JP2011203434A - Voice recognition device and voice recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce incorrect recognition and incorrect operation caused by noise and unintentional sound, in a voice recognition device.SOLUTION: The voice recognition device includes: a speech input unit for inputting a speech data; a word dictionary in which a plurality of words for voice recognition are registered by relating each of them to its reading information; a voice recognizer which performs voice recognition on input speech data that is input by a speech input unit by using the word dictionary, and which outputs a word corresponding to the reading information of an evaluation value of a predetermined value or more as a recognition result information; a recognition result storage unit for storing recognition result information which is output by the voice recognizer together with input time when the speech data are input to the speech input unit, as past recognition result information; and a final result determination unit for determining output information corresponding to new output recognition result information, based on the past recognition result information stored in the recognition result storage unit.

Description

  The present invention relates to a speech recognition apparatus and speech recognition method for recognizing speech signals.

  2. Description of the Related Art A speech recognition device that recognizes an input speech signal and outputs the recognition result is known. As an example of such a voice recognition device, the voice input through a sound collector such as a microphone or the voice of a digital voice file is used as input, and word reading information registered in a predetermined word dictionary in advance. Some of them determine similarity and output display information or identification information corresponding to reading information having similarity equal to or higher than a predetermined threshold as a recognition result.

  The recognition result recognized by the voice recognition device is input to, for example, a car navigation system or an automatic voice response system. In the device to which the recognition result of the speech recognition device is input, the corresponding processing is executed using the input recognition result as a trigger.

  For example, by outputting a control command for instructing operation of the device, a dial command corresponding to a person name or a number for designating the number of the called party as a recognition result of the speech recognition device, device control by voice input is performed. Can be done. The control command is a command for instructing a power on / off of the apparatus and an operation specific to the apparatus. The dial command is a command corresponding to a person name or a telephone number for specifying a communication destination when another device has a communication function through a line.

  In such a speech recognition apparatus, the input speech and the reading information of all the words registered in the word dictionary are compared, and the speech recognition result is obtained for those having high similarity. Therefore, when there are a large number of words registered in the word dictionary, a long time is required for the speech recognition processing, and the probability of misrecognition increases and the recognition accuracy decreases.

  In a speech recognition device, in order to improve the efficiency of speech recognition processing and improve recognition accuracy, a word that is frequently used is registered in a word dictionary based on past recognition results, and speech recognition is performed using this word dictionary. It has been proposed to perform processing (see Patent Document 1).

  In the speech recognition apparatus configured as described above, since words registered in the word dictionary are frequently used, the word coverage by the word dictionary can be increased even with a small vocabulary, and the recognition accuracy can be increased. Can be increased.

JP 2002-245078 A

  In an environment that is susceptible to noise from the surroundings, there is a possibility that the voice recognition device may be erroneously recognized by a voice that is not intended for voice input, or other devices may malfunction due to the recognition result of the voice recognition device. In particular, when the voice recognition device can always accept voice input, the voice recognition process may be started in response to ambient noise or the like. In this case, an unintended recognition result is obtained. It will be.

  For example, in a car navigation system including a voice recognition device, a driver can input a voice for designating a destination by uttering a place name. The voice recognition device performs voice recognition processing based on the input voice, and transmits the corresponding place name as a destination to the car navigation system. In the car navigation system, based on the input destination, the route to the destination and the prediction of the required time are calculated and displayed.

  At this time, it is conceivable that the voice recognition device starts the voice recognition processing by using the voice uttered by the passenger on the passenger seat or the rear seat as an input. When the speech recognition device starts speech recognition processing in response to such unintended speech, the output recognition result is discarded as misrecognition or activated based on the output recognition result Other processing will be reset as a malfunction.

  An object of the present invention is to reduce misrecognition or malfunction based on noise or voice that is not intended for voice input in a voice recognition device.

  A speech recognition device includes a speech input unit for inputting speech data, a word dictionary in which a plurality of words for speech recognition are registered in association with the reading information, and speech data input from the speech input unit is a word dictionary A speech recognition unit that calculates an evaluation value indicating a degree of similarity to the reading information of the word registered in the word and outputs a word corresponding to the reading information that is an evaluation value equal to or greater than a predetermined value as recognition result information, and a speech recognition unit When the recognition result information is newly output from the recognition result storage unit that stores the output recognition result information as past recognition result information together with the input time when the voice data is input to the voice input unit. A final result determination unit that determines output information corresponding to the newly output recognition result information based on past recognition result information stored in the recognition result storage unit.

