JP5396530B2 - Speech recognition apparatus and speech recognition method - Google Patents

Description

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

  A speech recognition technique is known that generates and outputs a character string indicating the content of speech from speech input from a microphone. Specifically, by referring to the acoustic model and the language model and performing matching processing between the input speech and the character string, a statistically likely character string is output as a recognition result. The acoustic model indicates a correspondence relationship between voice feature amounts and characters and statistical information thereof. The language model indicates the connection relationship between character strings and its statistical information.

  In such speech recognition technology, when the speech is distorted due to ambient noise, or when a vocabulary that is not registered in the language model is pronounced, the speech is misrecognized and the wrong character string Is output as a recognition result. In this case, the user has to manually correct an incorrect character string in the output recognition result by operating an input device such as a keyboard.

  Therefore, in Patent Document 1, for the purpose of reducing such an operation burden on the user, a word having a near-competition probability with this word is presented as a correction candidate word with respect to the word of the recognition result, and the recognition result Has been disclosed that allows the user to select whether to convert this word into a correction candidate word. According to this technique, the user can correct a word as a recognition result only by selecting a correction candidate word, so that the operation burden on the user can be reduced.

JP 2006-146008 A

  However, in the technique described in Patent Document 1, both the speech recognition process and the correction process are performed based on the contents of one language model, so that a correct word that is still a correct word is registered in the language model. If not, the correct word cannot be presented to the user. Therefore, in the end, the user has to manually correct an incorrect character string in the output recognition result by operating an input device such as a keyboard.

  An object of the present invention is to determine conversion candidates for a character string that is erroneously recognized in the voice recognition process according to a different ground from the voice recognition process.

In order to solve the above-described problems, a speech recognition apparatus according to the present invention shows a speech data acquisition unit that acquires speech data, an acoustic model that indicates the correspondence between speech feature values and characters, and a connection relationship between character strings. A speech recognition unit that performs speech recognition processing on the speech data with reference to the first language model and generates a recognized character string indicating the recognition result; a character string designated by the user among the recognized character strings; Conversion target that determines a character string having a reliability as a recognition result lower than a predetermined threshold or a character string formed by a combination of character strings having a reliability as a recognition result lower than a predetermined threshold as a conversion target character string A character string determining unit and a character in a unit specified by the user from a character string connected immediately before or immediately after the determined character string to be converted in the recognized character string Or a reference string determining unit for determining the number of strings as the reference string, the user refers to the second language model that indicates the connection relationship of the string between extracted from the character string entered in the past, the voice Candidate conversion candidates for converting the conversion target character string into character strings that do not depend on the recognition result of the voice recognition processing for data and that indicate the connection relationship with the determined reference character string A conversion candidate character string determination unit that is determined as a character string, and an output unit that outputs the determined conversion candidate character string.

  Preferably, in the second language model, when the number of conversion candidate character strings out of a predetermined range with respect to the reference character string is obtained, the reference character string determination unit may The number of character strings connected immediately before or after the conversion target character string is increased / decreased, and the number of character strings after the increase / decrease is determined as a new reference character string. A character string indicating a connection relationship with the new reference character string in the bilingual model may be determined as the conversion candidate character string.

  Preferably, the output unit may convert the conversion target character string in the recognized character string into the conversion candidate character string and output the converted recognized character string.

  Preferably, the conversion candidate character string determination unit calculates a degree of correlation with the conversion target character string for each of the character strings that indicate a connection relationship with the reference character string in the second language model, It is preferable that at least the character string having the highest correlation degree or the character string having the correlation degree higher than the threshold value is determined as a candidate conversion candidate character string for converting the conversion target character string.

  Preferably, the output unit converts a conversion candidate character string from which phonemes that do not match the phonemes of the conversion target character string are deleted from the phonemes of the conversion candidate character string, or the conversion candidate character string and the reference character string. It is preferable to output a conversion candidate character string from which unnecessary characters included in the conversion candidate character string specified by performing a grammatical check on the connection relation with the character string are deleted.

  Preferably, a communication unit that communicates with the server device that stores the acoustic model and the first language model is included, and the voice recognition unit is stored in the server device by communication with the server device by the communication unit. The speech recognition process may be performed with reference to the acoustic model and the first language model.

The speech recognition method according to the present invention is a speech recognition method by a speech recognition apparatus, and includes an audio data acquisition step for acquiring audio data, an acoustic model and a character string indicating a correspondence relationship between a feature amount of audio and characters. A speech recognition process for performing speech recognition processing on the speech data and generating a recognized character string indicating the recognition result, and a user of the recognized character strings The character string to be converted is a specified character string, a character string whose reliability as a recognition result is lower than a predetermined threshold, or a character string consisting of a combination of character strings whose reliability as a recognition result is lower than a predetermined threshold. A conversion target character string determination step for determining as a string, and a character string connected immediately before or immediately after the determined conversion target character string in the recognition character string Thus, the second language indicating the connection relationship between the reference character string determining step for determining the character string or the number of character strings specified as the reference character string and the character strings extracted from the character strings previously input by the user With reference to the model, a character string that does not depend on the recognition result of the voice recognition process for the voice data and that indicates the connection relationship with the determined reference character string is converted into the character to be converted. A conversion candidate character string determining step for determining a conversion candidate character string as a candidate for converting a column, and an output step for outputting the determined conversion candidate character string.

  According to the present invention, it is possible to determine conversion candidates for a character string that is erroneously recognized in the speech recognition process according to a different ground from the speech recognition process.

1 shows a configuration of a speech recognition apparatus 100 according to a first embodiment. An example of the hardware constitutions of the speech recognition apparatus 100 is shown. 2 shows a functional configuration of the speech recognition apparatus 100. An example of a 2nd language model is shown. The procedure of the process by the speech recognition apparatus 100 is shown. An example of the speech recognition process by the speech recognition unit 314 is shown. An example of the reliability calculated by the speech recognition unit 314 is shown. An example of the correlation degree of a conversion candidate character string is shown. The output example of a conversion candidate character string is shown. A conversion example of the conversion target character string is shown. An example of the 2nd language model concerning a 2nd embodiment is shown. An example of the 2nd language model concerning a 3rd embodiment is shown. The function structure of the speech recognition apparatus 100 concerning 4th Embodiment is shown.

  The present invention can be readily understood by considering the following detailed description with reference to the accompanying drawings shown for the embodiments. Subsequently, embodiments of the present invention will be described with reference to the accompanying drawings. Where possible, the same parts are denoted by the same reference numerals, and redundant description is omitted.

