaptive noise model, and the similarity is calculated. Then, when the similarity of the recognition word having the highest similarity is higher than the similarity of the environment adaptive noise model, the speech recognition device of the conventional technique 1 outputs the recognition word having the highest similarity as a recognition result.

  FIG. 18A and FIG. 18B are diagrams showing a range (hereinafter referred to as a detection range) BD2 in which the recognition word “Okayama” in Conventional Technique 1 is detected. FIG. 18 shows an example in which the environment adaptive noise model of the prior art 1 is regarded as a rejection word, the recognition word is “Okayama”, and the rejection words are “Wakayama” and “Toyama”. In the method of the prior art 1, as shown in FIG. 18, even if the threshold value (hereinafter referred to as the first threshold value) T1 is adjusted, the detection range BD2 is an ideal range in which a high reproduction rate and a high matching rate can be obtained. It is difficult to approach BD1 (hereinafter referred to as an ideal range). SC1 in the figure indicates the degree of similarity of the recognition word “Okayama” with respect to the utterance content. SC2 and SC3 in the figure respectively indicate the similarity of the reject word “Wakayama” and the similarity of the reject word “Toyama” with respect to the utterance content.

  When the first threshold T1 is lowered, as shown in FIG. 18A, the range in which the similarity SC1 of the recognized word “Okayama” is higher than the similarity of the rejected words (SC2 and SC3) is widened, so that a high recall is obtained. Although it becomes possible, the precision is deteriorated. On the other hand, when the first threshold value T1 is increased, as shown in FIG. 18B, the relevance ratio can be improved. However, the words that should be recognized cannot be covered and the reproduction ratio is deteriorated.

  In addition, the speech recognition device proposed in Patent Document 2 (hereinafter referred to as the speech recognition device of the prior art 2) stores a second threshold T2 for a recognized word separately from the first threshold T1, and the input speech The similarity between the feature quantity and each recognized word, the feature quantity of the input speech, and the similarity between each reject word are calculated. Then, the speech recognition apparatus according to the related art 2 has the highest similarity of the recognized word that is higher than the second threshold T2, or the highest similarity of the recognized word has the second threshold T2 or less. However, when the similarity is higher than the rejection word having the highest similarity with the first threshold T1, the recognition word having the highest similarity is output as the recognition result.

  19A and 19B are diagrams showing a detection range BD2 of the recognition word “Okayama” in the related art 2. In the method of the prior art 2, as shown in FIG. 19, by adjusting the second threshold T2, the detection range BD2 can be brought closer to the ideal range BD1 as compared with the case of the prior art 1, but it is still sufficient. is not.

  When the second threshold value T2 is lowered, as shown in FIG. 19A, the range in which the similarity SC1 of the recognized word “Okayama” is higher than the second threshold value T2 is widened, so that a high recall can be obtained. The precision will deteriorate. On the other hand, even if the second threshold T2 is increased, as shown in FIG. 19B, even when the similarity SC1 of the recognition word “Okayama” is equal to or lower than the second threshold T2, the highest similarity is obtained with the first threshold T1. When the similarity is higher than the rejection word similarity (SC2 or SC3), the recognition word “Okayama” is used as the recognition result, and therefore the relevance rate can only be improved to some extent.

  As described above, in the related arts 1 and 2, it may be very difficult to adjust the detection range BD2 so as to approach the ideal range BD1 in which a high reproduction rate and a high matching rate are obtained.

JP 2003-202887 A JP 2008-129263 A

  In one aspect, an object of the present invention is to provide a speech recognition device, a speech recognition method, and a program that can realize speech recognition that can achieve a high reproduction rate and a high matching rate.

The speech recognition apparatus according to an aspect includes a recognition word dictionary in which recognition words are registered and a rejection word dictionary in which rejection words are registered, calculates a similarity between an input speech and a recognition word, and has the highest similarity. A speech recognition device that rejects input speech when the similarity of a high recognition word is equal to or lower than a first threshold value set in advance, and calculates the similarity between the input speech and the rejection word, and has the highest similarity A first specifying means for specifying a reject word, a second specifying means for calculating the similarity between the input speech and the recognized word, and specifying the recognized word having the highest similarity, and the similarity of the specified reject word Is over a third threshold that is greater than the first threshold set in advance, the input speech is rejected even if the similarity of the identified recognition word exceeds the first threshold comprising a checking means, the collation means, recognition identified If the similarity of the word exceeds a second threshold greater than the first threshold, the identified recognition regardless of whether the similarity of the identified rejected word exceeds the third threshold or not It is characterized by outputting words .

A speech recognition method according to an aspect includes a recognition word dictionary in which recognition words are registered and a rejection word dictionary in which rejection words are registered, calculates a similarity between an input speech and a recognition word, and has the highest similarity. A speech recognition method of a speech recognition apparatus that rejects input speech when the similarity of a high recognition word is equal to or less than a first threshold value set in advance, and calculates the similarity between the input speech and the rejection word, identify the degree of similarity is high reject words, to calculate the degree of similarity between the input speech and the recognized word to identify the highest similarity recognition words, the similarity of the reject words specified, are previously set, If the third threshold value that is greater than the first threshold value is exceeded, the input speech is rejected even if the similarity level of the identified recognition word exceeds the first threshold value, and the similarity level of the identified recognition word Exceeds a second threshold greater than the first threshold Case, regardless of whether the similarity of the reject words identified exceeds a third threshold value, and outputs a recognized word specified is characterized in that.

The program according to one aspect includes a recognition word dictionary in which recognition words are registered and a rejection word dictionary in which rejection words are registered, calculates a similarity between an input voice and a recognition word, and recognizes the highest similarity The similarity of the input speech and the reject word is calculated in the computer of the speech recognition device that rejects the input speech when the word similarity is equal to or less than a first threshold value set in advance, and the reject word with the highest similarity is calculated identify and calculate the degree of similarity between the input speech and the recognized word to identify the highest similarity recognition words, the similarity of the reject words identified, are preset, larger than the first threshold value When the third threshold is exceeded, even if the similarity of the identified recognition word exceeds the first threshold, the input speech is rejected, and the similarity of the identified recognition word is the first threshold. If a larger second threshold is exceeded , Regardless of whether the similarity of the reject words identified exceeds a third threshold value, and outputs a recognized word specified it is characterized in that to execute the process.

  In one aspect, it is possible to realize speech recognition that provides a high recall and a high precision.

A is a functional block diagram illustrating a configuration example of the speech recognition apparatus according to the first embodiment, and B is a functional block diagram illustrating a configuration example of a storage unit according to the first embodiment. It is a figure which shows the structural example of the recognition word dictionary in Embodiment 1. FIG. It is a figure which shows the structural example of the rejection word dictionary in Embodiment 1. FIG. It is a figure which shows the range from which the recognition word in Embodiment 1 is detected. 3 is a part of an example of a flowchart for explaining a flow of a speech recognition process in the first embodiment. 6 is another part of an example of a flowchart for explaining the flow of the speech recognition process in the first embodiment. FIG. 4A is a diagram showing the recall rate and the matching rate for each first threshold value in the prior art 1, and B is a diagram showing the recall rate and the matching rate for each third threshold value in the first embodiment. It is a functional block diagram which shows the structural example of the speech recognition apparatus in Embodiment 2. It is a figure which shows the structural example of the rejection word dictionary in Embodiment 2. FIG. FIG. 10 is a diagram for explaining a third threshold value in the second embodiment. It is a figure which shows the range from which the recognition word in Embodiment 2 is detected. 10 is a part of an example of a flowchart for explaining a flow of speech recognition processing in the second embodiment. It is a figure which shows the range from which the recognition word in Embodiment 3 is detected. 10 is a part of an example of a flowchart for explaining a flow of voice recognition processing in the third embodiment. FIG. 7A is a diagram showing the recall rate and the matching rate for each second threshold value in Conventional Technology 2, and B is a diagram showing the recall rate and the matching rate for each third threshold value in the third embodiment. 14 is a part of an example of a flowchart for explaining a flow of voice recognition processing in the fourth embodiment. It is a figure which shows the example of the hardware constitutions of the speech recognition apparatus in embodiment. A and B are both diagrams showing a range in which a recognized word in the prior art 1 is detected. A and B are both diagrams showing a range in which a recognized word is detected in the related art 2.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

(Embodiment 1)
FIG. 1A is a functional block diagram illustrating a configuration example of the speech recognition apparatus 1 according to the first embodiment, and FIG. 1B is a functional block diagram illustrating a configuration example of the storage unit 20 according to the first embodiment. The speech recognition apparatus 1 according to the first exemplary embodiment compares the feature amount of the input speech with a recognition word registered in advance and a rejection word, and rejects the input speech according to the degree of similarity calculated, or the most similar It is a device that outputs recognition words with high degrees. As shown in FIG. 1, the voice recognition device 1 according to the first exemplary embodiment includes an input unit 10, a storage unit 20, an output unit 30, and a control unit 40.

