JP6437092B2 - Speech recognition apparatus, speech recognition method, and speech recognition program - Google Patents

Description

  The present invention relates to a keyword extraction type speech recognition technique using a statistical language model.

In the keyword extraction type speech recognition technology, it is not necessary to correctly recognize all of the user's utterance contents one by one, and it is required to correctly recognize only important words called keywords. For example, if the recognition target is a dish name and its accompanying expression, “sushi” is the keyword, and the recognition result for the utterance “I want to eat sushi” is “I ate sushi”, the recognition result is incorrect. Although it is included, the keyword “sushi” is correctly extracted and can be said to be correct. On the other hand, if the recognition result for the utterance “I want to play tennis” is “Sushi in hand”, the recognition result is an error because “sushi” that is not uttered by the user is extracted. That is, it is required to correctly extract the keyword for utterances including a certain keyword and not to extract the keyword erroneously for utterances not including the keyword.
In the following, utterances that are not subject to recognition, such as “I want to play tennis” described above, are referred to as utterances outside a task.

In speech recognition technology using a statistical language model that expresses the ease of connection between words as a language likelihood, it appears in the learning corpus at a high frequency by learning the language likelihood from the learning corpus. High language likelihood is given to a combination of words. Further, by using a class language model in which a plurality of vocabularies are expressed as one group called a class, it is possible to efficiently learn a language likelihood from a sparse (rough information) learning corpus.
For example, if there is a corpus “I want to eat sushi” in the learning corpus, “Sushi, Ramen, Curry and Rice” is the one class class <A>, and the expression “I want to eat <A>” By doing so, it is possible to learn the language likelihood for the word sequences “I want to eat sushi”, “I want to eat ramen” and “I want to eat curry and rice”.

  In speech recognition technology using a statistical language model, a method called back-off that complements high-order N-gram probabilities from low-order N-gram probabilities for unknown words based on N-gram establishment that does not exist in the learning corpus. Gives the language likelihood. By this back-off, an expression including a chain of words that is not included in the learning corpus can be accepted. However, in the task of extracting keywords from the recognition result, there is a problem that as a result of backoff, keyword misdetection increases for utterances not included in the learning corpus.

  For example, consider a case where a language model is learned as a keyword class for extracting “sushi, ramen, curry and rice”. In this case, if “tennis” is an unknown word that is not included in the learning corpus, it is acoustically matched to the out-task utterance “I want to play tennis” as “sushi in hand”, but N-gram The score of the recognition result composed of unknown words based on the establishment becomes high, and the keyword “sushi” is erroneously extracted. In this way, when a keyword is erroneously detected, a keyword that is not intended by the user or a function associated with the keyword is presented, which causes discomfort to the user. In particular, when the keyword length is short, the frequency with which the acoustic similarity with the keyword increases for a part of the utterance increases, and the erroneous detection of the keyword occurs frequently as in the example described above.

  By reducing the 1-gram probability of the keyword with respect to the above-described problem, erroneous keyword detection is suppressed.

  In addition, for the problem described above, Patent Document 1 gives a high word insertion penalty for short words that are likely to be erroneously inserted by local matching, and short words are continuously inserted. In order to suppress this, a speech recognition device that detects the word length of an inserted word and determines a word insertion penalty so that a larger penalty is imposed as the detected word length becomes shorter is disclosed.

JP 2011-164192 A

However, the above-described technique for reducing the 1-gram probability of a keyword has a problem that the recognition performance for an utterance of only a short keyword is lowered. Further, if the back-off method is not applied or the N-gram probability value for an unknown word calculated by the back-off is lowered, it is difficult to recognize expressions not included in the learning corpus, and the keyword extraction performance decreases. There was a problem.
Further, the technique disclosed in Patent Document 1 described above has a problem that the recognition performance of an utterance including a short keyword is deteriorated.

  The present invention has been made to solve the above-described problems, and an object thereof is to suppress a reduction in recognition performance for an utterance including a keyword having a short word length without reducing the keyword extraction performance. .

The speech recognition apparatus according to the present invention performs speech recognition of input speech using a language model that has been learned for a set recognition target and an acoustic model that is a model of speech characteristics, and the obtained recognition result is obtained. A speech recognition unit that calculates a recognition score of the recognition result from the language likelihood and acoustic likelihood calculated based on the unknown word included in the recognition result acquired by the speech recognition unit based on the N-gram probability of the language model When the unknown word number calculation unit for calculating the number of words, the keyword storage unit for storing keywords related to the set recognition target, and the keywords stored in the keyword storage unit are included in the recognition result acquired by the speech recognition unit, and keyword length calculation unit for calculating a keyword length indicating the length of the keyword, of reducing the recognition score in accordance with the increase in the number of unknown words unknown word count calculation unit has calculated , So as to reduce the recognition score in accordance with a decrease of the keyword length keyword length calculation section is calculated, recalculate the recognition score speech recognition unit is calculated, the voice recognition unit on the basis of the recalculated recognition score is acquired And a score recalculation unit that outputs the recognized result.