(First embodiment)
First, the first embodiment will be described. FIG. 1 shows a configuration of a speech recognition apparatus 100 according to the first embodiment. The speech recognition device 100 is a device that recognizes input speech and outputs characters according to the recognized speech. In the first embodiment, a personal computer having a voice recognition function is used as the voice recognition apparatus 100.

  The speech recognition apparatus 100 includes a main body 110, a microphone 120, a display 130, a speaker 140, a keyboard 150, and a mouse 160. For example, when the user utters voice, the voice is input to the main body 110 by the microphone 120. The main body 110 recognizes the input voice by performing voice recognition processing on the input voice, and outputs a character corresponding to the recognized voice. For example, the main body 110 displays an image indicating the recognized character on the display 130 or causes the speaker 140 to emit a sound corresponding to the recognized character. When performing voice recognition processing, the keyboard 150 and the mouse 160 are used as input devices for performing an input when an instruction from the user is required.

  FIG. 2 shows an exemplary hardware configuration of the speech recognition apparatus 100. In addition to the main body 110, the microphone 120, the display 130, the speaker 140, the keyboard 150, and the mouse 160 described above, the voice recognition device 100 includes a CPU 1505, ROM 1510, RAM 1520, external memory drive 1540, and external memory 1542 inside the main body 110. A communication interface 1550 and an input / output device interface 1560.

  The ROM 1510, the RAM 1520, and the external memory 1542 store various data and various programs. The CPU 1505 executes various types of data processing and various types of hardware control by executing programs stored in the ROM 1510, the RAM 1520, or the external memory 1542.

  The communication interface 1550 controls communication with an external device. The external memory drive 1540 is connected to the external memory 1542 and reads / writes data from / to the external memory 1542. Examples of the external memory 1542 include a CD (Compact Disc), a DVD (Digital Versatile Disc), and a memory card.

  The input / output device interface 1560 controls input / output of data by various input / output devices connected to the main body 110. As described above, the microphone 110, the display 130, the speaker 140, the keyboard 150, and the mouse 160 are connected to the main body 110 as input / output devices. Therefore, the input / output device interface 1560 controls the input / output of data by these input / output devices.

  For example, in the speech recognition apparatus 100, the ROM 1510, the RAM 1520, or the external memory 1542 functions as an acoustic model storage unit 302, a first language model storage unit 304, and a second language model storage unit 306 described in FIG. . In the speech recognition apparatus 100, the CPU 1505 executes a speech recognition program stored in the ROM 1510, the RAM 1520, or the external memory 1542, and controls the microphone 120, thereby obtaining speech data described in FIG. The unit 312 functions. Further, in the speech recognition apparatus 100, the CPU 1505 executes the speech recognition program, so that the speech recognition unit 314, the conversion target character string determination unit 316, the reference character string determination unit 318, and the conversion described later with reference to FIG. It functions as a candidate character string determination unit 320. In the speech recognition apparatus 100, the CPU 1505 executes the speech recognition program and controls the display 130 or the speaker 140, thereby functioning as an output unit 322 described in FIG. In the speech recognition apparatus 100, the CPU 1505 executes the speech recognition program and controls the communication interface 1550, thereby functioning as the communication unit 330 described in FIG.

  The voice recognition program is provided to the user in a state where it is installed in the voice recognition device 100, for example. As another example, the voice recognition program may be stored in a computer-readable recording medium, provided to the user, and installed in the voice recognition apparatus 100. In addition to the above, the voice recognition program may be installed in the voice recognition device 100 by providing the program executed by the CPU 1505 to the user from an external device via a communication network.

  FIG. 3 shows a functional configuration of the speech recognition apparatus 100. Here, it demonstrates centering on the function relevant to a speech recognition process among the functions with which the speech recognition apparatus 100 is provided. The speech recognition apparatus 100 includes an acoustic model storage unit 302, a first language model storage unit 304, a second language model storage unit 306, a speech data acquisition unit 312, a speech recognition unit 314, a conversion target character string determination unit 316, and a reference character string. A determination unit 318, a conversion candidate character string determination unit 320, and an output unit 322 are provided.

  The acoustic model storage unit 302 stores a so-called acoustic model in which the correspondence between voice feature amounts and characters and statistical information thereof are shown. Examples of the voice feature amount include MFCC (Mel Frequency Cepstrum Coefficient). The first language model storage unit 304 stores a so-called language model in which connection relations between character strings and statistical information thereof are shown. The second language model storage unit 306 stores at least a language model indicating a connection relationship between character strings.

  Here, the connection relationship and the statistical information will be described. For example, it can be said that the character string “me” is connected to the character string “me” and the character string “ni”. The character string “I ga” can be said to be a character string “I” connected to the character string “ga”. Therefore, in the first language model and the second language model, the connection relationship between the character string “I” and the character string “ni”, the connection relationship between the character string “I” and the character string “GA”, etc. The connection relationship between each character string is shown. The statistical information indicates the probability of connection of character strings in units such as morphemes and words in a large number of text data. For example, in the above example, the probability that “I” is followed by “ni” and the probability that “I” is followed by “ga” in a large number of text data corresponds to this. However, the language model is not limited to the connection relationship between the two morphemes as in the above example, but the connection relationship may be maintained with “me” or “I am” as one character string, and the number of connections is 1, 2 Any of 3, 3,... Suffice as long as the speech recognition process can be executed.

  The language model stored in the second language model storage unit 306 is a language model provided separately from the language model stored in the first language model storage unit 304, and is stored in the first language model storage unit 304. It is intended to supplement the stored language model, but the content may be different from or the same as the first language model. Hereinafter, in order to distinguish these, the language model stored in the first language model storage unit 304 is referred to as “first language model”. The language model stored in the second language model storage unit 306 is referred to as a “second language model”. In the above description, the fact that the first language model and the second language model are provided separately means logically, and may be physically provided in the same device or provided in different devices. It may be done.

  FIG. 4 shows an example of the second language model. FIG. 4 shows an example of the second language model stored in the second language model storage unit 306. Here, as already explained, the second language model shows at least the connection relationship of the character strings. For example, the second language model is generated in advance by analyzing a text database input by the user in the past, or is updated sequentially by user input. The second language model may indicate statistical information in addition to character string connection relationships. In the second language model, in order to suppress the amount of information, connection relationships with a low statistical probability and old connection relationships in time series may be deleted. In the example shown in FIG. 4, the connection relationship between character strings in units of morphemes is shown as the second language model. For example, in the example illustrated in FIG. 4, “fu”, “city”, “station”, and “no” are associated as the second character string connected behind the first character string “Kyoto”. ing. For example, these connection relationships are analyzed by analyzing the character string “Kyoto prefecture is located in Kyoto city” and the character string “Kyoto station is the center of Kyoto” entered by the user in the past. The connection relationship between character strings in morpheme units included in the character string is added to the second language model. In this way, by forming the user's original second language model based on the character string input by the user in the past, a character string that is likely to be used again by the user is presented as a conversion candidate character string. it can.