  The speech recognition apparatus can determine this word as output information when a word newly included in the recognition result information is included in past recognition result information accumulated in the recognition result accumulation unit. Therefore, if the same word is not included in the past recognition result information, the newly output recognition result information may be regarded as noise or voice data not intended to be input and not output. it can.

  This makes it possible to avoid erroneous recognition due to noise even in an environment that is easily affected by noise from the surroundings, and to reduce malfunctions caused by voice data that is not intended for input.

FIG. 1 is a block diagram illustrating an example of an in-vehicle device including a voice recognition device. FIG. 2 is a block diagram showing a hardware configuration of the speech recognition apparatus 10. FIG. 3 is a functional block diagram of the speech recognition apparatus according to the first embodiment. FIG. 4 is a flowchart showing processing in the speech recognition apparatus 10 of the first embodiment. FIG. 5 is an explanatory diagram of a word information table showing a correspondence relationship with word reading information registered in the word dictionary 303. FIG. 6 is an explanatory diagram of the recognition result information output from the voice recognition unit 302. FIG. 7 is an explanatory diagram of past recognition result information accumulated in the recognition result accumulation unit 304. FIG. 8 is a functional block diagram of the speech recognition apparatus according to the second embodiment. FIG. 9 is a flowchart showing processing in the speech recognition apparatus 10 of the second embodiment.

  An embodiment of a voice recognition device will be described with reference to the drawings.

<Outline configuration>
As an example of a voice recognition device, a case will be described in which a control command is recognized, generated, and output to a car navigation device or other device that is mounted on a vehicle and interlocks according to a voice uttered by a user.

  FIG. 1 is a block diagram illustrating an example of an in-vehicle device including a voice recognition device.

  In the vehicle, a voice recognition device 10, a car navigation device 20, an audio device 30, a communication device 40, and other electrical components 50 are connected via a network 60.

  In the car navigation apparatus 20, recognizable control commands such as power on / off, current position display, destination setting, route search, and destination change are registered in advance.

  In the audio device 30, recognizable control commands such as power on / off, playback, stop, pause, previous song, next song, and the like are registered in advance.

  The communication device 40 is, for example, an in-vehicle phone that can perform a hands-free call, in which control commands that can be recognized are registered in advance, such as call destination designation, call start, and call end.

  The other electrical components 50 are electrical components mounted on a vehicle such as a car air conditioner, a wiper, and a headlight, for example, and include control commands such as power on / off, operation mode, speed, and strength.

  The voice recognition device 10 executes voice recognition processing based on the voice uttered by the user, and controls any of the car navigation device 20, the audio device 30, the communication device 40, and other electrical components 50 based on the recognition result. Determine whether it is a command and output a control command to the corresponding device. For example, the voice recognition device 10 reads the reading information that specifies the control commands for the car navigation device 20 and the control commands for instructing power on, power off, current location display, destination setting, route search, destination change, etc. Register it in the word dictionary. The voice recognition device 10 converts voice input from a user into a digital voice signal and performs voice recognition processing.

  In speech recognition processing, the distance or similarity between the input digital speech signal and each phoneme is usually calculated using an acoustic model that represents the distribution of feature quantities of the digital speech signal for each phoneme, and is stored in the word dictionary. An evaluation value indicating the degree of coincidence with each word is calculated while referring to the reading information of the registered word. In the speech recognition device 10, a word corresponding to the reading information whose evaluation value exceeds a predetermined threshold value or a word corresponding to the reading information having the highest evaluation value is output to the car navigation device 20 as a control command inputted by voice. .

  For other devices, read information for specifying that it is a control command for each device and read information for specifying that it is a control command for each device are prepared and input. The speech signal converted from speech to digital acoustic signal is speech-recognized and compared with the reading information registered in the word dictionary, the word corresponding to the reading information whose evaluation value exceeds a predetermined threshold or the most evaluated value A word corresponding to high reading information is output to each device as recognition result information.

  FIG. 2 is a block diagram showing a hardware configuration of the speech recognition apparatus 10.