  The input unit 10 includes, for example, an input / output interface and receives a signal including an audio section (hereinafter referred to as an input signal) from a connected audio acquisition device (for example, a microphone). Then, the input unit 10 outputs the received input signal to the control unit 40. At this time, the input unit 10 temporarily stores the received input signal in a buffer memory (not shown), and the control unit 40 sequentially acquires the input signal from the buffer memory in units of frames in accordance with the processing timing. May be. In the following, the signal in the voice section of the input signal is referred to as a voice signal.

  The storage unit 20 includes, for example, a random access memory (RAM), a read only memory (ROM), and a hard disk drive (HDD). The storage unit 20 includes, for example, a work area of a central processing unit (CPU) provided in the control unit 40, a program area that stores various programs such as an operation program for controlling the entire speech recognition apparatus 1, and is conventionally used. It functions as a data area for storing various data such as the first threshold T1. This first threshold value T1 is used as one measure for rejecting the input speech, as is conventional.

  Moreover, the memory | storage part 20 functions as the recognition word dictionary 21, the rejection word dictionary 22, and the acoustic model memory | storage part 23, as shown to FIG. 1B.

  The recognition word dictionary 21 stores a plurality of recognition words and information related to the recognition words. FIG. 2 is a diagram illustrating a configuration example of the recognized word dictionary 21 according to the first embodiment. For example, as shown in FIG. 2, the recognition word dictionary 21 stores a plurality of recognition words and information related to the recognition words in a table form. As shown in FIG. 2, the recognized word dictionary 21 according to the first embodiment associates “word reading” with “phoneme series” for each “word notation” of the recognized word. The recognized word may include a recognized vocabulary.

  “Word notation” is information describing the corresponding recognition word. The “word reading” is information representing the corresponding recognition word in hiragana. The “phoneme series” is information representing the corresponding recognition word in phonemes, and is used when calculating the similarity with the feature amount of the input speech signal.

  The reject word dictionary 22 stores a plurality of reject words and information related to the reject words. FIG. 3 is a diagram illustrating a configuration example of the reject word dictionary 22 according to the first embodiment. For example, as shown in FIG. 3, the reject word dictionary 22 stores a plurality of reject words and information related to the reject words in a table. As shown in FIG. 3, the reject word dictionary 22 according to the first exemplary embodiment associates “word reading”, “phoneme series”, and “third threshold” for each “word notation” of the reject word. It is. The rejection word may include a rejection vocabulary.

  The “word notation” is information describing the corresponding reject word. “Reading a word” is information representing the corresponding reject word in hiragana. The “phoneme series” is information representing the corresponding reject word in phonemes, and is used when calculating the similarity with the feature amount of the input speech signal. The “third threshold value” is a threshold value of the corresponding rejection word, and is referred to as a third threshold value T3 in order to distinguish the first threshold value T1 and the second threshold value T2 described above. This third threshold value T3 is larger than the first threshold value T1 and is experimentally determined and set in advance so that if the value of the third threshold value T3 is exceeded, the corresponding rejection word is likely to be uttered. .

  Note that the third threshold T3 is not set for each rejection word, but may be a common threshold for all rejection words.

  The acoustic model storage unit 23 stores data obtained by statistically modeling the feature amount tendency for each phoneme. Examples of the acoustic model include, but are not limited to, Hidden Markov Model (HMM), Dynamic Programming (DP) matching, and the like.

  The output unit 30 includes, for example, an input / output interface and the like, and outputs a recognition result by the collation unit 42 (described later in detail). For example, when an information providing device (not shown) is connected, information based on the recognition result is provided to the user. Further, the recognition result may be passed to the upper program or may be output as a command to the CPU.

  The control unit 40 includes a CPU, for example, and executes an operation program stored in the program area of the storage unit 20 to function as an analysis unit 41 and a verification unit 42 as shown in FIG. 1A. Is realized. Further, the control unit 40 executes an operation program to execute a control process for controlling the entire voice recognition apparatus 1 and a process such as a voice recognition process described later in detail.

  The analysis unit 41 analyzes the input signal to detect a speech section, and further acoustically analyzes the speech signal that is a signal in the speech section to calculate a feature amount of the speech signal. Then, the analysis unit 41 outputs the calculated feature amount of the audio signal to the collation unit 42.

  The feature amount is not limited to these, but a Mel Frequency Cepstial Coefficient (MFCC), a linear predictive coding (LPC) cepstrum, a power spectrum, or the like can be used.

  The collation unit 42 calculates the similarity between the input speech and each recognized word stored in the recognized word dictionary 21 based on the input feature amount. The collation unit 42 also calculates the similarity between the input speech and each rejection word stored in the rejection word dictionary 22 based on the input feature amount.

  More specifically, the collation unit 42 compares the acoustic model stored in the acoustic model storage unit 23 with the input feature quantity, and extracts a phoneme string corresponding to the input speech. Then, the matching unit 42 calculates the similarity by comparing the extracted phoneme sequence and the phoneme sequence of each recognized word, and calculates the similarity by comparing the extracted phoneme sequence and the phoneme sequence of each rejected word. To do. Note that a conventionally used method can be used as the similarity calculation method. In the first embodiment, the similarity is normalized to a value between 0 and 100.

  And the collation part 42 specifies the rejection word with the highest similarity, and determines whether the similarity of the specified rejection word exceeds the corresponding 3rd threshold value T3. And when it determines with the similarity of the specified rejection word exceeding 3rd threshold value T3, the collation part 42 rejects the audio | voice signal of collation object. On the other hand, when it is determined that the similarity of the specified rejection word is equal to or less than the third threshold T3, the matching unit 42 specifies the recognized word having the highest similarity.

  And the collation part 42 determines whether the similarity of the identified recognition word is over 1st threshold value T1. And when it determines with the similarity of the identified recognition word being below 1st threshold value T1, the collation part 42 rejects the audio | voice signal of collation object. On the other hand, when it is determined that the similarity of the identified recognition word exceeds the first threshold T1, the collation unit 42 further exceeds the similarity of the rejection word that is identified by the similarity of the identified recognition word. It is determined whether or not.

  And when it determines with the similarity of the identified recognition word exceeding the similarity of the identified rejection word, the collation part 42 outputs via the output part 30 by making the identified recognition word into a recognition result. To do. On the other hand, when it determines with the similarity of the identified recognition word being below the similarity of the identified rejection word, the collation part 42 rejects the audio | voice signal of collation object.

  Next, with reference to FIG. 4, the process of the collation part 42 in this Embodiment 1 is further demonstrated according to a specific example. FIG. 4 is a diagram showing a range in which the recognition word “Okayama” is detected in the first embodiment. The example of FIG. 4 is an example when the input speech is compared with the recognized word “Okayama”, and is an example when the third threshold T3 is common to all rejected words.

  Referring to FIG. 4, if the user utters, for example, “△ ▲ Yama” (including the case where it can be heard as a result of noise from the surrounding environment), “△ ▲ Yama” is recognized as “OKyama”. Is lower than the similarity SC2 between the “△ ▲ yama” and the rejection word “Wakayama” (in this example, the similarity SC2 of the rejection word “Wakayama” is the highest among the rejection words). In other words, “Yama” is in the range to the left of the dotted line L1 in which the similarity SC1 of the recognition word “Okayama” is less than or equal to the similarity SC2 of the rejection word “Wakayama”. Therefore, in this case, the collation unit 42 rejects “Δ ▲ Yama”.

  For example, if the user utters “X Kayama”, the similarity SC1 between “X Kayama” and the recognized word “Okayama” is similar to “X Kayama” and the rejection word “Wakayama”. It is larger than SC2 (in this example, the similarity SC2 of the reject word “Wakayama” is the highest among the reject words). However, as shown in FIG. 4, the similarity SC2 of the rejection word “Wakayama” exceeds the third threshold T3. That is, “× Kayama” has the similarity SC1 of the recognition word “Okayama” greater than the similarity SC2 of the rejection word “Wakayama”, but the similarity SC2 of the rejection word “Wakayama” exceeds the third threshold T3. In the range between the dotted line L1 and the dotted line L2. Therefore, in this case, the collation unit 42 rejects “X Kayama”.