  FIG. 5 shows a processing procedure by the speech recognition apparatus 100. Hereinafter, according to the processing procedure shown in FIG. 5, details of each functional unit shown in FIG. 3 will be described.

(Step S502)
The voice data acquisition unit 312 acquires voice data to be subjected to voice recognition processing. Specifically, the audio data acquisition unit 312 acquires the audio signal input from the microphone 120 as audio data to be subjected to the audio recognition process. When the voice data is already stored in a storage medium such as a memory, the voice data acquisition unit 312 may acquire the voice data stored in the storage medium as voice data to be subjected to voice recognition processing. .

(Step S504)
The speech recognition unit 314 performs speech recognition processing based on the acoustic model stored in the acoustic model storage unit 302 and the first language model stored in the first language model storage unit 304, thereby acquiring the speech acquired in step S502. The speech represented by the data is recognized, and a plausible character string is generated as a recognized character string indicating the recognition result. Here, an example of voice recognition processing by the voice recognition unit 314 will be described. FIG. 6 shows an example of voice recognition processing by the voice recognition unit 314. In this example, the voice recognition unit 314 generates a recognition character string “I will go to Kyoto now” from voice data “I will go to Kyoto station now”. In this example, the voice recognition unit 314 erroneously recognizes “station” as “target”. The reason for this is, for example, that the word “station” is not shown in the first language model, and the speech is distorted due to noise included in the voice data, so that “station” is the recognition result and its candidate. There are various reasons such as the absence. The configuration of the acoustic model, the configuration of the first language model, and the specific method of speech recognition processing by the speech recognition unit 314 are various. In the present embodiment, it is only necessary to be able to generate a recognized character string from speech data, and any of these may be used as long as this can be realized.

(Step S506)
The voice recognition unit 314 calculates the reliability of voice recognition for each of the character strings constituting the recognized character string generated in step S504. Here, an example of the reliability calculated by the speech recognition unit 314 will be described. FIG. 7 shows an example of the reliability calculated by the voice recognition unit 314. FIG. 7 shows the speech recognition reliability calculated by the speech recognition unit 314 for each of the morphemes constituting the recognized character string “going to Kyoto now” generated by the speech recognition unit 314. Here, the reliability indicates the degree of possibility that the character string is correctly recognized from the voice data. The higher the reliability, the higher the possibility that the character string is correctly recognized. For example, in the example shown in FIG. 7, “Now”, “From”, “Kyoto”, “Target”, “Ni”, “To”, The reliability is shown for each morpheme of “mas”. In this example, the reliability of “now” is “95”, the reliability of “kara” is “92”, the reliability of “Kyoto” is “80”, the reliability of “target” is “50”, "Is" 70 "," going "is" 93 ", and" mas "is" 93 ".

  For example, the voice recognition unit 314 calculates a higher reliability of the character string as the degree of coincidence between the feature amount of the voice data and the feature amount of the acoustic model is higher. In addition, the voice recognition unit 314 calculates the reliability of the character string higher as the appearance probability and appearance frequency of the character string are higher based on the statistical information indicated in the first language model. Furthermore, it is conceivable that the reliability is determined in combination with reference to both the acoustic model and the language model. For example, the recognition character string consisting of “Kyoto”, “Target”, “Ni”, “Go”, “Mas” is often used in combination with “Kyoto”, “Ni”, “Go” and “Mas”. If this is the case, the probability of appearance of these character strings is high, so that the reliability of these character strings is increased. On the other hand, with regard to “target”, if the combination of “Kyoto”, “go”, and “mas” is rarely used, it means that the occurrence probability of this character string “target” is low. Reduce the reliability of the target.

  In addition to the above, the reliability may be calculated higher as the grammatical unnaturalness is smaller, or the reliability may be calculated lower as the number of homonyms is higher. In short, the lower the possibility of being wrong, the higher the reliability of the character string may be calculated, and any method may be used.

(Step S508)
The conversion target character string determination unit 316 determines a conversion target character string to be converted from the recognized character strings generated in step S504. For example, the conversion target character string determination unit 316 selects, from among the recognized character strings generated by the speech recognition unit 314, a character string whose reliability as a recognition result is lower than a predetermined threshold, or a character string formed by a combination thereof. It is determined as a conversion target character string. When this method is applied, a character string that is highly likely to be wrong can be determined as a conversion target character string. Further, since the user is not allowed to select a conversion target character string, the operation burden on the user can be reduced. For example, “60” is set as the predetermined threshold value, and as shown in FIG. 7, “Now”, “From”, “Kyoto”, which constitutes the recognition character string “I will go to Kyoto now”, are formed. , “Target”, “ni”, “go”, and “mas”, the conversion target character string determination unit 316 includes, as the reliability, “ The “target” for which “50” lower than “60” is calculated is determined as the conversion target character string.

  As another example, the conversion target character string determination unit 316 causes the user to specify an arbitrary character string among the recognized character strings generated by the speech recognition unit 314, and converts the character string specified by the user into the conversion target character string. It may be determined as a column. For example, the conversion target character string determination unit 316 causes the display 130 to display the recognized character string generated by the voice recognition unit 314. The user uses the keyboard 150 or the mouse 160 to specify an arbitrary character string that is determined to be erroneously recognized with respect to the recognized character string displayed on the display 130. Then, the conversion target character string determination unit 316 determines an arbitrary character string specified by the user as described above as a conversion target character string. When this method is applied, the user selects a conversion target character string, and therefore a character string that is surely wrong can be determined as the conversion target character string. That is, the conversion target character string can be determined with high accuracy. Note that the method for specifying an arbitrary character string by the user is not limited to the above. For example, when a touch panel is provided on the surface of the display 130, the user may specify an arbitrary character string using the touch panel.