  The hardware configuration to which the speech recognition device 10 is applied includes a CPU 11 constituted by a microprocessor, a ROM 12 that stores BIOS and various parameters, a RAM 13 that temporarily stores variable values and calculation values during application execution, It includes a hard disk drive (HDD) 14 in which various parameters necessary for application program data and device functions, various databases, word dictionaries, and the like are stored.

  The hardware configuration to which the voice recognition device 10 is applied includes a microphone 15 for collecting voice uttered by the user, a key input unit 16 for receiving input of control commands and various parameters, a liquid crystal display panel and plasma. A display unit 17 constituted by a display or an organic EL display, etc., music from an audio device, a speaker 18 for outputting voice guidance from a voice synthesis unit built in a car navigation device, etc. may be provided . Each part of such a hardware configuration is connected via a bus 19.

  The voice recognition device 10 can be application software that operates on the hardware configuration as described above, and can be a DLL (Dynamic Link Library) that is a program that can be used in common by a plurality of application software. Can do. All or a part of the hardware configuration can operate in cooperation with another device. For example, the hardware configuration can operate in cooperation with the hardware configuration of the car navigation device.

  The voice recognition device 10 described above can be mounted on a vehicle-mounted device together with a car navigation device, a voice recognition device incorporated in an automatic voice response device, a voice recognition device incorporated in a mobile phone or a PDA (Personal Digital Assistant), It can be realized as a digital signage or other speech recognition device that operates on a general personal computer.

<First Embodiment>
A first embodiment of a voice recognition device will be described.

  FIG. 3 is a functional block diagram of the speech recognition apparatus according to the first embodiment.

  The voice recognition device 10 includes a voice input unit 301, a voice recognition unit 302, a word dictionary 303, a recognition result storage unit 304, and a final result determination unit 305.

  The voice input unit 301 collects voice uttered by the user with a microphone, performs analog / digital conversion, and inputs the voice to the voice recognition unit 302 as a digital acoustic signal. The voice input unit 301 can also input voice data restored from a wav file or other digital voice data to the voice recognition unit 302.

  The voice recognition unit 302 performs voice recognition processing on the voice signal input from the voice input unit 301 using the word dictionary 303. The speech recognition unit 302 generates a phoneme model corresponding to the word reading information registered in the word dictionary 303 using a predetermined acoustic model (not shown), and compares the feature amount with the speech signal. An evaluation value indicating the degree of similarity is calculated. The speech recognition unit 302 outputs a word corresponding to reading information that is an evaluation value equal to or greater than a predetermined value as recognition result information.

  In the word dictionary 303, words for speech recognition are registered together with the reading information. The words registered in the word dictionary 303 correspond to control commands for instructing operations to the car navigation device or the like. For example, when a control command is output to the car navigation device 20, the control commands related to device operations such as power-on, power-off, current location display, destination search, route search, and destination change are used as words. The reading information corresponding to is registered in the word dictionary 303. Words registered in the word dictionary 303 can be selectively registered by the system designer of the device to which the control command is input, the user of the device, etc. It is also possible to register in the word dictionary 303 as a text file listing reading information, such as a (Comma Separated Values) format file.

  The recognition result accumulation unit 304 accumulates the recognition result information output by the voice recognition unit 302 as past recognition result information together with the input time when the voice data is input to the voice input unit 301. The recognition result accumulation unit 304 can be stored in a predetermined storage medium, such as a hard disk drive, flash memory, a network recording medium connected via a communication device, etc. Various configurations can be considered.

  When the recognition result information is newly output from the speech recognition unit 302, the final result determination unit 305 newly outputs the recognition result information based on the past recognition result information stored in the recognition result storage unit 304. The output information corresponding to is determined. For example, when the speech recognition unit 302 outputs new recognition result information, the final result determination unit 305 refers to the past speech recognition result information stored in the recognition result storage unit 304 and creates a new result within a predetermined time immediately before. If there is at least one recognition result information including the same word as the word included in the recognition result information, that word is determined as output information.

  FIG. 4 is a flowchart showing processing in the speech recognition apparatus 10 of the first embodiment.

  In step S401, the speech recognition apparatus 10 inputs a speech signal to the speech recognition unit 302 via the speech input unit 301. As described above, the voice input unit 301 collects a user's analog voice by a microphone and performs analog / digital conversion on a digital acoustic signal, or voice data restored from a wav file or other digital voice data. To enter.