  For example, if the user utters “Oyama”, the similarity SC1 between “Oyama” and the recognition word “Okayama” is the similarity SC3 between “Oyama” and the rejection word “Toyama” (this In the example, the similarity SC3 of the reject word “Toyama” is the highest among the reject words). That is, “Oyama” is in the range to the right of the dotted line L3, which is the range in which the similarity SC1 of the recognition word “Okayama” is less than or equal to the similarity SC3 of the rejection word “Toyama”. Therefore, in this case, the collation unit 42 rejects “Oyama”.

  For example, if the user utters “□ Kayama”, as shown in FIG. 4, the similarity SC3 between “□ Kayama” and the rejection word “Toyama” (in this example, the rejection word “Toyama”). “Similarity SC3 of“ is the highest among rejected words ”is equal to or less than the third threshold T3. Further, the similarity SC1 between “□ Kayama” and the recognized word “Okayama” is larger than the similarity SC3 of the reject word “Toyama” and the first threshold T1, as shown in FIG. That is, in “□ Kayama”, the similarity (SC2 or SC3) of the rejection word having the highest similarity does not exceed the third threshold T3, and the similarity SC1 of the recognition word “Okayama” has the highest similarity. The rejection word similarity (SC2 or SC3) is in a range larger than the first threshold T1, and is in a range between the dotted line L2 and the dotted line L3. In this specific example, the range between the dotted line L2 and the dotted line L3 is the detection range BD2 of the recognition word “Okayama”. Therefore, in this case, the collation unit 42 outputs the recognition word “Okayama” as a recognition result.

  Next, the flow of speech recognition processing in the first embodiment will be described with reference to FIGS. 5 and 6 are a part of an example of a flowchart for explaining the flow of the speech recognition processing in the first embodiment and another part, respectively. The voice recognition process is started with the input signal received by the input unit 10 being output to the control unit 40 as a trigger.

  The analysis unit 41 analyzes the input signal and detects an audio signal (step S001). Then, the analysis unit 41 calculates the feature amount of the detected audio signal, and outputs the calculated feature amount to the collation unit 42 (step S002).

  Then, the collation unit 42 calculates the similarity between the speech signal to be collated and each rejection word based on the input feature amount (step S003), and identifies the rejection word with the highest similarity (step S003). S004). And the collation part 42 determines whether the similarity of the specified rejection word exceeds the corresponding 3rd threshold value T3 (step S005).

  When it is determined that the identified rejection word similarity exceeds the third threshold T3 (step S005; YES), the collation unit 42 rejects the collation target speech signal (step S006), finish.

  On the other hand, when it is determined that the similarity of the specified rejection word is equal to or smaller than the third threshold T3 (step S005; NO), the collation unit 42 determines whether the collation target voice signal is based on the input feature amount. The degree of similarity with each recognized word is calculated (step S007), and the recognized word with the highest degree of similarity is specified (step S008). And the collation part 42 determines whether the similarity degree of the identified recognition word is over 1st threshold value T1 (step S009).

  If it is determined that the similarity of the identified recognition word is equal to or less than the first threshold T1 (step S009; NO), the collation unit 42 rejects the speech signal to be collated (step S006), and this process ends. To do. On the other hand, when it is determined that the similarity of the identified recognition word exceeds the first threshold T1 (step S009; YES), the collation unit 42 further rejects the similarity identified by the similarity of the identified recognition word. It is determined whether or not the similarity is exceeded (step S010).

  When it is determined that the similarity of the identified recognition word is equal to or less than the similarity of the specified rejection word (step S010; NO), the collation unit 42 rejects the speech signal to be collated (step S006), The process ends. On the other hand, if it is determined that the similarity of the identified recognition word exceeds the similarity of the specified rejection word (step S010; YES), the collation unit 42 uses the identified recognition word as a recognition result as an output unit. 30 is output (step S011). Then, this process ends.

  Next, effects of the first embodiment will be described with reference to FIG. FIG. 7A is a diagram showing the recall rate and the matching rate for each first threshold value T1 in the prior art 1, and FIG. 7B is a diagram showing the recall rate and the matching rate for each third threshold value T3 in the first embodiment. It is. Note that each value of the recall rate and the matching rate is a value based on the evaluation result with respect to the speech data for 300 words and about one hour.

  When the first threshold value T1 is increased from the reference value (T1 = 80) in the prior art 1, as shown in FIG. 7A, when the first threshold value T1 = 88, the recall rate is 84% and the matching rate is When the first threshold value T1 = 90, the recall rate is 81% and the matching rate is 86%. That is, in the case of the prior art 1, when the first threshold value T1 is increased from the reference value (T1 = 80), the matching rate can be increased, but the reproduction rate is deteriorated.

  On the other hand, when the method of the first embodiment is used and the third threshold T3 = 95, the recall is 85% and the matching ratio is 83%, and when the third threshold T3 = 94, the recall is 84% and the precision is 87%. That is, by adjusting the third threshold value T3, it is possible to improve the matching rate while maintaining the recall rate substantially. When evaluating the method of the first embodiment, the first threshold value T1 is fixed to the reference value (T1 = 80).

  According to the first embodiment, the speech recognition apparatus 1 calculates the feature amount of the input speech, calculates the similarity between the input speech and each rejected word based on the calculated feature amount of the input speech, and is the most similar Identify reject words with high degrees. Then, the speech recognition apparatus 1 rejects the input speech when the similarity of the identified rejection word exceeds the corresponding third threshold value T3. By doing so, in the range where the similarity of the recognition word with the highest similarity exceeds the first threshold T1, the similarity of the rejection word with the highest similarity exceeds the first threshold T1, and thus is erroneously detected. When a large number of occurrences occur, it is possible to prevent detection of rejected words that are erroneously detected by adjusting the third threshold value T3. That is, by adjusting the third threshold T3 when a false detection occurs due to a high correlation between the recognized word with the highest similarity and the recognized word with the highest similarity, It is possible to prevent erroneous detection. Therefore, it is possible to improve the matching rate while maintaining the recall rate.

  Moreover, according to the said Embodiment 1, the speech recognition apparatus 1 stores 3rd threshold value T3 for every rejection word. By doing so, it becomes possible to experimentally obtain a value that increases the possibility that the corresponding rejection word is uttered when the value is exceeded for each rejection word, and set it as the third threshold value T3. Therefore, fine adjustment is possible for each rejected word, and the detection accuracy can be further improved.

(Embodiment 2)
In the first embodiment, the third threshold T3 is set in advance. In the second embodiment, an example in which the third threshold value T3 is calculated will be described.

  FIG. 8 is a functional block diagram illustrating a configuration example of the speech recognition apparatus 1 according to the second embodiment. The basic configuration of the speech recognition apparatus 1 in the second embodiment is the same as that in the first embodiment. However, the configuration of the reject word dictionary 22 is slightly different from that in the first embodiment. Further, as shown in FIG. 8, the control unit 40 further includes a threshold value calculation unit 43, which is different from the case of the first embodiment.

  FIG. 9 is a diagram illustrating a configuration example of the reject word dictionary 22 according to the second embodiment. The configuration of the reject word dictionary 22 in the second embodiment is different from the configuration of the reject word dictionary 22 in the first embodiment in that the “third threshold” is not stored as shown in FIG. 9. This is because, as described above, in the second embodiment, the threshold calculation unit 43 calculates the third threshold T3.

  Returning to FIG. 8, the control unit 40 includes, for example, a CPU, executes an operation program stored in the program area of the storage unit 20, and collates with the analysis unit 41 as illustrated in FIG. 8. The function as the unit 42 and the threshold value calculation unit 43 is realized. Further, the control unit 40 executes an operation program to execute a control process for controlling the entire voice recognition apparatus 1 and a process such as a voice recognition process described later in detail.

  The threshold calculation unit 43 calculates the third threshold T3 of the reject word with the highest similarity. More specifically, the threshold calculation unit 43 calculates the inter-word similarity between the reject word with the highest similarity and the recognized word with the highest similarity. The similarity between words is not limited to these, but can be determined based on, for example, the editing distance of “word reading” or “phoneme series”. When the similarity between words is determined based on the editing distance, the similarity between words is set higher as the editing distance is shorter.

  For example, the similarity between words is determined based on the editing distance of “word reading”. The recognition word with the highest similarity and the rejection word with the highest similarity are “Okayama” and “Toyama”, respectively. To do. In this case, the “word reading” of the recognition word “Okayama” is “Okayama” and the “word reading” of the rejection word “Toyama” is “Toyama”, so to edit “Toyama” to “Okayama” Needs to add “o” and convert “to” to “ka”. That is, the edit distance in this case is “2”. On the other hand, if the rejection word with the highest degree of similarity is “Wakayama”, the “word reading” of the rejection word “Wakayama” is “Wakayama”, so to edit “Wakayama” to “Okayama” You just need to convert it to “O”. That is, the edit distance in this case is “1”.