  According to the present invention, it is possible to suppress a decrease in recognition performance of an utterance including a keyword having a short word length without reducing the keyword extraction performance. In addition, it is possible to suppress erroneous recognition of a keyword that is particularly short for an out-of-task utterance.

1 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 1. FIG. 2 is a diagram illustrating a hardware configuration of the speech recognition apparatus according to Embodiment 1. FIG. 3 is a flowchart showing an operation of the speech recognition apparatus according to the first embodiment. 6 is an example showing a recognition result of a voice recognition unit of the voice recognition device according to the first embodiment. 4 is a flowchart showing an operation of an unknown word number calculation unit of the speech recognition apparatus according to the first embodiment. 4 is a flowchart illustrating an operation of a keyword length calculation unit of the speech recognition apparatus according to the first embodiment. 4 is a flowchart showing an operation of a score recalculation unit of the speech recognition apparatus according to Embodiment 1. It is an example which shows the recognition result after the recognition score update of the speech recognition apparatus which concerns on Embodiment 1. FIG. 4 is a block diagram illustrating a configuration of a speech recognition apparatus according to Embodiment 2. FIG. 6 is a flowchart showing the operation of the speech recognition apparatus according to the second embodiment. 10 is a flowchart illustrating an operation of a keyword selection frequency calculation unit of the speech recognition apparatus according to the second embodiment. 10 is a flowchart illustrating an operation of a score recalculation unit of the speech recognition apparatus according to the second embodiment. It is an example which shows the recognition result after the recognition score update of the speech recognition apparatus which concerns on Embodiment 2. FIG.

Hereinafter, in order to explain the present invention in more detail, modes for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1 FIG.
FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus 10 according to the first embodiment.
The speech recognition device 10 is a device that performs speech recognition processing on an input signal and obtains a recognition result, and includes a speech recognition unit 1, an acoustic model storage unit 2, a language model storage unit 3, an unknown word number calculation unit 4, The keyword length calculation unit 5, the keyword storage unit 6, and the score recalculation unit 7 are configured.

  In the first embodiment, the speech recognition unit 1, the acoustic model storage unit 2, the language model storage unit 3, the unknown word number calculation unit 4, the keyword length calculation unit 5, the keyword storage unit 6, and the score recalculation unit 7 are dedicated to each. It is assumed that the circuit is configured by Note that the information processing method performed by the circuit is also a feature of the present invention. It can also be realized by a combination of a computer program and a control circuit constituted by other constituent circuits such as a general-purpose CPU.

FIG. 2 is a diagram illustrating a hardware configuration of the speech recognition apparatus 10 according to the first embodiment.
The speech recognition unit 1, the unknown word number calculation unit 4, the keyword length calculation unit 5, and the score recalculation unit 7 of the speech recognition apparatus 10 are realized by the processor 20 executing a program stored in the memory 30. The acoustic model storage unit 2, the language model storage unit 3, and the keyword storage unit 6 constitute a memory 30. Further, a plurality of processors 20 and a plurality of memories 30 may be configured to cooperate to execute the above-described functions.

Next, each configuration of the speech recognition apparatus 10 will be described. In the following, a case where the recognition target of the voice recognition device 10 is a dish name and an accompanying expression of the dish name will be described as an example.
The speech recognition unit 1 performs speech recognition on the input speech using the acoustic model stored in the acoustic model storage unit 2 and the language model stored in the language model storage unit 3, and acquires a recognition result. . Further, the obtained recognition results are ranked in descending order of recognition score, the notation of the top N recognition results is extracted, and the recognition score is calculated. Here, the recognition score is the sum of acoustic likelihood and language likelihood.

  The acoustic model accumulating unit 2 accumulates an acoustic model obtained by modeling a voice feature. The acoustic model is, for example, an HMM (Hidden Markov Model). The language model accumulating unit 3 accumulates a statistical language model (N-gram language model) obtained by learning a dish name and an accompanying expression of the dish name as a learning corpus. The cooking names to be accumulated are learned using a language model in which a plurality of vocabularies are expressed in one group called a class (hereinafter referred to as a keyword class). The incidental expression of the accumulated dish name is, for example, an expression “I want to eat <A>” or “I want to eat delicious <A>”.

In the expression “I want to eat <A>” or “I want to eat delicious <A>”, the keyword class is described as <A> on the learning corpus, and the keyword A included in the keyword class <A> is “Sushi , Ramen, curry and rice ". By accumulating the expressions, it is not necessary to develop additional expressions such as “I want to eat” and “I want to eat delicious” for all the keywords A, and the language model can be efficiently learned.
In addition, although the expressions “<A> wants to eat” and “delicious <A> wants to eat” are shown as supplementary expressions of the dish name, learning that covers expressions that the user is supposed to utter in addition to the expressions Use a corpus.