(Step S510)
The reference character string determination unit 318 determines, as a reference character string, a part of the character string connected before or after the conversion target character string determined in step S508 among the recognized character strings generated in step S504. To do. Specifically, the reference character string determination unit 318 refers to a character string that matches a predetermined condition among character strings connected before or after the recognized character string generated by the speech recognition unit 314. Determine as a string. This condition includes the direction of the reference character string based on the conversion target character string, the unit of the character string constituting the reference character string, and the number of character strings constituting the reference character string. For example, the direction of the reference character string based on the recognized character string is set to “front”, “rear”, or “both front and rear”. In addition, “character”, “word”, “morpheme”, “sentence”, etc. are set as the unit of the character string constituting the reference character string. Further, an arbitrary integer such as “1” is set as the number of character strings constituting the reference character string. By combining these, a condition such as “one morpheme in front of the character string to be converted as a reference character string” is stored in advance in a storage medium such as a memory. According to the condition that “one morpheme in front of the character string to be converted is a reference character string”, for example, as described above, the “target” among the recognized character strings “I will go to Kyoto now” Is determined as the conversion target character string, “Kyoto”, which is one morpheme in front of this “target”, is determined as the reference character string. Further, according to the condition that “one character in front of and behind the character string to be converted is a reference character string”, “Miyako” that is one character in front of this “target” and one character that is in the rear are “ "" Is determined as the reference character string.

(Step S512)
The conversion candidate character string determination unit 320 refers to the second language model, and determines a character string indicating a connection relationship with the reference character string determined in step S510 in the second language model in step S508. The conversion target character string is determined as a candidate conversion candidate character string to be converted. For example, as described above, it is assumed that “target” is determined as the character string to be converted and “Kyoto” is determined as the reference character string. Then, as shown in FIG. 4, it is assumed that the second language model shows the correspondence between the character string “Kyoto” and each of “fu”, “city”, “station”, and “no”. In this case, the conversion candidate character string determination unit 320 determines each of these “fu”, “city”, “station”, and “no” as conversion candidate character strings. The conversion candidate character string may include the same character string as the conversion target character string. For example, in the above example, the conversion target character string “target” may include the conversion candidate character string “target”.

  Here, the conversion candidate character string determination unit 320 may determine the plurality of conversion candidate character strings as candidate conversion candidate character strings for converting the conversion target character string. Further, the conversion candidate character string determination unit 320 sets a part of the plurality of conversion candidate character strings as candidate conversion candidate character strings for converting the conversion target character strings in order to suppress the number of output conversion candidate character strings. It may be determined.

  For example, the conversion candidate character string determination unit 320 calculates the degree of correlation with the conversion target character string for each of the plurality of character strings associated with the reference character string, and based on the calculated degree of correlation, A conversion candidate character string may be determined from the character string. For example, the conversion candidate character string determination unit 320 may determine the character string having the highest degree of correlation with the conversion target character string as the conversion candidate character string. Further, the conversion candidate character string determination unit 320 may determine a character string having a degree of correlation with the conversion target character string higher than a threshold as a conversion candidate character string. Alternatively, a predetermined number of character strings may be determined as conversion candidate character strings in descending order of correlation with the conversion target character string. Moreover, you may determine the character string whose correlation degree with a conversion object character string is more than a threshold value as a conversion candidate character string.

  FIG. 8 shows an example of the degree of correlation between conversion candidate character strings. FIG. 8 illustrates the conversion candidate character strings “fu”, “city”, “station”, and “no” determined by the conversion candidate character string determination unit 320 for the conversion target character string “target”. The correlation degree calculated by the conversion candidate character string determination unit 320 is shown. The degree of correlation here indicates the degree of similarity of the conversion target character string. For example, the conversion candidate character string determination unit 320 calculates a higher correlation degree between conversion candidate character strings having higher pronunciation similarities with the conversion target character string. In this example, the conversion candidate character string determination unit 320 calculates the similarity of pronunciation with the conversion target character string based on the number of phonemes (Roman characters indicating reading) corresponding to the conversion target character string. That is, the phoneme coincidence with the conversion target character string is calculated as the pronunciation similarity with the conversion target character string. For example, “target” phonemes are “t”, “e”, “k”, and “i”. On the other hand, the phonemes of “fu” are “h” and “u”. Thus, since the number of phonemes that match between “target” and “fu” is “0”, the conversion candidate character string determination unit 320 converts this “0” into the conversion candidate character string “fu”. Determined as the degree of correlation. On the other hand, phonemes of “station” are “e”, “k”, and “i”. Thus, since the number of phonemes that match “target” and “station” is “3”, the conversion candidate character string determination unit 320 converts this “3” into the conversion candidate character string “station”. Determined as the degree of correlation. In this way, the conversion candidate character string determination unit 320 calculates the degree of correlation of each of the plurality of conversion candidate character strings. Then, for example, the conversion candidate character string determination unit 320 selects a character string having the highest degree of correlation, a character string higher than the threshold value, or a predetermined number of character strings in order of the degree of correlation among the plurality of conversion candidate character strings. The conversion target character string is determined as a candidate conversion candidate character string to be converted. Note that the conversion candidate character string determination unit 320 calculates the similarity of pronunciation with the conversion target character string based on the number of readings that match the conversion target character string, and the like. It may be calculated by a method other than the method based on the degree of coincidence.

  The conversion candidate character string determination unit 320 does not determine the conversion candidate character string using the calculated correlation degree as it is, but multiplies or adds a predetermined coefficient to the calculated correlation degree. A conversion candidate character string may be determined. For example, the number of consonant matches is multiplied or added by a coefficient corresponding to this consonant, and the number of vowel matches is multiplied or added by a coefficient corresponding to this vowel. The degree of correlation may be determined automatically. For example, if the conversion target character string is “target” (teki) and the conversion candidate character string is “play” (geki), the correlation degree of the conversion candidate character string “play” is determined only by the number of phoneme matches. Becomes “3”. Of these, the number of consonant matches is “1”. The number of coincidence of vowels is “2”. For example, it is assumed that the coefficient corresponding to the consonant is “1” and the coefficient corresponding to the vowel is “2”. When the consonant coincidence number “1” is multiplied by a coefficient “1” corresponding to the consonant, the correlation degree based on the consonant coincidence number is “1”. Further, when “2” which is the number of coincidence of vowels is multiplied by a coefficient “2” corresponding to this vowel, the degree of correlation based on the number of coincidence of vowels is “4”. By summing up these correlations, the correlation of the final conversion candidate character string can be set to “5”.

(Step S514)
The output unit 322 outputs the conversion candidate character string determined in step S512. For example, the output unit 322 displays the conversion candidate character string together with the recognized character string on the display 130 so that the user can visually recognize the conversion candidate character string. Here, when there is one conversion candidate character string, the output unit 322 outputs in a form that the user can select whether or not to convert the conversion target character string in the recognized character string into a conversion candidate character string. When there are a plurality of conversion candidate character strings, the output unit 322 outputs the plurality of conversion candidate character strings in a form in which the user can select which conversion candidate character string is used to convert the conversion target character string. .