  Therefore, the similarity between words of the recognition word “Okayama” and the rejection word “Toyama” is smaller than the similarity between words of the recognition word “Okayama” and the rejection word “Wakayama”.

  Then, as shown in FIG. 10, the threshold value calculation 43 calculates the third threshold value T3 based on the function F that is defined such that the higher the inter-word similarity, the higher the value of the third threshold value T3. This is because if the similarity between words is high, the similarity of rejected words with high similarity between words tends to be high for utterances with high similarity of recognized words. It is because it will suppress too much. On the other hand, if the similarity between words is low, the similarity of the rejected word is unlikely to be high with respect to the utterance in which the similarity of the recognized word is high. Because there is no. As shown in FIG. 10, the third threshold value T3 is set to be equal to or greater than a predetermined value because if the third threshold value T3 is too low, the detection range BD2 becomes too narrow, that is, it is excessively suppressed. Because it ends up. Here, FIG. 10 is a diagram for explaining the third threshold value T3 (function F) in the second embodiment.

For example, the function F may be as shown in the following Expression 1. Note that T3_MIN in Equation 1 is the lowest value of the third threshold value T3 for preventing excessive suppression.
Third threshold = MAX (100−edit distance × 2, T3_MIN) (Expression 1)

  Note that T3_MIN may be determined for each combination of the recognition word for calculating the similarity between words and the rejection word, or may be common to all combinations. For example, T3_MIN may be set to be lower as the similarity between words is lower.

  Next, with reference to FIG. 11, processing of the matching unit 42 and the threshold value calculation unit 43 in the second embodiment will be described according to a specific example. FIG. 11 is a diagram illustrating a range in which the recognition word “Okayama” is detected in the second embodiment. In addition, the example of FIG. 11 is an example in the case of comparing the input speech with the recognition word “Okayama”. Further, T3-W1 in FIG. 11 is a third threshold value when the combination of the recognition word for calculating the similarity between words and the rejection word is the recognition word “Okayama” and the rejection word “Wakayama”. W2 is a third threshold value when the combination of the recognition word for calculating the similarity between words and the rejection word is the recognition word “Okayama” and the rejection word “Toyama”.

  In this specific example, as described above, the combination of the recognition word “Okayama” and the rejection word “Wakayama” has a higher degree of similarity between words than the combination of the recognition word “Okayama” and the rejection word “Toyama”. high. Therefore, as shown in FIG. 11, the third threshold value T3 calculated by the threshold value calculation unit 43 is greater than the third threshold value T3-W1 for the rejection word “Wakayama” than the third threshold value T3-W2 for the rejection word “Toyama”. Will be higher.

  Referring to FIG. 11, if the user utters, for example, “△ Yama” (including the case where it can be heard as a result of the influence of ambient environmental noise), “△ ▲ Yama” and the recognized word “Okayama” Is lower than the similarity SC2 between the “△ ▲ yama” and the rejection word “Wakayama” (in this example, the similarity SC2 of the rejection word “Wakayama” is the highest among the rejection words). In other words, “Yama” is in the range to the left of the dotted line L1 in which the similarity SC1 of the recognition word “Okayama” is less than or equal to the similarity SC2 of the rejection word “Wakayama”. Therefore, in this case, the collation unit 42 rejects “Δ ▲ Yama”.

  For example, if the user utters “X Kayama”, the similarity SC1 between “X Kayama” and the recognized word “Okayama” is similar to “X Kayama” and the rejection word “Wakayama”. It is larger than SC2 (in this example, the similarity SC2 of the reject word “Wakayama” is the highest among the reject words). However, as shown in FIG. 11, the similarity SC2 of the rejection word “Wakayama” exceeds the third threshold value T3-W1. That is, “× Kayama” has a similarity SC1 of the recognition word “Okayama” greater than a similarity SC2 of the rejection word “Wakayama”, but a similarity SC2 of the rejection word “Wakayama” exceeds the third threshold T3-W1. Is in the range between the dotted line L1 and the dotted line L2. Therefore, in this case, the collation unit 42 rejects “X Kayama”.

  For example, if the user utters “Oyama”, the similarity SC1 between “Oyama” and the recognition word “Okayama” is the similarity SC3 between “Oyama” and the rejection word “Toyama” (this In the example, the similarity SC3 of the reject word “Toyama” is the highest among the reject words). That is, “Oyama” is in the range to the right of the dotted line L3, which is the range in which the similarity SC1 of the recognition word “Okayama” is less than or equal to the similarity SC3 of the rejection word “Toyama”. Therefore, in this case, the collation unit 42 rejects “Oyama”.

  For example, if the user utters “□ Kayama”, the similarity SC1 between “□ Kayama” and the recognition word “Okayama” is similar to “□ Kayama” and the rejection word “Toyama”. It is larger than SC3 (in this example, the similarity SC3 of the rejection word “Toyama” is the highest among the rejection words). However, as shown in FIG. 11, the similarity SC3 of the rejection word “Okayama” exceeds the third threshold T3-W2. In other words, “□ Kayama” has a similarity SC1 of the recognition word “Okayama” greater than the similarity SC3 of the rejection word “Toyama”, but the similarity SC3 of the rejection word “Toyama” exceeds the third threshold T3-W2. Is in the range between the dotted line L4 and the dotted line L3. Therefore, in this case, the collation unit 42 rejects “□ Kayama”.

  For example, if the user utters “■ Kayama”, as shown in FIG. 11, the similarity SC3 between “■ Kayama” and the rejection word “Okayama” (in this example, the rejection word “Toyama”). “Similarity SC3 of“ is the highest among rejected words ”is equal to or less than the third threshold value T3-W2. Further, the similarity SC1 between “■ Kayama” and the recognized word “Okayama” is larger than the similarity SC3 of the reject word “Toyama” and the first threshold T1, as shown in FIG. That is, “■ Kayama” is similar to the recognition word “Okayama” because the similarity (SC2 or SC3) of the rejection word having the highest similarity does not exceed the third threshold (T3-W1 or T3-W2). The degree SC1 is in the range between the similarity (SC2 or SC3) of the reject word having the highest similarity and the first threshold T1, and the range between the dotted line L2 and the dotted line L4. In this specific example, the range between the dotted line L2 and the dotted line L4 is the detection range BD2 of the recognition word “Okayama”. Therefore, in this case, the collation unit 42 outputs the recognition word “Okayama” as a recognition result.

  In this way, by setting the third threshold value T3 for rejected words according to the similarity between words, the detection range BD2 can be made closer to the ideal range BD1.

  Next, with reference to FIG. 12, the flow of speech recognition processing in the second embodiment will be described. FIG. 12 is a part of an example of a flowchart for explaining the flow of speech recognition processing in the twelfth embodiment. The voice recognition process is started with the input signal received by the input unit 10 being output to the control unit 40 as a trigger. The processing after step S109 is the same as that in the first embodiment.

  The analysis unit 41 analyzes the input signal and detects an audio signal (step S101). And the analysis part 41 calculates the feature-value of the detected audio | voice signal, and outputs the calculated feature-value to the collation part 42 (step S102).

  Then, the collation unit 42 calculates the similarity between the speech signal to be collated and each rejection word based on the input feature amount (step S103), and identifies the rejection word with the highest similarity (step S103). S104). Then, the collation unit 42 calculates the similarity between the speech signal to be collated and each recognition word based on the input feature amount (step S105), and identifies the recognition word having the highest similarity (step S105). S106).

  Then, the threshold calculation unit 43 calculates the inter-word similarity between the specified rejection word and the specified recognition word (step S107), and the specified rejection word based on the calculated inter-word similarity. The third threshold value T3 is calculated (step S108).

  Then, the collation unit 42 determines whether or not the similarity of the identified rejection word exceeds the calculated third threshold T3 (step S109). When it is determined that the similarity of the identified rejection word exceeds the calculated third threshold T3 (step S109; YES), the collation unit 42 rejects the collation target speech signal (step S006), This process ends.

  On the other hand, when it is determined that the similarity of the specified rejection word is equal to or less than the calculated third threshold T3 (step S109; NO), the process proceeds to the process after step S009 described in the first embodiment. The collation unit 42 determines whether the similarity of the identified recognition word exceeds the first threshold T1 (step S009).

  If it is determined that the similarity of the identified recognition word is equal to or less than the first threshold T1 (step S009; NO), the collation unit 42 rejects the speech signal to be collated (step S006), and this process ends. To do. On the other hand, when it is determined that the similarity of the identified recognition word exceeds the first threshold T1 (step S009; YES), the collation unit 42 further rejects the similarity identified by the similarity of the identified recognition word. It is determined whether or not the similarity is exceeded (step S010).