  The unknown word number calculation unit 4 refers to the language model stored in the language model storage unit 3 and calculates the number of unknown words based on the N-gram probability from the words included in the notation of the N recognition results extracted by the speech recognition unit 1. calculate. The keyword length calculation unit 5 refers to the keyword descriptions and keyword readings stored in the keyword storage unit 6 and calculates the keyword lengths of the keywords included in the N recognition result descriptions output by the speech recognition unit 1. . The keyword accumulating unit 6 accumulates keyword notations and keyword readings according to the recognition result format output by the speech recognition unit 1. The keywords to be stored by the keyword storage unit 6 are keywords corresponding to the recognition targets of the speech recognition device 10. For example, when the recognition target is a dish name and an accompanying expression of the dish name, “sushi, ramen” , Curry and rice "accumulation and notation.

  The score recalculation unit 7 uses the recognition score of the recognition result extracted by the speech recognition unit 1, the number of unknown words calculated by the unknown word number calculation unit 4, and the keyword length of the keyword calculated by the keyword length calculation unit 5. Recalculate the recognition score for. The score recalculation unit 7 outputs a recognition result having the largest recognition score among the recalculated recognition scores.

Next, the operation of the speech recognition apparatus 10 will be described with reference to a flowchart and a specific example. FIG. 3 is a flowchart showing the operation of the speech recognition apparatus 10 according to the first embodiment. FIG. 4 is an example showing a recognition result of the speech recognition unit 1 of the speech recognition apparatus 10 according to the first embodiment.
When speech is input to the speech recognition device 10 (step ST1), the speech recognition unit 1 refers to the acoustic model stored in the acoustic model storage unit 2 and the language model stored in the language model storage unit 3, Voice recognition of the voice input at step ST1 is performed to obtain a recognition result (step ST2). Furthermore, the speech recognition unit 1 ranks the recognition results acquired in step ST2 in descending order of recognition scores, and extracts the top N recognition result descriptions and recognition scores (step ST3).

When an utterance outside task "I want to play tennis" is input as a voice in step ST1,
The voice recognition unit 1 performs the processing of step ST2 and step ST3, and extracts the recognition result shown in FIG. The notation of the recognition result of the recognition rank No. 1 is “Sushi in hand” and the recognition score is “110”. The notation of the recognition result of the second recognition rank is “Keep it in hand” and the recognition score is “105”. In the recognition result of FIG. 4, the notation “sushi to sushi” in which the kana of the reading matches “I want to play tennis” has the largest recognition score, and is extracted as the recognition result of the first recognition rank.

The unknown word number calculation unit 4 refers to the language model stored in the language model storage unit 3, and indicates a value C _n indicating the number of unknown words based on the N-gram probability included in the N recognition results n extracted in step ST3. (hereinafter, referred to as the unknown word number _{C n)} is calculated (step ST4). The unknown word count C _n calculated in step ST4 is output to the score recalculation unit 7. The detailed processing operation of step ST4 will be described later.

The keyword length calculation unit 5 refers to the keyword notations and keyword readings stored in the keyword storage unit 6 and refers to the values L _n indicating the lengths of the keywords included in the N recognition results n extracted in step ST3. (hereinafter, referred to as keyword length _{L n)} is calculated (step ST5). Keyword length L _n calculated in step ST5 are output to the score recalculation unit 7. The detailed processing operation of step ST5 will be described later.

Score recalculation unit 7 uses the keyword length L _n calculated in unknown word number C _n and step ST5 calculated in step ST4, updates the recognition score of the extracted N number of recognition results in step ST3 (Step ST6). The score recalculation unit 7 ranks the recognition results in descending order of the recognition score updated in step ST6, outputs the recognition result having the largest recognition score (step ST7), and ends the process. Detailed processing operations in steps ST6 and ST7 will be described later.

Next, the processing operation of the unknown word number calculation unit 4 shown in step ST4 will be described in more detail.
FIG. 5 is a flowchart showing the operation of the unknown word number calculation unit 4 of the speech recognition apparatus 10 according to the first embodiment. Hereinafter, N recognition results are referred to as recognition results n (n = 1, 2, 3,..., N). Further, a word included in the recognition result n is described as a word m (m = 1, 2, 3,..., M _n ). Note that n is an index of recognition results, and m is a word index. Furthermore, the example which counts the number of unknown words based on 3-gram probability as an unknown word number below is shown.

The unknown word number calculation unit 4 initializes the index n of the recognition result extracted in step ST3 to “1” (step ST11), initializes the index m of the word included in the recognition result n to “1”, and recognizes the recognition result. The number of unknown words C _n included in n is initialized to “0” (step ST12). Next, the unknown word number calculation unit 4 refers to the language model stored in the language model storage unit 3, and the 3-gram probability P (w _{n, m} | w _{n, m−2} w _{n, m−1} ) performing the set determined whether the threshold value _{P th} or less (step ST13). Here, w _{n, m} of the 3-gram probability P means the m-th word of the n-th recognition result. Further, by using the threshold value _Pth , when the 3-gram probability P is equal to or less than the threshold value _Pth , the word can be regarded as an unknown word.