  In step S403, the speech recognition apparatus 10 performs speech recognition processing on the speech signal input from the speech input unit 301 using the word dictionary 303. The speech recognition unit 302 refers to the word reading information registered in the word dictionary 303, compares the feature amount of the input speech signal with the feature amount of the phoneme model, and calculates an evaluation value indicating the degree of similarity To do. Furthermore, the speech recognition unit 302 determines that the word corresponding to the reading information whose evaluation value is equal to or greater than a predetermined value is a word that has been speech-recognized, and outputs this as recognition result information.

  FIG. 5 is an explanatory diagram of a word information table showing a correspondence relationship with word reading information registered in the word dictionary 303.

  The word information table 500 includes at least a notation information column 501 and a reading information column 503. The notation information column 501 stores information related to notation when a word is displayed on the display device. The reading information field 503 relates to word reading information and stores standard reading information in kana notation.

  The speech recognition unit 302 generates a phoneme model using the reading information of the words registered in the word dictionary 303, compares the feature amount with the feature amount of the input speech signal, and calculates an evaluation value. To do.

  When speech recognition processing based on standard reading information is performed according to the lack of pronunciation, the characteristics of pronunciation in each region, individual differences in pronunciation, etc., voice recognition may not be possible. The word information table 500 shown in FIG. 5 includes an extended reading information column 505 so that extended reading information can be registered based on pronunciation laziness, regionality of pronunciation, individual differences in pronunciation, and the like.

  FIG. 5 shows an example in which “Tokyo” is registered as standard reading information and “toki” is registered as extended reading information for the word “Tokyo”. The extended reading information is not limited to one, and two or more extended reading information can be registered.

  The voice recognition unit 302 uses both the reading information registered in the reading information column 503 that is standard reading information and the reading information registered in the extended reading information column 505 that is reading information in consideration of pronunciation laziness and the like. It is also possible to perform voice recognition processing by using and can also perform voice recognition processing using only one of them as necessary.

  In step S405, the speech recognition apparatus 10 stores the recognition result information output from the speech recognition unit 302 in the recognition result storage unit 304. As described above, the recognition result accumulation unit 304 can be a hard disk drive, a flash memory, a recording medium connected via a network, and other various recording media. As the recognition result information output from the voice recognition unit 302, The word information, its evaluation value, and the time input to the voice input unit 301 can be accumulated.

  In step S407, the speech recognition apparatus 10 refers to the recognition result information stored in the recognition result storage unit 304.

  The recognition result information output from the speech recognition unit 302 is stored in the recognition result storage unit 304 and also output to the final result determination unit 305. When new recognition result information is input from the speech recognition unit 302, the final result determination unit 305 refers to past recognition result information stored in the recognition result storage unit 304.

  The final result determination unit 305 determines whether or not the word included in the recognition result information newly output from the speech recognition unit 302 is included in the past recognition result information stored in the recognition result storage unit 304. If there is recognition result information including the same word in the past, the word is determined as output information and output as a final recognition result.

  The voice recognition unit 302 outputs recognition result information including a word that has been voice-recognized, its evaluation value (score), and the time when the voice signal is input to the voice input unit 301.

  FIG. 6 is an explanatory diagram of the recognition result information output from the voice recognition unit 302.

  The recognition result information output by the speech recognition unit 302 can be shown in a recognition result information table 600 as shown in FIG. 6, and the recognition result information table 600 includes a time information column 601, a first recognition result column 603, a second A recognition result column 605 and a third recognition result column 607 are provided. The time information column 601 stores the time when the audio signal is input to the audio input unit 301. The first recognition result column 603 to the third recognition result column 607 are provided with a word column 611 and a score column 613, and the word corresponding to the reading information that has been successfully recognized by voice recognition because the evaluation value is higher than a predetermined value. Is stored in the word column 611, and the corresponding evaluation value is stored in the score column 613.

  In the example shown in FIG. 6, the word “Kyoto” has the evaluation value “97” for the voice signal input by the voice input 301 at time 11:13:25, and is the word having the nearest reading information. And stored in the first recognition result field 603. Hereinafter, the reading information of the word “Hyogo” is stored in the second recognition result column 605 with the evaluation value “77”, and the recognition result information is not output in the third recognition result column 607 as not applicable.

  Here, as the evaluation value, it is assumed to use a normalized value so that the maximum feature value is 100 by adding the matching feature amount between the word reading information and the voice signal.