  When it is determined that the similarity of the identified recognition word is equal to or less than the similarity of the specified rejection word (step S010; NO), the collation unit 42 rejects the speech signal to be collated (step S006), The process ends. On the other hand, if it is determined that the similarity of the identified recognition word exceeds the similarity of the specified rejection word (step S010; YES), the collation unit 42 uses the identified recognition word as a recognition result as an output unit. 30 is output (step S011). Then, this process ends.

  According to the second embodiment, the speech recognition apparatus 1 calculates the inter-word similarity (similarity between words) between the recognition word with the highest similarity and the rejection word with the highest similarity, and the calculated inter-word The higher the similarity is, the higher the third threshold value T3 of the corresponding rejection word (the rejection word with the highest similarity) is set. In this way, since the third threshold T3 can be adjusted according to the correlation between the reject word with the highest similarity and the recognized word with the highest similarity, it is possible to prevent excessive suppression. Therefore, when improving the relevance rate, it is possible to prevent the recall rate from deteriorating too much.

  Further, according to the second embodiment, the speech recognition apparatus 1 sets the third threshold value T3 to a predetermined value or more. Thereby, even if the similarity between words of the recognition word with the highest similarity and the rejection word with the highest similarity is very low, the third threshold T3 is set to a predetermined value or more to suppress the suppression. It becomes possible to prevent excessively.

(Embodiment 3)
In the third embodiment, an example in which the method described in the first embodiment is applied to the related art 2 will be described. Of course, the method described in the second embodiment can be applied to the related art 2.

  The basic configuration of the speech recognition apparatus 1 in the third embodiment is the same as that in the first embodiment. However, the second threshold value T2 corresponding to the recognition word described above is further stored in the data area of the storage unit 20, which is different from the case of the first embodiment. Further, the function performed by the collation unit 42 is slightly different from that in the first embodiment. Note that the second threshold T2 stored in the data area may be a second threshold T2 that is different for each recognized word, or may be a second threshold T2 that is common to all recognized words.

  The collation unit 42 according to the third embodiment compares the acoustic model stored in the acoustic model storage unit 23 with the input feature amount, and extracts a phoneme string corresponding to the input speech. Then, the matching unit 42 calculates the similarity by comparing the extracted phoneme sequence and the phoneme sequence of each recognized word, and calculates the similarity by comparing the extracted phoneme sequence and the phoneme sequence of each rejected word. To do. Note that a conventionally used method can be used as the similarity calculation method. In the third embodiment, the similarity is normalized to a value between 0 and 100.

  And the collation part 42 specifies the rejection word with the highest similarity, and determines whether the similarity of the specified rejection word exceeds the corresponding 3rd threshold value T3. And when it determines with the similarity of the specified rejection word exceeding 3rd threshold value T3, the collation part 42 rejects the audio | voice signal of collation object. On the other hand, when it is determined that the similarity of the specified rejection word is equal to or less than the third threshold T3, the matching unit 42 specifies the recognized word having the highest similarity.

  Then, the collation unit 42 determines whether or not the similarity of the identified recognition word exceeds the corresponding second threshold value T2. If it is determined that the similarity of the identified recognition word exceeds the corresponding second threshold T2, the collation unit 42 outputs the identified recognition word as a recognition result via the output unit 30. . On the other hand, when it is determined that the similarity of the identified recognition word is equal to or less than the corresponding second threshold T2, the collation unit 42 further determines whether the similarity of the identified recognition word exceeds the first threshold T1. Determine whether. And when it determines with the similarity of the identified recognition word being below 1st threshold value T1, the collation part 42 rejects the audio | voice signal of collation object. On the other hand, when it is determined that the similarity of the identified recognition word exceeds the first threshold T1, the collation unit 42 further exceeds the similarity of the rejection word that is identified by the similarity of the identified recognition word. It is determined whether or not.

  And when it determines with the similarity of the identified recognition word exceeding the similarity of the identified rejection word, the collation part 42 outputs via the output part 30 by making the identified recognition word into a recognition result. To do. On the other hand, when it determines with the similarity of the identified recognition word being below the similarity of the identified rejection word, the collation part 42 rejects the audio | voice signal of collation object.

  Next, with reference to FIG. 13, the process of the collation part 42 in this Embodiment 3 is further demonstrated according to a specific example. FIG. 13 is a diagram illustrating a range in which the recognition word “Okayama” is detected in the third embodiment. Note that the example of FIG. 13 is an example in which the input speech is compared with the recognition word “Okayama”, and is an example in which the third threshold T3 is common to all rejection words.

  Referring to FIG. 13, if the user utters, for example, “△ Yama” (including the case where it can be heard as a result of the influence of ambient environmental noise), “△ ▲ Yama” and the recognized word “Okayama” Is lower than the similarity SC2 between the “△ ▲ yama” and the rejection word “Wakayama” (in this example, the similarity SC2 of the rejection word “Wakayama” is the highest among the rejection words). In other words, “Yama” is in the range to the left of the dotted line L1 in which the similarity SC1 of the recognition word “Okayama” is less than or equal to the similarity SC2 of the rejection word “Wakayama”. Therefore, in this case, the collation unit 42 rejects “Δ ▲ Yama”. On the other hand, in the case of the prior art 2, the range between the dotted line L5 and the dotted line L1 to which “ΔYama” belongs is the similarity of the recognized word having the highest similarity (the similarity of the recognized word “Okayama”). Since SC1) exceeds the second threshold T2, the recognition word “Okayama” is output as the recognition result.

  For example, if the user utters “X Kayama”, the similarity SC1 between “X Kayama” and the recognized word “Okayama” is similar to “X Kayama” and the rejection word “Wakayama”. It is larger than SC2 (in this example, the similarity SC2 of the reject word “Wakayama” is the highest among the reject words). However, as shown in FIG. 13, the similarity SC2 of the reject word “Wakayama” exceeds the third threshold T3. That is, “× Kayama” has the similarity SC1 of the recognition word “Okayama” greater than the similarity SC2 of the rejection word “Wakayama”, but the similarity SC2 of the rejection word “Wakayama” exceeds the third threshold T3. In the range between the dotted line L1 and the dotted line L2. Therefore, in this case, the collation unit 42 rejects “X Kayama”. On the other hand, in the case of the prior art 2, the range between the dotted line L1 and the dotted line L2 to which “X Kayama” belongs is the similarity of the recognition word having the highest similarity (the similarity of the recognition word “Okayama”). Since SC1) exceeds the second threshold T2, the recognition word “Okayama” is output as the recognition result.

  For example, if the user utters “Oyama”, the similarity SC1 between “Oyama” and the recognition word “Okayama” is the similarity SC3 between “Oyama” and the rejection word “Toyama” (this In the example, the similarity SC3 of the reject word “Toyama” is the highest among the reject words). That is, “Oyama” is in the range to the right of the dotted line L3, which is the range in which the similarity SC1 of the recognition word “Okayama” is less than or equal to the similarity SC3 of the rejection word “Toyama”. Therefore, in this case, the collation unit 42 rejects “Oyama”.

  For example, if the user utters “□ Kayama”, as shown in FIG. 13, the similarity SC3 between “□ Kayama” and the rejection word “Toyama” (in this example, the rejection word “Toyama”). “Similarity SC3 of“ is the highest among rejected words ”is equal to or less than the third threshold T3. Further, the similarity SC1 between “□ Kayama” and the recognition word “Okayama” is equal to or lower than the second threshold T2, as shown in FIG. However, the similarity SC1 of the recognition word “Okayama” is larger than the similarity SC3 of the rejection word “Toyama” and the first threshold T1, as shown in FIG. That is, in “□ Kayama”, the similarity (SC2 or SC3) of the rejection word with the highest similarity does not exceed the third threshold T3, and the similarity SC1 of the recognition word “Okayama” with the highest similarity is Range of exceeding the second threshold T2 or similarity of the rejection word having the highest similarity (SC2 or SC3) does not exceed the third threshold T3 and the similarity of the recognized word “Okayama” having the highest similarity The degree SC1 is less than or equal to the second threshold T2, but the similarity SC1 of the recognized word “Okayama” has the highest similarity (SC2 or SC3) and the range greater than the first threshold T1, dotted line L2 and dotted line It is in the range between L3. In this specific example, the range between the dotted line L2 and the dotted line L3 is the detection range BD2 of the recognition word “Okayama”. Therefore, in this case, the collation unit 42 outputs the recognition word “Okayama” as a recognition result.

  As described above, by applying the method described in the first embodiment to the related art 2, it becomes possible to remove the range between the dotted line L5 and the dotted line L2 from the detection range BD2, which is compared with the case of the related art 2. Thus, the detection range BD2 can be brought closer to the ideal range BD1.