When the 3-gram probability P is equal to or less than the threshold value P _th (step ST13; YES), the unknown word number calculation unit 4 adds 1 to the unknown word number C _n (step ST14), and the index m of the word included in the recognition result n. 1 is added to (step ST15). On the other hand, 3-gram probability if P is greater than the threshold _{P th} (step ST13; NO), the process proceeds to step ST15 described above. After step ST15, the unknown word number calculation unit 4 determines whether or not the index m of the word included in the recognition result n is equal to or less than the number _{Mn of} words included in the recognition result n (step ST16).

When the index m of the word included in the recognition result n is less than or equal to the number _{Mn of} words included in the recognition result n (step ST16; YES), the process returns to step ST13. On the other hand, when the word index m included in the recognition result n is larger than the number _{Mn of} words included in the recognition result n (step ST16; NO), the unknown word number calculation unit 4 adds 1 to the index n of the recognition result. (Step ST17). Furthermore, the unknown word number calculation unit 4 determines whether or not the index n of the recognition result is equal to or less than the number N of recognition results (step ST18). If the index n of the recognition result is equal to or less than the number N of recognition results (step ST18; YES), the process returns to step ST12. On the other hand, when the index n of the recognition result is larger than the number N of the recognition results (step ST18; NO), the unknown word number calculation unit 4 outputs the unknown word number C _n obtained by the above-described processing to the score recalculation unit 7. (Step ST19), the process proceeds to Step ST5 of the flowchart of FIG.

  In the above description, an example of counting the number of unknown words based on the 3-gram probability is shown, but the number of unknown words based on the 2-gram probability may be counted, or the number of unknown words based on the 3-gram probability and 2 -You may comprise so that the total value of both of the number of unknown words based on -gram probability may be counted.

Next, the processing operation of the keyword length calculation unit 5 shown in step ST5 will be described in more detail.
FIG. 6 is a flowchart showing the operation of the keyword length calculation unit 5 of the speech recognition apparatus 10 according to the first embodiment.
The keyword length calculation unit 5 initializes the index n of the recognition result extracted in step ST3 to “1” (step ST21), initializes the index m of the word included in the recognition result n to “1”, and recognizes the recognition result. A value L _n (hereinafter referred to as keyword length L _n ) indicating the length of the keyword included in _n is initialized to “0” (step ST22). Next, the keyword length calculation unit 5 refers to the keyword notation and keyword reading stored in the keyword storage unit 6, _{and the mth} word W _{n, m} of the _nth recognition result is stored in the keyword storage unit 6. It is determined whether or not the keyword is stored (step S23).

When W _{n, m} is a keyword (step ST23; YES), the keyword length calculation unit 5 updates the keyword length L _n of the keyword based on the following equations (1) and (2) (step) ST24).
L ' _n = UpdateLength (L _n , length (w _{n, m} )) (1)
L _n = L ' _n (2)
In Expression (1) and Expression (2), L ′ _n is a keyword length after update, length (w) is a function that returns the length of keyword w, and UpdateLength (L _n , A) is an update of keyword length L _n . The function to perform. In the first embodiment, length (w) is described as a function for calculating the mora length of the keyword w, and UpdateLength (L _n , A) is described as a function for calculating the minimum value of the keyword lengths L _n and A.

The keyword length calculation unit 5 adds 1 to the word index m included in the recognition result n (step ST25). On the other hand, if the word W _{n, m} is not a keyword (step ST23; NO), the process proceeds to step ST25 described above. Following step ST25, the keyword length calculation unit 5 determines whether or not the index m of the word included in the recognition result n is equal to or less than the number _{Mn of} words included in the recognition result n (step ST26). When the index m of the word included in the recognition result n is _{equal to} or less than the number _{Mn of} words included in the recognition result n (step ST26; YES), the process returns to step ST23.

On the other hand, when the index m of the word included in the recognition result n is larger than the number M _{n of} words included in the recognition result n (step ST26; NO), the keyword length calculation unit 5 adds 1 to the index n of the recognition result. (Step ST27). Further, the keyword length calculation unit 5 determines whether or not the recognition result index n is equal to or less than the number N of recognition results (step ST28). If the index n of the recognition result is less than or equal to the number N of recognition results (step ST28; YES), the process returns to step ST22. On the other hand, greater than the number N of index n of the recognition result is a recognition result (step ST28; NO), the keyword length calculating unit 5, and outputs a keyword length L _n obtained by the process described above the score recalculation section 7 (Step ST29), the process proceeds to Step ST6 of the flowchart of FIG.

Next, the processing operation of the score recalculation unit 7 shown in step ST6 and step ST7 will be described in more detail.
FIG. 7 is a flowchart showing the operation of the score recalculation unit 7 of the speech recognition apparatus 10 according to the first embodiment.
The score recalculation unit 7 initializes the index n of the recognition result extracted in step ST3 to “1” (step ST31). The score recalculation unit 7 uses the unknown word count C _n calculated in step ST4 and the keyword length L _n calculated in step ST5 to calculate the recognition score Score _n of the recognition result n extracted in step ST3 as follows: Update based on (3) and equation (4) (step ST32)
Score ' _n = Rescore1 (Score _n , C _n , L _n ) (3)
Score _n = Score ' _n (4)
In Equations (3) and (4), Score _n represents the updated recognition score, and Rescore 1 (Score _n , C _n , L _n ) is a function for updating the score.