  Here, if there is one conversion candidate character string determined by the conversion candidate character string determination unit 320, the output unit 322 converts the conversion target character string in the recognized character string into this conversion candidate character string, and after conversion The recognized character string may be output. If there are a plurality of conversion candidate character strings, the output unit 322 may output the plurality of conversion candidate character strings in descending order of correlation. In this case, the output unit 322 may output only the conversion candidate character string having the highest degree of correlation, or may output a predetermined number of conversion candidate character strings in order of the degree of correlation. Conversion candidate character strings having a correlation degree equal to or higher than a threshold may be output in descending order. The output unit 322 may convert the conversion target character string in the recognized character string into a conversion candidate character string having the highest degree of correlation, and output the converted recognized character string.

  FIG. 9 shows an output example of the conversion candidate character string. A screen 900 illustrated in FIG. 9 is a screen displayed on the display 130 by the output processing by the output unit 322. This screen 900 is a screen for converting a conversion target character string included in the recognized character string into a conversion candidate character string. In this screen 900, “I will go to Kyoto now” is displayed as a recognized character string. Among these, the “target” that is the conversion target character string is highlighted by bold and underline so that the user can recognize that it is the conversion target character string. The screen 900 displays “fu”, “city”, “station”, and “no” as conversion candidate character strings. Here, as shown in FIG. 8, since the correlation degrees are obtained in advance for these conversion candidate character strings, these conversion candidate character strings are displayed on the screen 900 in descending order of the correlation degrees. ing. The user can select an arbitrary conversion candidate character string from a plurality of conversion candidate character strings, thereby converting the conversion target character string into the selected conversion candidate character string and confirming the recognized character string. . In addition, the user selects the “Confirm” button 910 without selecting any conversion candidate character string, so that the conversion target character string is not converted. Can be confirmed.

(Step S516)
When the character string to be converted is changed or not changed, and the recognized character string is confirmed, the speech recognition apparatus 100 stores the recognized character string in a storage medium such as a memory so that the user can use it. Or a series of speech recognition processing is completed.

  FIG. 10 shows a conversion example of the conversion target character string. On the screen 900 shown in FIG. 10, “I will go to Kyoto station now” is displayed as the recognition character string. This is because the conversion target character string “target” is converted into the conversion candidate character string “station” by the user selecting the conversion candidate character string “station” on the screen 900 shown in FIG. This is because the recognized character string is displayed. The user can confirm the recognized character string “I will go to Kyoto Station now” as the recognized character string by selecting the “confirm” button 910. Note that the speech recognition apparatus 100 does not convert the conversion target character string “target” by the user's selection as described above, but converts the conversion target character string “target” to the conversion candidate having the highest correlation. The character string “station” may be automatically converted, and the converted character string “I will go to Kyoto station now” after the conversion may be displayed from the beginning.

  As described above, the speech recognition apparatus 100 according to the first embodiment determines the reference character string connected to the conversion target character string, and the character whose connection relationship with the reference character string is indicated in the second language model. The sequence is output as a conversion candidate character string. That is, the speech recognition apparatus 100 according to the first embodiment can determine conversion candidates for a character string that is erroneously recognized in the speech recognition process according to a different basis from the speech recognition process. Thereby, even if the correct character string for converting the character string to be converted is not registered in the first language model referred to in the speech recognition process, the correct character string is registered in the second language model. If so, it can be output as a conversion candidate character string. Therefore, the speech recognition apparatus 100 according to the first embodiment can output a more appropriate character string as a correction candidate for a character string that is erroneously recognized in the speech recognition process.

(Second Embodiment)
Next, a second embodiment will be described. In the first embodiment, the connection relationship between character strings is shown in morpheme units in the second language model. On the other hand, in the second embodiment, the connection relationship between character strings in the second language model is shown in units of user input. When a user inputs a character string, the character string may be divided into a plurality of partial character strings and input step by step. For example, if you enter the string "Kyoto Prefecture is located in Kyoto City", this string will be changed to "Kyoto", "Funo", "Prefectural Office Location", "Ha", "Kyoto", "City" ”, Etc., and enter them step by step.

  FIG. 11 shows an example of the second language model according to the second embodiment. FIG. 11 shows an example of the second language model stored in the second language model storage unit 306. In this example, the connection relationship between character strings in user input units is shown as the second language model. Has been. For example, in the example shown in FIG. 11, as the second character string connected behind the first character string “Kyoto”, each of “funo”, “city is”, “station is”, “no” It is associated. For example, these connection relationships are analyzed by analyzing the character string “Kyoto prefecture is located in Kyoto city” and the character string “Kyoto station is the center of Kyoto” entered by the user in the past. Is added to the second language model as a connection relationship of character strings in user input units included in the character string.

  The partial character string delimiter is determined by, for example, the timing when the user presses the ENTER key, the unit in which the character string is converted, or the like. For example, if the input is made in the order of “Kyoto”, ENTER key pressed, “Fu no”, and Enter key pressed, a partial character string is determined in units of “Kyoto”, “Fu no”. Similarly, if "Kyoto Funo" is entered and then partial conversions are made in the order of "Kyoto Funo" and "Kyoto Prefecture", the unit is also "Kyoto" and "Fuino". A partial string is determined.

  Here, “target” is determined as the conversion target character string by the conversion target character string determination unit 316 from the recognized character string “I will go to Kyoto now”, and the reference character string determination unit 318 determines “ Assume that “Kyoto” is determined. As shown in FIG. 11, in the second language model, “fu”, “city”, “station”, and “no” are associated as character strings following the character string “Kyoto”. Suppose that Then, it is assumed that the conversion candidate character string determination unit 320 determines “station is” as a conversion candidate character string from these character strings. In this case, if the conversion target character string “Target” is converted to the conversion candidate character string “Station is” for the recognition character string “I will go to Kyoto now”, the converted recognition character string is “From now on I will go to Kyoto station”, and “ha” will be included extra. This extra character may be manually deleted by the user, but in this case, it is necessary to delete unnecessary characters manually in consideration of the context of the conversion candidate character string. End up.

  Therefore, in the second embodiment, after the output unit 322 automatically deletes the extra character string included in the conversion candidate character string in this way, the converted conversion candidate character string is output. It was. Hereinafter, the specific method will be described. The processing until the conversion candidate character string determining unit 320 determines the conversion candidate character string is as described above.

  First, based on the correlation between the conversion candidate character string and the conversion target character string, an unnecessary character included in the conversion candidate character string is automatically deleted, and the first method and the second method for deleting this are automatically recognized. A method will be described.