  Next, the flow of speech recognition processing in the third embodiment will be described with reference to FIG. FIG. 14 is a part of an example of a flowchart for explaining the flow of speech recognition processing in the third embodiment. The voice recognition process is started with the input signal received by the input unit 10 being output to the control unit 40 as a trigger. Note that the subsequent processing when “NO” is determined in the process of step S205 and the subsequent processing when “NO” is determined in the processing of step S208 are the same as those described in the first embodiment. The same.

  The analysis unit 41 analyzes the input signal and detects an audio signal (step S201). And the analysis part 41 calculates the feature-value of the detected audio | voice signal, and outputs the calculated feature-value to the collation part 42 (step S202).

  Then, the collation unit 42 calculates the similarity between the speech signal to be collated and each rejection word based on the input feature amount (step S203), and identifies the rejection word with the highest similarity (step S203). S204). And the collation part 42 determines whether the similarity of the specified rejection word exceeds the corresponding 3rd threshold value T3 (step S205).

  If it is determined that the similarity of the identified rejection word exceeds the third threshold T3 (step S205; YES), the process proceeds to the process of step S006 described in the first embodiment, The voice signal to be verified is rejected (step S006), and this process ends.

  On the other hand, when it is determined that the similarity of the specified rejection word is equal to or less than the third threshold T3 (step S205; NO), the collation unit 42 determines whether the collation target speech signal is based on the input feature amount. The similarity with each recognized word is calculated (step S206), and the recognized word with the highest similarity is specified (step S207). Then, the collation unit 42 determines whether or not the similarity of the identified recognized word exceeds the corresponding second threshold T2 (step S208).

  If it is determined that the similarity of the identified recognition word exceeds the corresponding second threshold T2 (step S208; YES), the collation unit 42 uses the identified recognition word as a recognition result via the output unit 30. Is output (step S209). Then, this process ends. On the other hand, when it is determined that the similarity of the identified recognized word is equal to or less than the corresponding second threshold value T2 (step S208; NO), the process proceeds to the process of step S009 described in the first embodiment, and the collating unit 42 further determines whether or not the similarity of the identified recognition word exceeds the first threshold value T1 (step S009).

  If it is determined that the similarity of the identified recognition word is equal to or less than the first threshold T1 (step S009; NO), the collation unit 42 rejects the speech signal to be collated (step S006), and this process ends. To do. On the other hand, when it is determined that the similarity of the identified recognition word exceeds the first threshold T1 (step S009; YES), the collation unit 42 further rejects the similarity identified by the similarity of the identified recognition word. It is determined whether or not the similarity is exceeded (step S010).

  When it is determined that the similarity of the identified recognition word is equal to or less than the similarity of the specified rejection word (step S010; NO), the collation unit 42 rejects the speech signal to be collated (step S006), The process ends. On the other hand, if it is determined that the similarity of the identified recognition word exceeds the similarity of the specified rejection word (step S010; YES), the collation unit 42 uses the identified recognition word as a recognition result as an output unit. 30 is output (step S011). Then, this process ends.

  Next, effects of the third embodiment will be described with reference to FIG. FIG. 15A is a diagram showing the recall rate and the matching rate for each second threshold value T2 in the prior art 2, and FIG. 15B is a diagram showing the recall rate and the matching rate for each third threshold value T3 in the third embodiment. It is. Note that each value of the recall rate and the matching rate is a value based on the evaluation result with respect to the speech data for 300 words and about one hour.

  When the second threshold value T2 is increased from the reference value (T2 = 80) in the prior art 2, as shown in FIG. 15A, when the second threshold value T2 = 91, the recall is 89% and the precision is When the second threshold value T2 = 93, the recall rate is 87% and the matching rate is 79%. When the second threshold value T2 = 95, the recall rate is 85% and the matching rate is 80%. It becomes. In other words, in the case of the prior art 2, when the second threshold value T2 is increased from the reference value (T2 = 80), the matching rate can be increased, but the recall rate is also deteriorated.

  In contrast, when the method of the first embodiment is applied to the conventional technique 2, when the third threshold value T3 = 95, the recall rate is 91% and the matching rate is 79%, and when the third threshold value T3 = 94, When the third threshold value T3 = 93, the recall rate is 89% and the match rate is 84%. That is, by adjusting the third threshold value T3, it is possible to improve the matching rate while maintaining the recall rate substantially. When evaluating the case where the method of the first embodiment is applied to the related art 2 (the third embodiment), the second threshold value T2 is fixed to the reference value (T2 = 80).

(Embodiment 4)
In the fourth embodiment, another example in which the method described in the first embodiment is applied to the related art 2 will be described. In the speech recognition processing according to the third embodiment, the determination processing as to whether or not the similarity of the reject word having the highest similarity exceeds the corresponding third threshold T3 is performed, and the similarity of the recognition word having the highest similarity is the second. This was performed before the process of determining whether or not the threshold value T2 was exceeded.

  In the fourth embodiment, it is determined whether or not the similarity of the reject word having the highest similarity exceeds the corresponding third threshold T3, and the similarity of the recognized word having the highest similarity is the second threshold T2. This is performed after the process of determining whether or not the threshold is exceeded. By doing so, it becomes possible to make a setting with more emphasis on the recall.

  The basic configuration of the speech recognition apparatus 1 in the fourth embodiment is the same as that in the third embodiment. However, the function performed by the matching unit 42 is slightly different from that in the third embodiment.

  The matching unit 42 according to the fourth embodiment compares the acoustic model stored in the acoustic model storage unit 23 with the input feature quantity, and extracts a phoneme string corresponding to the input speech. Then, the matching unit 42 calculates the similarity by comparing the extracted phoneme sequence and the phoneme sequence of each recognized word, and calculates the similarity by comparing the extracted phoneme sequence and the phoneme sequence of each rejected word. To do. Note that a conventionally used method can be used as the similarity calculation method. In the fourth embodiment, the similarity is normalized to a value between 0 and 100.

  And the collation part 42 specifies the recognition word with the highest similarity, and determines whether the similarity of the specified recognition word exceeds the corresponding 2nd threshold value T2. If it is determined that the similarity of the identified recognition word exceeds the corresponding second threshold T2, the collation unit 42 outputs the identified recognition word as a recognition result via the output unit 30. . On the other hand, if it is determined that the similarity of the identified recognition word is equal to or less than the corresponding second threshold value T2, the matching unit 42 identifies the rejection word having the highest similarity, and the similarity of the identified rejection word is It is determined whether or not the corresponding third threshold value T3 is exceeded.

  And when it determines with the similarity of the identified rejection word exceeding the corresponding 3rd threshold value T3, the collation part 42 rejects the audio | voice signal of collation object. On the other hand, if it is determined that the similarity of the identified rejection word is equal to or less than the corresponding third threshold T3, the collation unit 42 further determines whether the similarity of the identified recognition word exceeds the first threshold T1. Determine whether. And when it determines with the similarity of the identified recognition word being below 1st threshold value T1, the collation part 42 rejects the audio | voice signal of collation object. On the other hand, when it is determined that the similarity of the identified recognition word exceeds the first threshold T1, the collation unit 42 further exceeds the similarity of the rejection word that is identified by the similarity of the identified recognition word. It is determined whether or not.

  And when it determines with the similarity of the identified recognition word exceeding the similarity of the identified rejection word, the collation part 42 outputs via the output part 30 by making the identified recognition word into a recognition result. To do. On the other hand, when it determines with the similarity of the identified recognition word being below the similarity of the identified rejection word, the collation part 42 rejects the audio | voice signal of collation object.

  Next, with reference to FIG. 16, the flow of speech recognition processing in the fourth embodiment will be described. FIG. 16 is a part of an example of a flowchart for explaining the flow of speech recognition processing in the fourth embodiment. The voice recognition process is started with the input signal received by the input unit 10 being output to the control unit 40 as a trigger. Note that the processing after step S309 is the same as the processing described in the first embodiment.

  The analysis unit 41 analyzes the input signal and detects an audio signal (step S301). And the analysis part 41 calculates the feature-value of the detected audio | voice signal, and outputs the calculated feature-value to the collation part 42 (step S302).

  Then, the collation unit 42 calculates the similarity between the speech signal to be collated and each recognized word based on the input feature amount (step S303), and identifies the recognized word having the highest similarity (step S303). S304). Then, the collation unit 42 determines whether or not the similarity of the identified recognized word exceeds the corresponding second threshold T2 (step S305).