上述した式（５）の第２項および第３項は、従来技術である認識スコアに対するペナルティに相当する。α、β、ＴＨ_Ｌは予め実験により決定されたパラメータである。この実施の形態１では、例えばα＝１、β＝１０、ＴＨ_Ｌ＝０とする。In the first embodiment, the function shown in the following formula (5) is used as a function for updating the score.

The second term and the third term of the above-described formula (5) correspond to a penalty for the recognition score, which is a conventional technique. α, β, and TH _L are parameters determined in advance by experiments. In the first embodiment, for example, α = 1, β = 10, and TH _L = 0.

As the function for updating the score, functions other than Expression (5) can be applied as long as it is a function of Score _n , C _n , and L _n . For example, you may comprise as the following formula | equation (5a).

The score recalculation unit 7 adds 1 to the index n of the recognition result (step ST33), and determines whether or not the index n of the recognition result is equal to or less than the number N of recognition results (step ST34). If the index n of the recognition result is less than or equal to the number N of recognition results (step ST34; YES), the process returns to step ST32. On the other hand, if the index n of the recognition result is larger than the number N of recognition results (step ST34; NO), the score recalculation unit 7 refers to the recognition score Score _n obtained by the above-described processing, and the recognition score Score _n The recognition results n are rearranged in descending order (step ST35), the recognition result n having the largest recognition score is output (step ST36), and the process ends.

FIG. 8 is an example illustrating a recognition result after the recognition score is updated in the speech recognition apparatus 10 according to the first embodiment. FIG. 8 shows the recognition rank, the recognition result, the recognition score before update, the number of unknown words C _n , the keyword length L _n, and the recognition score after update.
The number of unknown words C _n of the recognition results “handed sushi” and “sushi sushi” is both “2”. On the other hand, the keyword length L _n included in the recognition result “hand-made rice cake” is “0”, and the keyword length L _n contained in “hand-made sushi rice” is “2”. Substituting these results into the above formula (5), the recognition score after updating the recognition result “sushi in hand” decreases to “110− (1 × 2) − (10/2) = 103”. . On the other hand, the updated recognition score of the recognition result “Hold in hand” is “105” because L _n = 0. As a result, the recognition order based on the updated recognition score is reversed, and the first place becomes the recognition result “hands on rice” and the second place the recognition result “hands on sushi rice”.

  As described above, according to the first embodiment, the speech recognition of the input speech is performed, the speech recognition unit 1 that extracts the recognition result having the highest recognition score, and the recognition result of the recognition result by referring to the accumulated language model. An unknown word number calculating unit 4 that calculates the number of unknown words based on the notation, a keyword length calculating unit 5 that calculates the keyword length of the keyword included in the recognition result based on the notation of the recognition result with reference to the accumulated keyword, Since the score recalculation unit 7 that updates the recognition score of the recognition result extracted by the speech recognition unit 1 using the calculated number of unknown words and the keyword length is provided, the number of unknown words is large and the keyword length is short. The recognition score of the recognition result in which the keyword exists can be reduced. Thereby, it can suppress that many keywords with a short keyword length appear with respect to the utterance outside a task.

Embodiment 2. FIG.
In the second embodiment, the number of times the recognition result presented to the user by the voice recognition device 10 is selected or the number of times the function corresponding to the recognition result output by the voice recognition device 10 is selected is fed back to the out-task utterance. On the other hand, the structure which suppresses appearance of the keyword which a user does not intend is shown more.

FIG. 9 is a block diagram illustrating a configuration of the speech recognition apparatus 10a according to the second embodiment.
The voice recognition device 10a of the second embodiment is provided with a keyword selection frequency calculation unit 8 and a keyword selection frequency storage unit 9 added to the voice recognition device 10 shown in the first embodiment. In the following description, the same or corresponding parts as the components of the speech recognition apparatus 10 according to the first embodiment are denoted by the same reference numerals as those used in the first embodiment, and description thereof is omitted or simplified. .

  The keyword selection frequency calculation unit 8 refers to the keyword selection frequency stored in the keyword selection frequency storage unit 9 and indicates the number of times a certain keyword included in the recognition result extracted by the speech recognition unit 1 has been selected by the user. Calculate the selection frequency. The keyword selection frequency accumulation unit 9 accumulates the number of times a keyword included in the recognition result is selected based on whether the recognition result output from the score recalculation unit 7a or a function corresponding to the recognition result is selected. . The accumulation method can be appropriately configured. For example, by inputting the recognition result output from the score recalculation unit 7 to the keyword selection frequency calculation unit 8, the keyword selection frequency calculation unit 8 acquires and accumulates the keywords presented to the user. Further, the keyword selection frequency calculation unit 8 acquires information input from an external device that accepts an input operation from the user, specifies the keyword selected by the user, and adds the number of times the specified keyword is selected.