  FIG. 7 is an explanatory diagram of past recognition result information accumulated in the recognition result accumulation unit 304.

  The past recognition result information table 701 stored in the recognition result storage unit 304 is similar to the recognition result information table 600 shown in FIG. 6 in the time information column 701, the first recognition result column 703, the second recognition result column 705, A third recognition result column 707 is provided. The time information column 701 stores the time when the audio signal is input to the audio input unit 301. The first recognition result column 703 to the third recognition result column 707 include a word column 711 and a score column 713, and word notation information for the recognition result information output from the speech recognition unit 302 is stored in the word column 711. The corresponding evaluation value is stored in the score column 713.

  As shown in FIG. 7, the recognition result information output from the speech recognition unit 302 is accumulated in time series based on the time when the speech signal is input to the speech input unit 301. In the example shown in FIG. 7, the evaluation value of the reading information corresponding to the word “Tokyo” is “95” for the audio signal input by the audio input unit 301 at the time 11:11:56, The past recognition result information that the evaluation value of the reading information corresponding to the word “Kyoto” is “75” is accumulated. In addition, for the audio signal input by the audio input unit 301 at time 11:11:59, the evaluation value of the reading information corresponding to the word “Wakayama” is “81”, and the word “Okayama” Past recognition result information in which the evaluation value of the corresponding reading information is “80” and the evaluation value of the reading information corresponding to the word “Takayama” is “72” is accumulated. Similarly, the evaluation value of the reading information corresponding to the word “Wakayama” is “87” for the audio signal input by the audio input unit 301 at time 11:13:07, and the word “Okayama” Corresponds to the word “Kyoto” for the past recognition result information whose reading information evaluation value corresponding to is “70” and the voice signal input by the voice input unit 301 at the time 11:13:20 The past recognition result information in which the evaluation value of the reading information is “84” and the evaluation value of the reading information corresponding to the word “Tokyo” is “80” is accumulated.

  In step S409, the speech recognition apparatus 10 determines output information based on the newly output recognition result information and the past recognition result information stored in the recognition result storage unit 304, and outputs this.

  When new recognition result information is output from the speech recognition unit 302, the final result determination unit 305 refers to the recognition result storage unit 304, and includes a past result including the same word as the word included in the new recognition result information. It is determined whether the recognition result information is accumulated.

  The final result determination unit 305 can determine whether or not the same word is included in the recognition result information within a predetermined time from the time of the new recognition result information. For example, the final result determination unit 305 determines whether there is past recognition result information input within the past one minute from the time of new recognition result information among the recognition result information accumulated in the recognition result accumulation unit 304. Can be determined.

  When the recognition result information as shown in FIG. 6 is output from the speech recognition unit 302, the final result determination unit 305 refers to the recognition result storage unit 304 to obtain past recognition result information as shown in FIG. Suppose. In this case, the words included in the recognition result information newly output from the speech recognition unit 302 are “Kyoto” and “Hyogo”, and the evaluation values are “97” and “77”, respectively. Of the past recognition result information stored in the recognition result storage unit 304, the recognition result information at time 11:13:20 includes the word “Kyoto”. “Kyoto” included in the recognition result information at 11:13:20 at this time is an evaluation value “84”, which indicates the highest evaluation value in the speech recognition processing at this time. In such a case, the final result determination unit 305 determines the new recognition result information and “Kyoto” included in the recognition result information within the past one minute as output information.

  In the example shown in FIG. 7, the past recognition result information input at time 11:11:56 contains the word “Kyoto” as the second recognition result, but it has been newly output. When the recognition result information is output within the past one minute of the recognition result information, the recognition result information is not adopted.

  In this way, by determining the output information based on the recognition result information output in the past within a predetermined time and the newly output recognition result information, sudden noise or speech that is not intended for input can be obtained. It is possible to avoid voice recognition and output as output information.

  For example, when a device such as a car navigation system accepts a voice command input, a specification is made so that the user continuously utters the same voice command twice.

  In this case, when the voice recognition unit 302 recognizes a word corresponding to the same command within one minute, the final result determination unit 305 can use the word as output information for the car navigation system. Therefore, even if a word recognized by the speech recognition unit 302 based on a word registered in the word dictionary 303 corresponds to a command of the car navigation system, the same word is recognized in the past within one minute. If not, the final result determination unit 305 regards the recognition result information output from the speech recognition unit 302 as being based on noise or an unintended speech signal and does not output it as output information.