  If it is determined that the similarity of the identified recognition word exceeds the corresponding second threshold T2 (step S305; YES), the collation unit 42 uses the identified recognition word as a recognition result via the output unit 30. Is output (step S306). Then, this process ends. On the other hand, when it is determined that the similarity of the identified recognition word is equal to or less than the corresponding second threshold value T2 (step S305; NO), the collation unit 42 determines the speech to be collated based on the input feature amount. The similarity between the signal and each rejection word is calculated (step S307), and the rejection word with the highest similarity is specified (step S308).

  And the collation part 42 determines whether the similarity of the specified rejection word exceeds the corresponding 3rd threshold value T3 (step S309). When it is determined that the similarity of the specified rejection word exceeds the third threshold T3 (step S309; YES), the process proceeds to the process of step S006 described in the first embodiment. The voice signal to be verified is rejected (step S006), and this process ends.

  On the other hand, if it is determined that the similarity of the specified rejection word is equal to or less than the third threshold T3 (step S309; NO), the process proceeds to the process of step S009 described in the first embodiment, and the collation unit 42 Further, it is determined whether the similarity of the identified recognition word exceeds the first threshold T1 (step S009).

  If it is determined that the similarity of the identified recognition word is equal to or less than the first threshold T1 (step S009; NO), the collation unit 42 rejects the speech signal to be collated (step S006), and this process ends. To do. On the other hand, when it is determined that the similarity of the identified recognition word exceeds the first threshold T1 (step S009; YES), the collation unit 42 further rejects the similarity identified by the similarity of the identified recognition word. It is determined whether or not the similarity is exceeded (step S010).

  When it is determined that the similarity of the identified recognition word is equal to or less than the similarity of the specified rejection word (step S010; NO), the collation unit 42 rejects the speech signal to be collated (step S006), The process ends. On the other hand, if it is determined that the similarity of the identified recognition word exceeds the similarity of the specified rejection word (step S010; YES), the collation unit 42 uses the identified recognition word as a recognition result as an output unit. 30 is output (step S011). Then, this process ends.

  FIG. 17 is a diagram illustrating an example of a hardware configuration of the speech recognition apparatus 1 according to the first to fourth embodiments. The voice recognition device 1 shown in FIG. 1 and the like may be realized by various hardware shown in FIG. 17, for example. In the example of FIG. 17, the speech recognition apparatus 1 includes a CPU 201, a RAM 202, a ROM 203, an HDD 204, an input / output interface 205, and a communication module 206, and these hardware are connected via a bus 207.

  For example, the CPU 201 loads an operation program stored in the HDD 204 into the RAM 202 and executes various processes while using the RAM 202 as a working memory. The CPU 201 can implement each function unit of the control unit 40 illustrated in FIG. 1 and the like by executing the operation program.

  Note that an operation program for executing the above operation is a computer-readable recording medium such as a flexible disk, Compact Disk-Read Only Memory (CD-ROM), Digital Versatile Disk (DVD), or Magneto Optical disk (MO). It may be configured to execute the above-described processing by storing and distributing it (not shown), reading it with a reading device (not shown) of the speech recognition apparatus 1 and installing it in a computer. Furthermore, an operation program may be stored in a disk device or the like included in a server device on the Internet, and the operation program may be downloaded to the computer of the speech recognition apparatus 1 via the communication module 206.

  Depending on the embodiment, other types of storage devices other than the RAM 202, the ROM 203, and the HDD 204 may be used. For example, the speech recognition apparatus 1 may include a storage device such as a Content Addressable Memory (CAM), a Static Random Access Memory (SRAM), or a Synchronous Dynamic Random Access Memory (SDRAM).

  Depending on the embodiment, the hardware configuration of the speech recognition apparatus 1 may be different from that in FIG. 17, and other hardware other than the standards and types illustrated in FIG. 17 may be applied to the speech recognition apparatus 1. You can also.

  For example, each functional unit of the control unit 40 of the voice recognition device 1 illustrated in FIG. 1 and the like may be realized by a hardware circuit. Specifically, each functional unit of the control unit 40 shown in FIG. 1 is realized by a reconfigurable circuit such as Field Programmable Gate Array (FPGA) or Application Specific Integrated Circuit (ASIC) instead of the CPU 201. May be. Of course, these functional units may be realized by both the CPU 201 and the hardware circuit.

  In the above, several embodiments have been described. However, the embodiments are not limited to the above-described embodiments, and should be understood as including various modifications and alternatives of the above-described embodiments. For example, it will be understood that various embodiments can be embodied by modifying the components without departing from the spirit and scope thereof. It will be understood that various embodiments can be made by appropriately combining a plurality of components disclosed in the above-described embodiments. Further, various embodiments may be implemented by deleting or replacing some components from all the components shown in the embodiments, or adding some components to the components shown in the embodiments. Those skilled in the art will appreciate that this can be done.

The following additional notes are further disclosed with respect to the embodiments including the first to fourth embodiments.
(Appendix 1)
A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition device that rejects the input speech when it is equal to or lower than a first threshold value set in advance,
Calculating a similarity between the input voice and the reject word, and specifying a reject word having the highest similarity;
Calculating a similarity between the input speech and the recognized word, and specifying a recognized word having the highest similarity;
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first Collating means for rejecting the input speech even if the threshold is exceeded;
Comprising
A speech recognition apparatus characterized by that.
(Appendix 2)
When the similarity of the identified rejection word is equal to or less than the third threshold and the similarity of the identified recognition word satisfies a predetermined condition, the collating unit identifies the identified recognition If the similarity of the identified rejection word is less than or equal to the third threshold and the identified similarity of the recognized word does not satisfy the predetermined condition, the input speech is output To reject,
The speech recognition apparatus according to supplementary note 1, wherein:
(Appendix 3)
The predetermined condition is that the similarity of the identified recognition word exceeds the similarity of the identified rejection word and exceeds the first threshold value.
The speech recognition apparatus according to Supplementary Note 2, wherein
(Appendix 4)
When the similarity of the identified recognition word exceeds a second threshold value that is greater than the first threshold value, the matching unit determines that the similarity of the identified rejection word is the third threshold value. Output the identified recognition word regardless of whether or not
The speech recognition apparatus according to any one of appendices 1 to 3, wherein
(Appendix 5)
The predetermined condition is that the degree of similarity of the identified recognized word exceeds the degree of similarity of the identified rejection word and exceeds the first threshold value, or the identified recognition The similarity of the words exceeds a second threshold greater than the first threshold;
The speech recognition apparatus according to Supplementary Note 2, wherein
(Appendix 6)
A plurality of third thresholds having different values are preset, and any one of the plurality of third thresholds is associated with each rejection word,
The collation means rejects the input speech when the similarity of the identified rejection word exceeds the corresponding third threshold;
The speech recognition device according to any one of appendices 1 to 5, characterized in that:
(Appendix 7)
The third threshold is determined so that the rejection word specified when the third threshold is exceeded is likely to be spoken.
The speech recognition device according to any one of appendices 1 to 6, characterized in that:
(Appendix 8)
A setting means for setting the third threshold for the identified rejection word higher as the correlation between the identified rejection word and the identified recognized word is higher;
The collation means rejects the input speech when the similarity of the identified rejection word exceeds the corresponding third threshold;
The speech recognition device according to any one of appendices 1 to 5, characterized in that:
(Appendix 9)
The setting means calculates a similarity between words of the identified rejection word and the identified recognition word, and the higher the similarity between the calculated words is, the higher the similarity between the identified rejection words, the third Set a high threshold for
The speech recognition apparatus according to appendix 8, wherein
(Appendix 10)
The setting means sets the third threshold value to be equal to or greater than a predetermined value;
The speech recognition apparatus according to appendix 8 or 9, characterized in that.
(Appendix 11)
A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition method for a speech recognition apparatus that rejects the input speech when the input speech is equal to or lower than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is less than the first threshold. Even if it exceeds, reject the input speech,
A speech recognition method characterized by the above.
(Appendix 12)
A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is In the computer of the speech recognition device that rejects the input speech when it is equal to or less than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is less than the first threshold. Even if it exceeds, reject the input speech,
To execute the process,
A program characterized by that.
(Appendix 13)
A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is In the computer of the speech recognition device that rejects the input speech when it is equal to or less than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is less than the first threshold. Even if it exceeds, reject the input speech,
A recording medium storing a program for executing processing.