Next, the operation of the voice recognition device 10a will be described.
FIG. 10 is a flowchart showing the operation of the speech recognition apparatus 10a according to Embodiment 2 of the present invention. In the following, the same steps as those of the speech recognition apparatus 10 according to Embodiment 1 are denoted by the same reference numerals as those used in FIG. 3, and the description thereof is omitted or simplified.
When the keyword length calculation unit 5 calculates the keyword length L _n in step ST5, the keyword selection frequency calculation unit 8 refers to the keyword selection frequency stored in the keyword selection frequency storage unit 9, and N extracted in step ST3. calculating a keyword selection frequency F _n of keywords contained in the pieces of recognition result (step ST41). Selection frequency F _n calculated in step ST41 is outputted to the score recalculation unit 7a. The detailed processing operation of step ST41 will be described later.

The score recalculation unit 7a is extracted in step ST3 using the number of unknown words C _n calculated in step ST4, the keyword length L _n calculated in step ST5, and the selection frequency F _n calculated in step ST41. The recognition score of N recognition results is updated (step ST42). The score recalculation unit 7a ranks the recognition results in descending order of the recognition score updated in step ST42, outputs the recognition result having the largest recognition score (step ST43), and ends the process. Detailed processing operations of step ST42 and step ST43 will be described later.

Next, the operation of the keyword selection frequency calculation unit 8 shown in step ST41 will be described in more detail.
FIG. 11 is a flowchart showing the operation of the keyword selection frequency calculation unit 8 of the speech recognition apparatus 10a according to the second embodiment.
The keyword selection frequency calculation unit 8 initializes the index n of the recognition result extracted in step ST3 to “1” (step ST51), initializes the index m of the word included in the recognition result n to “1”, and recognizes it. The keyword selection frequency F _n included in the result n is initialized to “−1” (step ST52).

Next, the keyword selection frequency calculation unit 8 refers to the keyword notation and keyword reading stored in the keyword storage unit 6, _{and the mth} word W _{n, m} of the _nth recognition result is the keyword storage unit 6. It is determined whether or not the keyword is stored in (Step ST53). When the word W _{n, m} is a keyword (step ST53; YES), the keyword selection frequency calculation unit 8 determines the keyword selection frequency F _n of the keyword included in the recognition result _n by the following expressions (6) and (7) ) Based on (step ST54).
F ' _n = UpdateFreq (F _n , Freq (w _{n, m} )) (6)
F _n = F ' _n (7)
In Formula (6) and Formula (7), F ′ _n is a keyword selection frequency after update, Freq (w) is a function that returns the selection frequency of keyword w, and UpdateLength (F _n , A) is the keyword selection frequency F _n . A function that performs an update.

式（８）においてＳ（ｗ）はキーワードｗの選択回数、Ｒ（ｗ）はキーワードｗの認識回数である。また、この実施の形態２ではUpdateFreq(F_n,A)はキーワード選択頻度Ｆ_ｎとＡの最大値演算を行う関数とする。In the first embodiment, a function represented by the following equation (8) is used as Freq (w).

In equation (8), S (w) is the number of times the keyword w is selected, and R (w) is the number of times the keyword w is recognized. In the second embodiment, UpdateFreq (F _n , A) is a function for calculating the maximum value of the keyword selection frequencies F _n and A.

After that, the keyword selection frequency calculation unit 8 adds 1 to the word index m included in the recognition result n (step ST55). On the other hand, when the word W _{n, m} is not a keyword (step ST53; NO), the process proceeds to step ST55 described above. Following step ST55, the keyword selection frequency calculation unit 8 determines whether or not the index m of the word included in the recognition result n is equal to or less than the number _{Mn of} words included in the recognition result n (step ST56). . When the index m of the word included in the recognition result n is less than or equal to the number _{Mn of} words included in the recognition result n (step ST56; YES), the process returns to step ST53.

On the other hand, when the index m of the word included in the recognition result n is larger than the number M _{n of} words included in the recognition result n (step ST56; NO), the keyword selection frequency calculation unit 8 adds 1 to the index n of the recognition result. (Step ST57). Further, the keyword selection frequency calculation unit 8 determines whether or not the recognition result index n is equal to or less than the number N of recognition results (step ST58). If the index n of the recognition result is equal to or less than the number N of recognition results (step ST58; YES), the process returns to step ST52. On the other hand, when the index n of the recognition result is larger than the number N of the recognition results (step ST58; NO), the keyword selection frequency calculation unit 8 uses the keyword selection frequency F _n obtained by the above-described processing as the keyword selection frequency calculation unit 8. (Step ST59), the process proceeds to step ST42 in the flowchart of FIG.