  The first method is a method of deleting unnecessary characters from the conversion candidate character string in accordance with the number of characters in the conversion target character string. For example, in the above example, the conversion target character string “target” is 2 characters in the reading character, and the conversion candidate character string “station” is 3 characters in the reading character. Accordingly, the conversion candidate character string “ For example, “ha”, which is the last character of “station”, is deleted.

  The second method is a method of leaving a portion having a high correlation with the conversion target character string from the conversion candidate character string and deleting the other portions. For example, in the above example, the phonemes (Roman characters indicating the reading) of the conversion target character string “ma” are “t”, “e”, “k”, and “i”. Moreover, the phonemes of the conversion candidate character string “station is” are “e”, “k”, “i”, “w”, and “a”. Among the phonemes of the conversion candidate character string “station is”, leave “e”, “k”, “i” as phonemes highly correlated with the phoneme of the conversion target character string “target”, and use other phonemes Delete certain “w” and “a”. As a result, “ha” is deleted from the conversion candidate character string “station is”.

  Next, a third method for automatically recognizing and deleting unnecessary characters included in the conversion candidate character string based on the connection relationship between the conversion candidate character string and the reference character string will be described.

  In the third method, a recognition character string after conversion is provisionally generated, and a grammar check is performed on this (particularly, the connection relationship between the conversion candidate character string and the reference character string). This is a method of identifying unnecessary characters and deleting them. For example, in the above example, for the recognized character string “I will go to Kyoto now”, the conversion target character string “target” is converted to the conversion candidate character string “station is”, and the recognition after conversion Temporarily generate "Kyoto station will go to now" as a character string. If the grammar check process is performed on this, it is clear that “ha” is unnecessary, and therefore “ha” is deleted from the conversion candidate character string “station”.

  As described above, the speech recognition apparatus according to the second embodiment is based on the correlation between the conversion candidate character string and the conversion target character string or the connection relationship between the conversion candidate character string and the reference character string. Unnecessary characters are automatically recognized and deleted, and then the conversion candidate character string after deletion is output, so the unnecessary characters are manually deleted in consideration of the context of the conversion candidate character string. It is possible to save the user's troubles such as.

(Third embodiment)
Next, a third embodiment will be described. In the first embodiment, when the connection relationship between a reference character string and a plurality of character strings is indicated in the second language model, a part or all of the plurality of character strings are output as conversion candidate character strings. It was. Here, if there are a large number of character strings that indicate the connection relationship with the reference character string, if all of the plurality of character strings are output as conversion candidate character strings, the user is confused. Also, if some of these multiple character strings are output as conversion candidate character strings, it takes time to determine conversion candidate character strings from a large number of character strings, as well as appropriate conversion. Candidate character strings cannot be determined.

  FIG. 12 shows an example of the second language model according to the third embodiment. FIG. 12 shows an example of the second language model stored in the second language model storage unit 306. For example, in the example shown in FIG. 12, the second character string “ni” is “Tokyo”, “Kyoto”, “Vice Head”, “Dog”, “Voice”, “School”, It has a connection relationship with a huge number of character strings such as “somewhere”. For this purpose, the second character string “ni” is used as a reference character string, and all of the first character strings having a connection relationship with the second character string “ni” are output as conversion candidate character strings. If this happens, the user will be confused. Further, if a part of the first character string having the connection relation with the second character string “ni” is output as a conversion candidate character string, it takes time to determine the conversion candidate character string. In addition, an appropriate conversion candidate character string cannot be determined.

  As described above, an appropriate conversion candidate character string may not be determined from one reference character string. Therefore, in the third embodiment, a plurality of reference character strings having different character string ranges are determined, and a reference character string from which an appropriate conversion candidate character string is obtained among the determined plurality of reference character strings is used. Thus, an appropriate conversion candidate character string can be determined by determining a candidate conversion candidate character string for converting the conversion target character string. Hereinafter, as an example of the specific method, when there are a large number of character strings having a connection relationship with the reference character string as described above, the range of the reference character string is expanded, and this is replaced with a new reference character string. Then, a description will be given of an example in which part or all of a character string having a connection relationship with a reference character string is output as a conversion candidate character string.

  In the third embodiment, in the second language model, when the reference character string determined by the reference character string determining unit 318 has a connection relationship with a predetermined number or more of character strings, the reference character string determining unit 318 The range of the character string as a reference character string in the sequence is expanded, and the expanded character string is determined as a new reference character string. Then, the conversion candidate character string determination unit 320 refers to the second language model, and uses the character string that indicates the connection relationship with the new reference character string in the second language model as the conversion candidate character string. decide.

  For example, the reference character string determination unit 318 converts from the recognized character string “I will go to Kyoto now” according to the condition “use one morpheme behind the character string to be converted as a reference character string”. Assume that “ni” following the target character string “target” is determined as a reference character string.

  In this case, the conversion candidate character string determination unit 320 refers to the second language model, and the connection relationship with the reference character string “ni” as a character string followed by the reference character string “ni” in the second language model. Are determined as conversion candidate character strings. For example, according to the second language model shown in FIG. 12, an enormous number of characters such as “Tokyo”, “Kyoto”, “President”, “Dog”, “Speech”, “School”, “Somewhere”,. A column is provisionally determined as a conversion candidate character string.

  Here, when the number of conversion candidate character strings tentatively determined is smaller than a predetermined number, part or all of the temporarily determined conversion candidate character strings are determined as formal conversion candidate character strings. On the other hand, when the number of conversion candidate character strings tentatively determined is larger than the predetermined number, the reference character string determination unit 318 sets the reference character string among the recognized character strings “I will go to Kyoto now”. The range of the character string is expanded, and the expanded character string is determined as a new reference character string. Specifically, the reference character string determination unit 318 determines a new reference character string by increasing the number of morphemes in the condition that “one morpheme behind the conversion target character string is a reference character string”. . For example, the reference character string determination unit 318 increments the number of morphemes in the condition that “one morpheme after the conversion target character string is a reference character string” by one, and “2 after the conversion target character string” The morpheme is used as a reference character string ”condition, and the corresponding“ ni-go ”is used as a new reference character string.

  In this case, the conversion candidate character string determination unit 320 refers to the second language model and sets the reference character string “ni-go” as a character string followed by the reference character string “ni-go” in the second language model. A plurality of first character strings for which connection relations are indicated are determined as conversion candidate character strings. For example, according to the second language model shown in FIG. 12, four character strings such as “Tokyo”, “Kyoto”, “school”, and “somewhere” are provisionally determined as conversion candidate character strings.