DESCRIPTION OF SYMBOLS 1 Speech recognition apparatus 10 Input part 20 Storage part 21 Recognition word dictionary 22 Reject word dictionary 23 Acoustic model storage part 30 Output part 40 Control part 41 Analysis part 42 Collation part 43 Threshold calculation part T1 1st threshold value T2 2nd threshold value T3 3rd Threshold T3-W1 Third threshold of reject word “Wakayama” T3-W2 Third threshold of reject word “Toyama” BD1 Ideal range BD2 Detection range SC1 Similarity of recognition word “Okayama” SC2 Similarity of reject word “Wakayama” SC3 Similarity of reject word “Toyama” L1-L5 dotted line F function 201 CPU
202 RAM
203 ROM
204 HDD
205 I / O interface 206 Communication module 207 Bus

Claims (15)

A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition device that rejects the input speech when it is equal to or lower than a first threshold value set in advance,
Calculating a similarity between the input voice and the reject word, and specifying a reject word having the highest similarity;
Calculating a similarity between the input speech and the recognized word, and specifying a recognized word having the highest similarity;
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first Collating means for rejecting the input speech even if the threshold is exceeded;
Equipped with a,
When the similarity of the identified recognition word exceeds a second threshold value that is greater than the first threshold value, the matching unit determines that the similarity of the identified rejection word is the third threshold value. Output the identified recognition word regardless of whether or not
A speech recognition apparatus characterized by that. When the similarity of the identified rejection word is equal to or less than the third threshold and the similarity of the identified recognition word satisfies a predetermined condition, the collating unit identifies the identified recognition If the similarity of the identified rejection word is less than or equal to the third threshold and the identified similarity of the recognized word does not satisfy the predetermined condition, the input speech is output To reject,
The speech recognition apparatus according to claim 1. The predetermined condition is that the similarity of the identified recognition word exceeds the similarity of the identified rejection word and exceeds the first threshold value.
The speech recognition apparatus according to claim 2. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition device that rejects the input speech when it is equal to or lower than a first threshold value set in advance,
Calculating a similarity between the input voice and the reject word, and specifying a reject word having the highest similarity;
Calculating a similarity between the input speech and the recognized word, and specifying a recognized word having the highest similarity;
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first Collating means for rejecting the input speech even if the threshold is exceeded;
With
The verification means includes
If the similarity of the identified rejection word is less than or equal to the third threshold and the similarity of the identified recognition word satisfies a predetermined condition, the identified recognition word is output,
When the similarity of the identified rejection word is equal to or less than the third threshold and the similarity of the identified recognition word does not satisfy the predetermined condition, the input speech is rejected,
The predetermined condition is:
The similarity of the identified recognition word is less than or equal to a second threshold greater than the first threshold, and the similarity of the identified recognition word exceeds the similarity of the identified rejection word And a first condition that the first threshold is exceeded ;
The similarity of the recognition word identified, a second condition that exceeds the second threshold value,
Which is either
A speech recognition apparatus characterized by that. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition device that rejects the input speech when it is equal to or lower than a first threshold value set in advance,
Calculating a similarity between the input voice and the reject word, and specifying a reject word having the highest similarity;
Calculating a similarity between the input speech and the recognized word, and specifying a recognized word having the highest similarity;
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first Collating means for rejecting the input speech even if the threshold is exceeded;
With
A plurality of third thresholds having different values are preset, and any one of the plurality of third thresholds is associated with each rejection word,
The collation means rejects the input speech when the similarity of the identified rejection word exceeds the corresponding third threshold;
A speech recognition apparatus characterized by that. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition device that rejects the input speech when it is equal to or lower than a first threshold value set in advance,
Calculating a similarity between the input voice and the reject word, and specifying a reject word having the highest similarity;
Calculating a similarity between the input speech and the recognized word, and specifying a recognized word having the highest similarity;
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first Collating means for rejecting the input speech even if the threshold is exceeded;
The similarity between the identified rejection word and the identified recognition word is calculated, and the higher the similarity between the calculated words , the higher the third threshold for the identified rejection word is set. Setting means to
With
The collation means rejects the input speech when the similarity of the identified rejection word exceeds the corresponding third threshold;
A speech recognition apparatus characterized by that. The setting means sets the third threshold value to be equal to or greater than a predetermined value;
The speech recognition apparatus according to claim 6 . A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition method for a speech recognition apparatus that rejects the input speech when the input speech is equal to or lower than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
The similarity of the reject word specified, are previously set, when said first exceeds the threshold value larger than the third threshold value, the similarity of the recognition word identified the first threshold value even if not exceed, dismissed the input voice,
If the degree of similarity of the identified recognition word exceeds a second threshold value that is greater than the first threshold value, whether the degree of similarity of the identified rejection word exceeds the third threshold value Regardless of whether the identified recognition word is output,
A speech recognition method characterized by the above. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition method for a speech recognition apparatus that rejects the input speech when the input speech is equal to or lower than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first threshold. Even if it exceeds, the input speech is rejected,
If the similarity of the identified rejection word is less than or equal to the third threshold and the similarity of the identified recognition word satisfies a predetermined condition, the identified recognition word is output,
When the similarity of the identified rejection word is equal to or less than the third threshold and the similarity of the identified recognition word does not satisfy the predetermined condition, the input speech is rejected,
The predetermined condition is:
The similarity of the identified recognition word is less than or equal to a second threshold greater than the first threshold, and the similarity of the identified recognition word exceeds the similarity of the identified rejection word And a first condition that the first threshold is exceeded;
A second condition that the similarity of the identified recognized word exceeds the second threshold;
Which is either
A speech recognition method characterized by the above. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition method for a speech recognition apparatus that rejects the input speech when the input speech is equal to or lower than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first threshold. Even if it exceeds, the input speech is rejected,
A plurality of third thresholds having different values are preset, and any one of the plurality of third thresholds is associated with each rejection word,
In the rejection of the input speech, if the similarity of the identified rejection word exceeds the corresponding third threshold, the input speech is rejected.
A speech recognition method characterized by the above. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is A speech recognition method for a speech recognition apparatus that rejects the input speech when the input speech is equal to or lower than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first threshold. Even if it exceeds, the input speech is rejected,
The similarity between the identified rejection word and the identified recognition word is calculated, and the higher the similarity between the calculated words, the higher the third threshold for the identified rejection word is set. And
In the rejection of the input speech, if the similarity of the identified rejection word exceeds the corresponding third threshold, the input speech is rejected.
A speech recognition method characterized by the above. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is In the computer of the speech recognition device that rejects the input speech when it is equal to or less than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
The similarity of the reject word specified, are previously set, when said first exceeds the threshold value larger than the third threshold value, the similarity of the recognition word identified the first threshold value even if not exceed, dismissed the input voice,
If the degree of similarity of the identified recognition word exceeds a second threshold value that is greater than the first threshold value, whether the degree of similarity of the identified rejection word exceeds the third threshold value Regardless of whether the identified recognition word is output,
To execute the process,
A program characterized by that. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is In the computer of the speech recognition device that rejects the input speech when it is equal to or less than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first threshold. Even if it exceeds, the input speech is rejected,
If the similarity of the identified rejection word is less than or equal to the third threshold and the similarity of the identified recognition word satisfies a predetermined condition, the identified recognition word is output,
If the similarity of the identified rejection word is less than or equal to the third threshold and the similarity of the identified recognition word does not satisfy the predetermined condition, reject the input speech;
Let the process run,
The predetermined condition is:
The similarity of the identified recognition word is less than or equal to a second threshold greater than the first threshold, and the similarity of the identified recognition word exceeds the similarity of the identified rejection word And a first condition that the first threshold is exceeded;
A second condition that the similarity of the identified recognized word exceeds the second threshold;
Which is either
A program characterized by that. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is In the computer of the speech recognition device that rejects the input speech when it is equal to or less than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first threshold. The input speech is rejected even if
Let the process run,
A plurality of third thresholds having different values are preset, and any one of the plurality of third thresholds is associated with each rejection word,
In the rejection of the input speech, if the similarity of the identified rejection word exceeds the corresponding third threshold, the input speech is rejected.
A program characterized by that. A recognition word dictionary in which a recognition word is registered and a rejection word dictionary in which a rejection word is registered; the similarity between the input speech and the recognition word is calculated, and the similarity of the recognition word having the highest similarity is In the computer of the speech recognition device that rejects the input speech when it is equal to or less than a first threshold value set in advance,
Calculate the similarity between the input speech and the reject word, identify the reject word with the highest similarity,
Calculating the similarity between the input speech and the recognized word, identifying the recognized word having the highest similarity,
When the similarity of the identified rejection word exceeds a third threshold that is set in advance and is greater than the first threshold, the similarity of the identified recognition word is the first threshold. Even if it exceeds, the input speech is rejected,
The similarity between the identified rejection word and the identified recognition word is calculated, and the higher the similarity between the calculated words, the higher the third threshold for the identified rejection word is set. To
Let the process run,
In the rejection of the input speech, if the similarity of the identified rejection word exceeds the corresponding third threshold, the input speech is rejected.
A program characterized by that.