  When the car navigation system is operated mainly by instructions from the user in the driver's seat, it is not intended based on the voice uttered by the passenger in the passenger seat or the rear seat. Even if the command is recognized by voice, it can be prevented from malfunctioning.

  In this way, even when sudden noise or speech from another speaker is input, it is not output as the final result of speech recognition, so the speech recognition device 10 can reduce misrecognition. In addition, malfunctions based on output information output from the speech recognition apparatus 10 can be reduced.

  6 and 7 show a case where a speech-recognized word and an evaluation value obtained by the speech recognition process are output as recognition result information output by the speech recognition unit 302. FIG. In such a case, the final result determination unit 305 may aggregate the evaluation values of recognition result information recognized in the past within a predetermined time for each word, and use the largest value as output information. it can.

  The word with the highest evaluation value of the recognition result information may be different for each utterance of the user. For example, in the utterance of a user who has made a voice input to the speech recognition device 10, the utterance is different between the first time and the second time, or the sound is slightly different due to changes in the physical state of the user. Can be considered.

  By summing the evaluation values included in the recognition result information for each word and determining the largest one as output information, even if there is a difference for each utterance of the user, appropriate output information can be obtained Can do.

  In the example of the car navigation system described above, a command input by voice is determined based on two utterances, but it is also possible to set the command to three or more times.

  Compared with handwritten character input using a keyboard or mouse, software keyboard or handwritten character input using a touch panel, or other character input methods, user input is less burdensome and multiple input is relatively easy. . Therefore, by performing voice recognition processing based on a plurality of utterances, it is possible to prevent erroneous recognition based on noise and voice that is not intended for input with a relatively simple configuration.

Second Embodiment
FIG. 8 is a functional block diagram of the speech recognition apparatus according to the second embodiment.

  Similar to the first embodiment, the speech recognition apparatus 10 includes a speech input unit 301, a speech recognition unit 302, a word dictionary 303, a score determination unit 801, a recognition result storage unit 304, and a final result determination unit 305.

  The voice input unit 301 inputs to the voice recognition unit 302 digital audio signals obtained by analog / digital conversion of voice uttered by the user, or voice data restored from a wav file or other digital voice data.

  The speech recognition unit 302 performs speech recognition processing on the speech signal input from the speech input unit 301 using the word dictionary 303, and outputs a word corresponding to the reading information that is an evaluation value equal to or greater than a predetermined value as recognition result information. .

  In the word dictionary 303, words for speech recognition are registered together with their reading information.

  The recognition result accumulation unit 304 accumulates the recognition result information output by the voice recognition unit 302 as past recognition result information together with the input time when the voice data is input to the voice input unit 301.

  The score determination unit 801 determines whether or not the word evaluation value included in the recognition result information output from the speech recognition unit 302 is equal to or greater than a predetermined value, and transmits the determination result to the final result determination unit 305.

  When the recognition result information is newly output from the speech recognition unit 302, the final result determination unit 305 newly outputs the recognition result information based on the past recognition result information stored in the recognition result storage unit 304. The output information corresponding to is determined. At this time, if the determination result output from the score determination unit 801 is a determination result indicating that the evaluation value of the word included in the recognition result information is greater than or equal to a predetermined value, the final result determination unit 305 Without referring to the storage unit 304, a word having an evaluation value equal to or greater than a predetermined value is determined as output information.

  FIG. 9 is a flowchart showing processing in the speech recognition apparatus 10 of the second embodiment.

  In step S901, the speech recognition apparatus 10 inputs a speech signal to the speech recognition unit 302 via the speech input unit 301. As described above, the voice input unit 301 collects a user's analog voice by a microphone and performs analog / digital conversion on a digital acoustic signal, or voice data restored from a wav file or other digital voice data. To enter.

  In step S903, the speech recognition apparatus 10 performs speech recognition processing on the speech signal input from the speech input unit 301 using the word dictionary 303. The speech recognition unit 302 refers to the word reading information registered in the word dictionary 303, compares the feature amount of the input speech signal with the feature amount of the phoneme model, and calculates an evaluation value indicating the degree of similarity To do. Furthermore, the speech recognition unit 302 determines that the word corresponding to the reading information whose evaluation value is equal to or greater than a predetermined value is a word that has been speech-recognized, and outputs this as recognition result information.