  Here, when the number of conversion candidate character strings tentatively determined is smaller than a predetermined number, part or all of the temporarily determined conversion candidate character strings are determined as formal conversion candidate character strings. If the number of conversion candidate character strings tentatively determined is larger than the predetermined number, the reference character string determination unit 318 determines that the reference character string of the recognized character strings “going to Kyoto now” and The expanded character string range is further expanded, and the expanded character string is determined as a new reference character string. For example, the reference character string determination unit 318 further increases the number of morphemes in the changed condition “use the two morphemes behind the conversion target character string as the reference character string”, The three morphemes behind are used as reference character strings, and “go to” corresponding to this is used as a new reference character string.

  In this case, the conversion candidate character string determination unit 320 refers to the second language model, and the reference character string “goes to” as the character string followed by the reference character string “goes to” in the second language model. Are determined as conversion candidate character strings. For example, according to the second language model shown in FIG. 12, two character strings such as “Tokyo” and “Kyoto” are provisionally determined as conversion candidate character strings.

  As described above, the speech recognition apparatus according to the third embodiment performs a simple process of extending the range of the reference character string until the number of character strings associated with the reference character string is less than a predetermined number. Thus, it is possible to narrow down an enormous number of conversion candidate character strings to conversion candidate character strings having an appropriate number and appropriate contents.

(Fourth embodiment)
Next, a fourth embodiment will be described. A part of each functional unit described in the embodiment is provided in an external information processing apparatus, and the speech recognition apparatus 100 refers to a part of data from the external information processing apparatus or external information processing apparatus. Alternatively, a part of the processing may be performed. That is, the speech recognition apparatus 100 is not limited to performing speech recognition processing alone as in the first embodiment, but performs communication with other information processing apparatuses to perform speech recognition processing. Also good. In the fourth embodiment, an example will be described in which the speech recognition apparatus 100 is configured to perform speech recognition processing by communicating with another information processing apparatus.

  FIG. 13 shows a functional configuration of the speech recognition apparatus 100 according to the fourth embodiment. In the fourth embodiment, the voice recognition device 100 becomes the client device 11 and constitutes the voice recognition system 10 together with the server device 12. In response to a request from the voice recognition device 100, the server device 12 performs voice recognition processing and transmits the processing result to the voice recognition device 100.

  FIG. 13 shows a functional configuration of the speech recognition apparatus 100 according to the fourth embodiment. In the fourth embodiment, the voice recognition device 100 includes a communication unit 330 that communicates with the server device 12. In addition, the acoustic model storage unit 302, the first language model storage unit 304, and the speech recognition unit 314 are stored in the server device 12. In response to this, the voice recognition device 100 passes voice data to the voice recognition unit 314 of the server device 12 through communication with the server device 12 by the communication unit 330, performs voice recognition processing, and the recognition result (recognized character). Column) is received from the server device 12.

  According to the speech recognition system 10 configured as described above, the acoustic model storage unit 302, the first language model storage unit 304, and the speech recognition unit 314 are provided in the server device 12, for example, a plurality of client devices 11 Is provided in the voice recognition system, the plurality of client apparatuses 11 can share the voice recognition process. As a result, the voice recognition processing result does not differ depending on the client device 11, and a uniform voice recognition processing result can be provided to the plurality of client devices 11. In addition, since the acoustic model and the first language model are unified, the acoustic model and the first language model can be easily maintained.

  Further, since the second language model storage unit 306 is not provided in the server device 12 but is provided in the speech recognition device 100, for example, when a plurality of client devices 11 are provided in the speech recognition system, the plurality of client devices 11 are provided. The contents of each of the second models can be made different according to the tendency of the user of each client device 11. Thereby, when there is a request for voice recognition processing in a certain client device 11, the voice recognition system 10 applies the second model according to the tendency of the user of the client device 11, and the text input by the user in the past Since the conversion candidate character string can be determined from among the character strings included in, a more appropriate processing result can be provided to the user of the client device 11.

(Modification)
The present invention is not limited to the embodiment described above, and may be modified as follows. Further, the following modifications may be combined.

(Modification 1)
In each embodiment, a personal computer having a voice recognition function is used as an example of a voice recognition device. However, the present invention is not limited to this, and the voice recognition device has the same voice recognition function as the voice recognition device 100 described in the embodiment. Any device such as a cellular phone, a PDA (Personal Digital Assistant), a navigation device, a portable music player, a notebook PC (Personal Computer), or a home appliance may be used as long as it can be realized.

(Modification 2)
In the third embodiment, a plurality of reference character strings having different character string ranges are determined, and conversion is performed using a reference character string from which the appropriate conversion candidate character string is obtained among the determined plurality of reference character strings. As a specific example of the configuration for determining the candidate conversion candidate character string for converting the target character string, the reference character string range is moved backward until the number of character strings associated with the reference character string is less than a predetermined number. Although the example of narrowing down the number of conversion candidate character strings by extending to the above has been described, the present invention is not limited to this. For example, the number of conversion candidate character strings may be narrowed down by extending the range of the reference character string forward or by extending it both forward and backward. If the number of character strings associated with the reference character string is less than the lower limit value, the number of conversion candidate character strings is reduced by reducing the range of the reference character string until it exceeds the lower limit value. You may make it increase. Also, reference is made until the number of character strings having a correlation degree equal to or higher than a predetermined correlation degree among the character strings associated with the reference character string is less than the upper limit value or more than the lower limit value. The number of conversion candidate character strings may be narrowed by expanding the range of character strings, or the number of conversion candidate character strings may be increased by reducing the range of reference character strings. Further, if the correlation degree of all the character strings is lower than a certain threshold value, the number of conversion candidate character strings is reduced by reducing the range of the reference character string until a character string having a higher correlation degree than the threshold value appears. May be increased. Also, if the correlation degree of all the character strings is higher than a certain threshold, the number of conversion candidate character strings is expanded by expanding the range of the reference character string until a character string having a lower correlation degree than this threshold appears. You may make it narrow down.

(Modification 3)
In the fourth embodiment, by sharing the voice recognition processing by the voice recognition unit 314 with a plurality of client devices 11, a configuration that can provide a uniform processing result to each of the plurality of client devices 11. For example, the speech recognition system 10 may have any device configuration, and any function may be provided for each device. For example, in the fourth embodiment, the speech recognition unit 314 is provided in the speech recognition device 100, and the speech recognition unit 314 is an acoustic model storage unit provided in the server device 12 through communication with the server device 12 by the communication unit 330. The acoustic model and the first language model may be referred to from the 302 and the first language model storage unit 304. Further, not only the voice recognition unit 314 but also the conversion target character string determination unit 316 and the reference character string determination unit 318 may be provided in the server device 12. In addition, the voice recognition device 100 is provided separately from the server device 12 and the client device 11. The voice recognition device 100 acquires voice data from the client device 11, and uses the second language model provided in the client device 11. It is also possible to refer to the speech recognition process described so far, and output the recognized character string and the conversion candidate character string to the client device 11 as the processing result.

