JP2015055653A - Speech recognition device and method and electronic apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent the reduction in recognition rate of speech recognition processing regardless of difference of contents of choices or conditions such as a voice quality.SOLUTION: A speech recognition device includes: an acoustic model storage unit in which a plurality of sets of acoustic models obtained by collecting distributed states of frequency components of a plurality of phonemes used in a prescribed language, under a plurality of different conditions are stored; a choice list storage unit in which a choice list including choice data and acoustic model specification information is stored; an acoustic model selection unit which selects a set of acoustic models specified by the acoustic model specification information: a signal processing unit which extracts frequency components of a voice signal to generate a feature pattern; and a coincidence detection unit which performs speech recognition processing by comparing a feature pattern generated from at least a part of the voice signal with acoustic models corresponding to at least a part of the choice data, out of the selected set of acoustic models and outputs the speech recognition result.

Description

  The present invention relates to a speech recognition apparatus and a speech recognition method for recognizing speech uttered by a speaker and performing a response and processing corresponding to a speech recognition result. Furthermore, the present invention relates to an electronic device equipped with such a voice recognition device.

  For example, a voice recognition device having a voice playback function and a voice recognition function includes a host interface that communicates with a host system, and generates voice by receiving commands and voice playback data from the host system. The voice recognition device receives a command, a choice list or choice list designation information from the host system, and recognizes a voice by detecting a choice closest to the uttered voice from a plurality of choices included in the choice list. To transmit the voice recognition result to the host system.

  In addition, the speech recognition apparatus usually uses a sound model (also referred to as “standard pattern” or “template”) prepared in a storage unit such as a memory based on prerecorded speech, and the speech uttered by the speaker. Speech recognition processing is performed by performing pattern matching between a feature pattern obtained by analyzing the above and an acoustic model.

  However, depending on the speaker, when an acoustic model prepared in advance is used, the recognition rate in the speech recognition process may not be good. In such a case, the recognition rate can be improved or misrecognition can be improved by the speaker adaptation function that trains the acoustic model to adapt to the voice of the speaker.

  However, in general, since the speech recognition process is performed using the same acoustic model for unspecified speakers regardless of the contents of a plurality of options included in the option list, the contents of options or speakers The recognition rate may decrease depending on the voice quality and other conditions. Therefore, it is conceivable to perform speech recognition processing using a plurality of acoustic models.

  As a related art, Patent Document 1 discloses a speech recognition system that aims to improve recognition accuracy and simplify an operation until obtaining a correct recognition result. This speech recognition system uses a first speech recognition unit that performs speech recognition processing on speech stored in the speech storage unit using a speech recognition unit that stores speech uttered by a speaker and a first recognition dictionary. And a second voice recognition means for performing voice recognition processing on the voice stored in the voice storage means using a second recognition dictionary different from the first recognition dictionary, and the first and second voices A recognition result determining unit that determines a recognition candidate corresponding to the voice stored in the voice storage unit based on the recognition result of the recognition unit.

JP 2012-168349 A (paragraphs 0005-0006, 0011, FIG. 1)

  However, in the speech recognition system that performs speech recognition processing on the same speech using the first and second recognition dictionaries and the first and second speech recognition means, the first recognition dictionary is When the recognition result obtained by using and the recognition result obtained by using the second recognition dictionary are different from each other, there is a problem that it is difficult to determine a correct recognition result.

  Further, in Patent Document 1, when a plurality of recognition candidates are narrowed down by the speech recognition processing by the second speech recognition means, a first recognition dictionary corresponding to these recognition candidates is created and the first recognition is performed. It is also disclosed to extract the closest candidate to the input speech from among the recognition candidates by the speech recognition processing of the first speech recognition means using a dictionary. However, even if the input speech can be correctly recognized using the first recognition dictionary, if it is difficult to correctly recognize the input speech using the second recognition dictionary, a correct recognition result can be obtained. Most likely not.

  Accordingly, in view of the above points, one of the objects of the present invention is to select the content of an option or the speaker's content in a speech recognition process for detecting an option close to the uttered speech from among a plurality of options included in the option list. Even if conditions such as voice quality are different, the recognition rate is prevented from decreasing.

  In order to solve the above problems, the speech recognition apparatus according to the first aspect of the present invention is obtained by collecting frequency component distribution states of a plurality of phonemes used in a predetermined language under a plurality of different conditions. An option for storing an option list including an acoustic model storage unit for storing a plurality of sets of acoustic models, option data representing options, and acoustic model specifying information for specifying one set of acoustic models from the plurality of sets of acoustic models Features that represent a list storage unit, an acoustic model selection unit that selects a set of acoustic models specified by acoustic model specification information, and a frequency component of an input audio signal, and represents a distribution state of the frequency component of the audio signal A signal processing unit for generating a pattern, and a feature pattern generated from at least a part of the audio signal are selected from a set of selected acoustic models. Performs speech recognition processing as compared with the acoustic model corresponding to the part, and a coincidence detecting section for outputting a speech recognition result.