Next, the processing operation of the score recalculation unit 7a shown in step ST42 and step ST43 will be described in more detail.
FIG. 12 is a flowchart showing the operation of the score recalculation unit 7a of the speech recognition apparatus 10 according to the second embodiment. In the following, the same steps as those of the score recalculation unit 7 of the speech recognition apparatus 10 according to the first embodiment are denoted by the same reference numerals as those used in FIG. 7, and the description thereof is omitted or simplified.

When the index n of the recognition result is initialized to “1” in step ST31, the score recalculation unit 7a calculates the number of unknown words C _n calculated in step ST4, the keyword length L _n calculated in step ST5, and step ST41. The recognition score Score _n of the recognition result n extracted in step ST3 is updated based on the following formula (9) and formula (10) using the keyword selection frequency F _n thus performed (step S61).
Score ' _n = Rescore2 (Score _n , C _n , L _n , F _n ) (9)
Score _n = Score ' _n (10)
In Equation (9) and Equation (10), Score ′ _n represents the recognition score after update, and Rescore ₁ (Score _n , C _n , L _n ) is a function for updating the score.

上述した式（１１）の第２項から第４項は、従来技術である認識スコアに対するペナルティに相当する。α、β、γ、ＴＨ_Ｌ、ＴＨ_Ｆは予め実験により決定されたパラメータである。この実施の形態２では、例えばα＝１、β＝１０、γ＝５、ＴＨ_Ｌ＝０、ＴＨ_Ｆ＝０．５とする。なお、スコアの更新を行う関数は、Score_n、Ｃ_ｎ、Ｌ_ｎ、Ｆ_ｎの関数になっていれば式（１１）以外の関数を適用することが可能である。In the first embodiment, a function shown in the following equation (11) is used as a function for updating the score.

The second to fourth terms of the above-described formula (11) correspond to a penalty for the recognition score, which is a conventional technique. α, β, γ, TH _L , and TH _F are parameters determined in advance by experiments. In the second embodiment, for example, α = 1, β = 10, γ = 5, TH _L = 0, and TH _F = 0.5. As the function for updating the score, functions other than Expression (11) can be applied as long as the functions are Score _n , C _n , L _n , and F _n .

  Thereafter, the score recalculation unit 7a performs the processing from step ST33 to step ST36, outputs the recognition result, and ends the processing.

FIG. 13 is an example showing a recognition result after updating the recognition score in the speech recognition apparatus 10a according to the second embodiment. FIG. 13 shows the recognition rank, the recognition result, the recognition score before update, the number of unknown words C _n , the keyword length L _n , the keyword selection frequency F _n, and the recognition score after update.
The number of unknown words C _n of the recognition results “handed sushi” and “sushi sushi” is both “2”. On the other hand, the keyword length L _n included in the recognition result “hand-made rice cake” is “0”, and the keyword length L _n contained in “hand-made sushi rice” is “2”. Furthermore, the keyword selection frequency F _n included in the recognition result “hand-made rice cake” is “−1 (means that no keyword is included)”, and the keyword included in “hand-made sushi rice cake”. The selection frequency F _n is “0”.

Substituting these results into the above-mentioned formula (11), the recognition score after updating the recognition result “sushi on hand” is “110- (1 × 2)-(10/2) -5 = 98”. descend. On the other hand, the updated recognition score of the recognition result “Hold in hand” is “105” because L _n = 0. As a result, the recognition order based on the updated recognition score is reversed, and the first place becomes the recognition result “hands on rice” and the second place the recognition result “hands on sushi rice”. Moreover, when compared with the recognition result example of FIG. 6 described in Embodiment 1, by the penalty is added on the basis of the keyword selection frequency F _n that keyword selection frequency calculating unit 8 is calculated, the recognition result of FIG. 11, "hand The recognition score for “Nishi Sushi” is lower.

  As described above, according to the second embodiment, the keyword selection frequency calculation unit 8 that calculates the keyword selection frequency indicating the number of times a certain keyword included in the recognition result extracted by the speech recognition unit 1 is selected by the user is provided. Even if a keyword unintended by the user appears temporarily for an out-of-task utterance, a penalty will be added if the recognition result containing the keyword is not frequently selected by the user. The recognition score can be recalculated. Thereby, it can suppress that many keywords which a user does not intend with respect to the utterance outside a task appear.

  In Embodiments 1 and 2 described above, the case where the recognition target is a dish name and an accompanying expression of the dish name has been described as an example. However, the present invention is not limited to these recognition targets. .

  In the above-described first embodiment, the case where processing is performed in the order of calculation of the number of unknown words and calculation of the keyword length will be described as an example. In the above-described second embodiment, calculation of the number of unknown words, calculation of the keyword length, Although the case where processing is performed in the order of keyword selection frequency calculation has been described as an example, the calculation order is not limited to this.

  In addition to the above, within the scope of the present invention, the present invention can be freely combined with each embodiment, modified any component of each embodiment, or omitted any component in each embodiment. Is possible.

  The speech recognition apparatus according to the present invention is suitable for improving the recognition performance of an utterance including a keyword having a short keyword length, for example, by applying a speech recognition program as a software to a navigation system equipped with a speech recognition function or a device to which the speech recognition program is applied. ing.