(Modification 4)
In each embodiment, the speech recognition unit 314 may generate a plurality of recognized character strings. For example, the conversion target character string determination unit 316 uses the first candidate “I will go to Kyoto now” and the second candidate “Now” from the voice data “I will go to Kyoto station now”. For example, the recognition character string “I will go to Kyoto drama” will be generated.

  Here, the conversion target character string determination unit 316 determines “target” from the first candidate recognized character string as the first candidate conversion target character string, and sets “drama” as the second candidate recognized character string. Assume that the candidate character string to be converted is determined, and the reference character string determining unit 318 determines “Kyoto” common to these character strings to be converted as the reference character string. In this case, the conversion candidate character string determination unit 320 determines the correlation between the first candidate conversion target character string “target” from the character strings having the connection relationship with the reference character string “Kyoto”, and the second candidate. The conversion candidate character string may be determined based on both the degree of correlation with the conversion target character string “play”.

  In the above, for example, when each of “station”, “city”, “fu”, and “no” is determined as the conversion candidate character string, the output unit 322 uses the first candidate “from now on as Kyoto” as the recognition character string. The recognition character string “I will go” is output, and in addition to “station”, “city”, “fu”, “no” as the conversion candidate character string, the conversion candidate character string “play” of the second candidate May be output. Also, if the conversion candidate character string of the second candidate is the same “station” as the conversion candidate character string, the priority of the conversion candidate character string “station” is increased and the conversion candidate character string “ ”,“ City ”,“ fu ”, and“ no ”may be output. According to this, for example, even when the word “station” is pronounced, but the “station” becomes the second candidate conversion target character string due to the distortion of the voice, the priority is set in this way. By increasing, the possibility of being correctly corrected to “station” can be increased.

  DESCRIPTION OF SYMBOLS 100 ... Voice recognition apparatus, 110 ... Main body, 120 ... Microphone, 130 ... Display, 140 ... Speaker, 150 ... Keyboard, 160 ... Mouse, 302 ... Acoustic model storage part, 304 ... First language model storage part, 306 ... Second Language model storage unit, 312 ... voice data acquisition unit, 314 ... voice recognition unit, 316 ... conversion target character string determination unit, 318 ... reference character string determination unit, 320 ... conversion candidate character string determination unit, 322 ... output unit, 330 ... communication department

Claims (7)

An audio data acquisition unit for acquiring audio data;
A recognition character string indicating a recognition result by performing a voice recognition process on the voice data with reference to an acoustic model indicating a correspondence relation between a voice feature and a character and a first language model indicating a connection relation between the character strings. A voice recognition unit for generating
Among the recognized character strings, a character string specified by the user, a character string whose reliability as a recognition result is lower than a predetermined threshold, or a combination of character strings whose reliability as a recognition result is lower than a predetermined threshold A conversion target character string determination unit that determines a character string consisting of as a conversion target character string;
In the recognized character string, a reference character string that determines a character string of a unit or number specified by the user as a reference character string from a character string connected immediately before or after the determined character string to be converted A decision unit;
A character string that does not depend on the recognition result of the speech recognition process for the speech data , with reference to a second language model that indicates a connection relationship between the character strings extracted from a character string previously input by the user, and is determined A conversion candidate character string determining unit that determines a character string indicating a connection relationship with the reference character string as a conversion candidate character string of a candidate for converting the conversion target character string;
An output unit that outputs the determined conversion candidate character string. In the second language model, when the number of conversion candidate character strings out of a predetermined range with respect to the reference character string is obtained,
The reference character string determination unit
In the recognized character string, increase or decrease the number of character strings connected immediately before or immediately after the conversion target character string, determine the number of character strings after the increase or decrease as a new reference character string,
The conversion candidate character string determination unit
The speech recognition apparatus according to claim 1, wherein a character string that indicates a connection relationship with the new reference character string in the second language model is determined as the conversion candidate character string. The output unit is
The speech recognition apparatus according to claim 1, wherein the character string to be converted in the recognized character string is converted into the conversion candidate character string, and the converted recognized character string is output. The conversion candidate character string determination unit
For each of the character strings for which the connection relationship with the reference character string is indicated in the second language model, the degree of correlation with the conversion target character string is calculated, and at least the character string having the highest degree of correlation or the The speech recognition apparatus according to any one of claims 1 to 3, wherein a character string having a correlation degree higher than a threshold value is determined as a candidate conversion candidate character string for converting the conversion target character string. The output unit is
Among the phonemes of the conversion candidate character string, a conversion candidate character string from which a phoneme that does not match the phoneme of the conversion target character string is deleted, or a grammatical check on a connection relationship between the conversion candidate character string and the reference character string The speech recognition apparatus according to any one of claims 1 to 4, wherein a conversion candidate character string from which unnecessary characters included in the conversion candidate character string specified by performing is deleted is output. A communication unit that communicates with the server device storing the acoustic model and the first language model;
The voice recognition unit
6. The voice recognition process is performed by referring to the acoustic model and the first language model stored in the server device by communication with the server device by the communication unit. The speech recognition device according to any one of the above. A speech recognition method by a speech recognition device,
An audio data acquisition process for acquiring audio data;
A recognition character string indicating a recognition result by performing a voice recognition process on the voice data with reference to an acoustic model indicating a correspondence relation between a voice feature and a character and a first language model indicating a connection relation between the character strings. A speech recognition process for generating
Among the recognized character strings, a character string specified by the user, a character string whose reliability as a recognition result is lower than a predetermined threshold, or a combination of character strings whose reliability as a recognition result is lower than a predetermined threshold A conversion target character string determination step for determining a character string consisting of as a conversion target character string;
In the recognized character string, a reference character string that determines a character string of a unit or number specified by the user as a reference character string from a character string connected immediately before or after the determined character string to be converted A decision process;
A character string that does not depend on the recognition result of the speech recognition process for the speech data , with reference to a second language model that indicates a connection relationship between the character strings extracted from a character string previously input by the user, and is determined A conversion candidate character string determining step for determining a character string indicating a connection relationship with the reference character string as a conversion candidate character string of a candidate for converting the conversion target character string;
An output step of outputting the determined conversion candidate character string.

Images