  Further, the speech recognition method according to the first aspect of the present invention includes: an option list storage unit that stores an option list including option data representing options and acoustic model specifying information for specifying a set of acoustic models; The step (a) of reading out the acoustic model specifying information, and from among a plurality of sets of acoustic models obtained by collecting the distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions, A step (b) of selecting a set of acoustic models specified by the acoustic model specifying information, a step of extracting a frequency component of the input voice signal, and generating a feature pattern representing a distribution state of the frequency component of the voice signal (C) and a feature pattern generated from at least a part of the audio signal is read in step (a) within the selected set of acoustic models. Compared with the acoustic model corresponding to at least part of the selection data issued performs speech recognition processing comprises a step (d) for outputting a speech recognition result.

  According to the first aspect of the present invention, a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions are prepared. One set of acoustic models specified by the acoustic model specifying information included in the option list is selected from the set of acoustic models. Thereby, since the speech recognition process is performed using the acoustic model set according to the conditions such as the contents of the option data included in the option list, the recognition rate in the speech recognition process can be improved.

  In the speech recognition apparatus according to the first aspect of the present invention, acoustics obtained by collecting a distribution state of phoneme frequency components included in unspecified types of terms used in a predetermined language by a plurality of acoustic models. A set of models and a set of acoustic models obtained by collecting distribution states of frequency components of phonemes included in specific types of terms used in a predetermined language may be included. In that case, speech recognition processing can be performed using an acoustic model set according to the type of term.

  The speech recognition apparatus according to the second aspect of the present invention stores a plurality of sets of acoustic models obtained by collecting frequency component distribution states of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities. An acoustic model storage unit, an option list storage unit that stores an option list including option data representing options, and a feature pattern that extracts a frequency component of an input audio signal and represents a distribution state of the frequency component of the audio signal A signal processing unit to be generated, a voice quality determination unit that determines the type of voice quality in the audio signal, and a set of acoustic models corresponding to the type of voice quality determined by the voice quality determination unit are selected from a plurality of sets of acoustic models An acoustic model selection unit, and a feature pattern generated from at least a part of the audio signal, at least of the selection data in the set of selected acoustic models. Performs speech recognition processing as compared with the acoustic model corresponding to the part comprises a coincidence detection unit for outputting a speech recognition result.

  The speech recognition method according to the second aspect of the present invention includes a step (a) of reading out option data from an option list storage unit that stores an option list including option data representing options, and an input voice signal A step (b) of extracting a frequency component and generating a feature pattern representing a distribution state of the frequency component of the audio signal; a step (c) of determining a voice quality type in the audio signal; and a plurality of types used in a predetermined language A set of acoustic models corresponding to the type of voice quality determined in step (c) is obtained from a plurality of sets of acoustic models obtained by collecting the distribution of the frequency components of phonemes for a plurality of different types of voice qualities. Selecting (d) a feature pattern generated from at least a portion of the audio signal within a set of selected acoustic models, step (a) Performs speech recognition processing as compared with the acoustic model corresponding to at least a portion of the selection data read Oite comprises a step (e) for outputting a speech recognition result.

  According to the second aspect of the present invention, a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities are prepared. A set of acoustic models corresponding to the type of voice quality determined based on the input audio signal is selected from the set of acoustic models. Thereby, since the speech recognition process is performed using the acoustic model set according to the type of voice quality of the speaker, the recognition rate in the speech recognition process can be improved.

  In the speech recognition apparatus according to the second aspect of the present invention, the option list storage unit stores an option list further including control information for permitting or prohibiting speech recognition processing for a specific voice quality type, and the match detection unit is Based on the type of voice quality determined by the voice quality determination unit and the control information, it may be determined whether or not to start the speech recognition process for the option list. In that case, it is possible to control the security level in operation of the electronic device by voice recognition. For example, by prohibiting children from setting the temperature of the air conditioner, it is possible to prevent dangerous operation of the air conditioner.

  The speech recognition apparatus according to the third aspect of the present invention stores a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions. An acoustic model storage unit, an option list storage unit that stores an option list including option data representing options, and a feature pattern that extracts a frequency component of an input audio signal and represents a distribution state of the frequency component of the audio signal Performs speech recognition processing by comparing the feature pattern generated from the signal processing unit to be generated and at least a part of the speech signal with an acoustic model corresponding to at least a part of the choice data among a plurality of sets of acoustic models. The coincidence detection unit that outputs a plurality of speech recognition results and a plurality of recognition accuracy corresponding to a plurality of sets of acoustic models, and the highest recognition accuracy among the plurality of recognition accuracy were obtained. The voice recognition result comprises a recognition accuracy determining unit which outputs as a final speech recognition result.