  In step S905, the speech recognition apparatus 10 stores the recognition result information output from the speech recognition unit 302 in the recognition result storage unit 304. As described above, the recognition result accumulation unit 304 can be a hard disk drive, a flash memory, a recording medium connected via a network, and other various recording media. As the recognition result information output from the voice recognition unit 302, The word information, its evaluation value, and the time input to the voice input unit 301 can be accumulated.

  In step S907, the speech recognition apparatus 10 determines whether or not the evaluation value included in the recognition result information is greater than or equal to a predetermined value. The score determination unit 801 determines whether or not the word evaluation value included in the recognition result information output from the speech recognition unit 302 is greater than or equal to a predetermined value, and outputs the determination result to the final result determination unit 305. If the evaluation value of the word included in the recognition result information is greater than or equal to the predetermined value, the process proceeds to step S911, and if not, the process proceeds to step S909.

  In step S909, the speech recognition apparatus 10 refers to the recognition result information stored in the recognition result storage unit 304. When new recognition result information is input from the speech recognition unit 302, the final result determination unit 305 refers to past recognition result information stored in the recognition result storage unit 304.

  In step S911, the speech recognition apparatus 10 determines output information to be output as a speech recognition result.

  When the determination result by the score determination unit 801 is a determination result that a word having an evaluation value equal to or greater than a predetermined value is not included in the recognition result information, the final result determination unit 305 is the same as in the first embodiment. In addition, output information is determined based on past recognition result information stored in the recognition result storage unit 304. When the determination result by the score determination unit 801 is a determination result that a word having an evaluation value equal to or greater than a predetermined value is included in the recognition result information, the final result determination unit 305 outputs the word Information is determined and output.

  For example, when an operation command is input to a car navigation system or the like by voice input, it is assumed that the user utters a word corresponding to the command relatively clearly. Therefore, when the user speaks to input an operation command, the recognition result information output from the speech recognition unit 302 includes a word having a relatively high evaluation value. For example, when normalization is performed so that the maximum value of the evaluation value is 100, if an evaluation value of 90 or more, or 95 or more is obtained, the word included in the recognition result information is used for normal speech input. Likely based on.

  On the other hand, when the speech recognition unit 302 performs speech recognition processing based on noise or speech that is not intended to be input, it is considered that the evaluation value included in the recognition result information is not so high. For example, when normalization is performed so that the maximum evaluation value becomes 100, it is unlikely that a word having an evaluation value of 95 or more is included in the recognition result information of the speech recognition unit 302 for suddenly generated noise. It is assumed.

  From this, when the evaluation value of the word included in the recognition result information is a certain high value, the final result determination unit 305 regards the word as being uttered by the user for the purpose of voice input. The output information can be output without referring to the recognition result storage unit 304.

  For example, in the example illustrated in FIG. 6, the recognition result information output from the speech recognition unit 302 includes the word “Kyoto” corresponding to the reading information whose evaluation value is “97”.

  For example, for a word having an evaluation value of 95 or higher with respect to an evaluation value having a maximum value of 100, the score determination unit 801 indicates a determination result indicating that the word has a high evaluation value as a final result determination unit 305. Will be output.

  As described above, for a word “Kyoto” whose reading information evaluation value is “97” as a word included in the recognition result information output from the speech recognition unit 302, the score determination unit 801 has an evaluation value of It is determined that the word is a high word, and a determination result to that effect is output to the final result determination unit 305.

  Since the final result determination unit 305 has obtained a determination result indicating that the word “Kyoto” included in the recognition result information is a word having a high evaluation value, the final result determination unit 305 does not refer to the recognition result accumulation unit 304. , “Kyoto” is determined as output information and output.

  As a result of performing speech recognition processing in the speech recognition unit 302, it is considered that a word corresponding to reading information with a high evaluation value is unlikely to be erroneously recognized as the same word within a predetermined time, and this is also used as output information. The possibility of misrecognition is low.

  As described above, a word corresponding to reading information having a high evaluation value can be determined as output information only by one utterance, and it is possible to reduce the time of the speech recognition processing, and the user's utterance. Can save labor.

<Other embodiments>
In the speech recognition apparatus according to the above-described embodiment, it is possible to provide a noise level measuring unit that measures the background noise level.