  DESCRIPTION OF SYMBOLS 1 Speech recognition part, 2 Acoustic model storage part, 3 Language model storage part, 4 Unknown word number calculation part, 5 Keyword storage part, 6 Keyword length calculation part, 7, 7a Score recalculation part, 8 Keyword selection frequency calculation part, 9 Keyword selection frequency storage unit, 10, 10a speech recognition device, 20 memory, 30 processor.

Claims (7)

Language likelihood calculated based on the recognition result obtained by performing speech recognition of the input speech using a language model that has been learned for the set recognition target and an acoustic model that is modeled on the features of speech A speech recognition unit that calculates a recognition score of the recognition result from acoustic likelihood;
An unknown word number calculation unit that calculates the number of unknown words included in the recognition result acquired by the speech recognition unit based on the N-gram probability of the language model;
A keyword storage unit for storing keywords related to the set recognition target;
A keyword length calculation unit that calculates a keyword length indicating the length of the keyword when the keyword stored in the keyword storage unit is included in the recognition result acquired by the voice recognition unit;
The speech is reduced so that the recognition score is reduced according to an increase in the number of unknown words calculated by the unknown word number calculation unit, and the recognition score is reduced according to a decrease in keyword length calculated by the keyword length calculation unit. A speech recognition apparatus comprising: a score recalculation unit that recalculates a recognition score calculated by a recognition unit and outputs a recognition result acquired by the speech recognition unit based on the recalculated recognition score. The score recalculation unit, when the keyword length calculated by the keyword length calculation unit is larger than a set threshold, a value obtained by multiplying the number of unknown words calculated by the unknown word number calculation unit and a parameter, The speech recognition apparatus according to claim 1 , wherein a value obtained by dividing the keyword length is subtracted from a recognition score calculated by the speech recognition unit. The score recalculation unit, when the keyword length calculated by the keyword length calculation unit is larger than a set threshold, a value obtained by multiplying the parameter by the number of unknown words calculated by the unknown word number calculation unit and the parameter The speech recognition apparatus according to claim 1 , wherein a value obtained by multiplying the value divided by the keyword length is subtracted from the recognition score calculated by the speech recognition unit. A keyword selection frequency accumulating unit that accumulates the number of times the user selected a recognition result output by the score recalculation unit or a function corresponding to the recognition result;
When the recognition result acquired by the voice recognition unit includes the keyword stored in the keyword storage unit, the selection frequency indicating the number of times the keyword is selected is selected as the keyword selection stored in the keyword selection frequency storage unit. A keyword selection frequency calculation unit that calculates by referring to the frequency,
The score recalculation unit recalculates the recognition score calculated by the voice recognition unit based on the number of unknown words, the keyword length, and the selection frequency calculated by the keyword selection frequency calculation unit. The speech recognition apparatus according to claim 1. The score recalculation unit reduces the recognition score according to an increase in the number of unknown words calculated by the unknown word number calculation unit, and the keyword length calculated by the keyword length calculation unit or the keyword selection frequency calculation unit calculates The speech recognition apparatus according to claim 4, wherein the recognition score is reduced according to a decrease in the selected frequency. The speech recognition unit performs speech recognition of the input speech using the language model learned for the set recognition target and the acoustic model that models the features of the speech, and is calculated based on the obtained recognition results A recognition score of the recognition result is calculated from the language likelihood and the acoustic likelihood,
An unknown word number calculation unit calculates the number of unknown words included in the recognition result based on the N-gram probability of the language model,
The keyword length calculation unit calculates a keyword length indicating the length of the keyword when the keyword related to the set recognition target accumulated in advance is included in the recognition result,
Score recalculation unit reduces the recognition score in accordance with the increase in the number of unknown words the calculated, so as to reduce the recognition score in accordance with a decrease of the calculated keyword length, the recognition score A speech recognition method for recalculating and outputting the recognition result based on the recalculated recognition score. Language likelihood calculated based on the recognition result obtained by performing speech recognition of the input speech using a language model that has been learned for the set recognition target and an acoustic model that is modeled on the features of speech A speech recognition unit that calculates a recognition score of the recognition result from acoustic likelihood;
An unknown word number calculation unit that calculates the number of unknown words included in the recognition result acquired by the speech recognition unit based on the N-gram probability of the language model;
A keyword storage unit for storing keywords related to the set recognition target;
A keyword length calculation unit that calculates a keyword length indicating the length of the keyword when the keyword stored in the keyword storage unit is included in the recognition result acquired by the voice recognition unit;
The speech is reduced so that the recognition score is reduced according to an increase in the number of unknown words calculated by the unknown word number calculation unit, and the recognition score is reduced according to a decrease in keyword length calculated by the keyword length calculation unit. A speech recognition program for causing a computer to function as a score recalculation unit that recalculates a recognition score calculated by a recognition unit and outputs the recognition result based on the recalculated recognition score.

Images