  In this case, when the background noise level measured by the noise level measurement unit exceeds a predetermined value, as described in the first embodiment, referring to past recognition result information within a predetermined time, the same word is If so, it can be configured to determine this as output information.

  Further, when the background noise level measured by the noise level measurement unit is less than or equal to a predetermined value, the word included in the recognition result information output from the speech recognition unit 302 is output as output information as it is. it can.

  When the background noise level is high to some extent, there is a possibility that the voice recognition process in the voice recognition unit 302 may be erroneously recognized even when the voice is intended for input. Accordingly, the final result determination unit 305 refers to the past recognition result information stored in the recognition result storage unit 304 and determines output information through multiple speech recognitions, thereby reducing the probability of erroneous recognition. it can.

  Further, when the background noise level is low, it is considered that erroneous recognition of the speech recognition processing in the speech recognition unit 302 is reduced. Therefore, even if the final result determination unit 305 determines output information based on the recognition result information of one speech recognition process, it is assumed that there is a low possibility of erroneous recognition.

  When the background noise level measured by the noise level measurement unit exceeds a predetermined value, an instruction for prompting the user to input the same voice a plurality of times may be displayed or notified by voice.

  In this way, when the background noise level is low, the user can speak only once, reducing the time for voice recognition processing and saving the user's time for voice input. It becomes.

  The present invention can be applied to devices capable of giving operation instructions by voice input, such as car navigation systems, digital signage, and mobile phones.

10: Voice recognition device
301: Voice input part
302: Voice recognition unit
303: Word dictionary
304: Recognition result storage
305: Final result judgment unit
801: Score determination unit

Claims (6)

A voice input unit for inputting voice data;
A word dictionary in which a plurality of words for speech recognition are registered in association with the reading information;
An evaluation value indicating a degree of similarity of speech data input from the speech input unit to reading information of a word registered in the word dictionary is calculated, and a word corresponding to the reading information having an evaluation value equal to or greater than a predetermined value is recognized. A voice recognition unit that outputs the result information;
A recognition result storage unit that stores recognition result information output by the voice recognition unit as past recognition result information together with an input time when the voice data is input to the voice input unit;
When recognition result information is newly output from the speech recognition unit, output information corresponding to the newly output recognition result information is obtained based on past recognition result information stored in the recognition result storage unit. A final result determination unit to be determined;
A speech recognition apparatus comprising:   The final result determination unit, when a word included in the recognition result information output by the speech recognition unit is included in at least one past recognition result information stored in the recognition result storage unit, The speech recognition apparatus according to claim 1, wherein words included in the recognition result information and the newly output recognition result information are determined as output information.   The final result determination unit includes at least one word included in at least one past recognition result information output within a predetermined time before the newly output recognition result information and the newly output recognition result information. The speech recognition apparatus according to claim 2, wherein when the word included in the word matches, the matched word is determined as output information.   The recognition result output from the speech recognition unit includes an evaluation value calculated for the reading information of the word, and the final result determination unit includes at least one past recognition stored in the recognition result storage unit. If there are two or more words included in the result information and words included in the newly output recognition result information, the evaluation values are totaled for each word, and the word having the highest total evaluation value is output. The speech recognition apparatus according to claim 3, wherein the speech recognition apparatus determines information. A score determination unit that determines whether or not an evaluation value of a word included in the newly output recognition result information is equal to or greater than a predetermined value;
The speech recognition apparatus according to claim 4, wherein the final result determination unit determines, as the output information, a word determined by the score determination unit as an evaluation value equal to or greater than a predetermined value. A speech recognition apparatus program for performing speech recognition processing on input speech data,
A voice input unit for inputting voice data;
A word dictionary in which a plurality of words for speech recognition are registered in association with the reading information;
An evaluation value indicating a degree of similarity of speech data input from the speech input unit to reading information of a word registered in the word dictionary is calculated, and a word corresponding to the reading information having an evaluation value equal to or greater than a predetermined value is recognized. A voice recognition unit that outputs the result information;
A recognition result storage unit that stores recognition result information output by the voice recognition unit as past recognition result information together with an input time when the voice data is input to the voice input unit;
When recognition result information is newly output from the speech recognition unit, output information corresponding to the newly output recognition result information is obtained based on past recognition result information stored in the recognition result storage unit. A final result determination unit to be determined;
A program for causing a computer to function as a speech recognition apparatus.

Images