  The speech recognition method according to the third aspect of the present invention includes a step (a) of reading out option data from an option list storage unit that stores an option list including option data representing options, and an input voice signal A step (b) of extracting a frequency component and generating a feature pattern representing a distribution state of the frequency component of the audio signal; and a feature pattern generated from at least a part of the audio signal, a plurality of phonemes used in a predetermined language An acoustic model corresponding to at least a part of the option data read out in step (a) among a plurality of sets of acoustic models obtained by collecting the distribution state of the frequency components under a plurality of different conditions; A step (c) of performing a speech recognition process in comparison and outputting a plurality of speech recognition results and a plurality of recognition accuracy corresponding to a plurality of sets of acoustic models; The speech recognition result with the highest recognition certainty is obtained within the recognition accuracy, it comprises a step (d) to output as a final speech recognition result.

  According to the third aspect of the present invention, a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions are prepared. By performing speech recognition processing simultaneously using a set of acoustic models, a plurality of speech recognition results and a plurality of recognition accuracy corresponding to a plurality of sets of acoustic models can be obtained. Furthermore, the speech recognition result that provides the highest recognition accuracy among the plurality of recognition accuracy is output as the final speech recognition result. As a result, a speech recognition result with the highest recognition accuracy can be obtained in speech recognition processing performed simultaneously using a plurality of sets of acoustic models set according to a plurality of different conditions. The rate can be improved.

  In the speech recognition apparatus according to the third aspect of the present invention, a plurality of sets of acoustic models are obtained by collecting distribution states of frequency components of phonemes included in unspecified types of terms used in a predetermined language. A set of acoustic models and a set of acoustic models obtained by collecting the distribution states of the frequency components of phonemes included in specific types of terms used in a predetermined language may be included. In that case, speech recognition processing can be performed using an acoustic model set according to the type of term.

  Alternatively, the plurality of sets of acoustic models may be composed of a plurality of acoustic models obtained by collecting frequency component distribution states of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities. good. In that case, the speech recognition process can be performed using an acoustic model set according to the type of voice quality.

  An electronic apparatus according to one aspect of the present invention includes any one of the above speech recognition apparatuses. Thereby, operation by voice recognition can be realized in electronic devices such as home appliances, housing equipment, in-vehicle devices (navigation devices, etc.), vending machines, or mobile terminals.

The figure of the electronic device carrying the voice recognition device concerning a 1st embodiment of the present invention. The figure of the electronic equipment carrying the voice recognition device concerning the modification of a 1st embodiment. The figure which shows the speech recognition method implemented by the speech recognition apparatus shown in FIG. The figure which shows the example of the choice list | wrist stored in the choice list storage part shown in FIG. The figure of the electronic device carrying the speech recognition apparatus which concerns on the 2nd Embodiment of this invention. The figure which shows a part of structure of the speech recognition apparatus which concerns on the modification of 2nd Embodiment. The figure which shows the speech recognition method implemented by the speech recognition apparatus shown in FIG. The figure of the electronic device carrying the speech recognition apparatus which concerns on the 3rd Embodiment of this invention. The figure which shows the speech recognition method implemented by the speech recognition apparatus shown in FIG.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In addition, the same referential mark is attached | subjected to the same component and the overlapping description is abbreviate | omitted.
<Voice Recognition Device According to First Embodiment>
FIG. 1 is a block diagram showing a part of the configuration of an electronic apparatus equipped with the speech recognition apparatus according to the first embodiment of the present invention. This electronic device is, for example, a home appliance, a housing facility, a vehicle-mounted device (navigation device, etc.), a vending machine, a portable terminal, or the like. FIG. Has been.

  As shown in FIG. 1, the electronic device includes a voice recognition device 100 and a control unit 200 of the host system. The voice recognition device 100 issues a question or a message to the user according to the command received from the control unit 200, recognizes the user's voice, and transmits the voice recognition result to the control unit 200.

  The voice recognition apparatus 100 includes a command analysis unit 10, a voice input unit 20, an A / D converter 30, a voice recognition processing unit 40, a memory 50, a voice reproduction unit 60, and a D / A converter 70. And an audio output unit 80. A part of the command analysis unit 10 to the voice output unit 80 may be built in the semiconductor integrated circuit device.

  The control unit 200 includes a host CPU (central processing unit) 91 and a storage unit 92. The host CPU 91 operates based on software (voice recognition control program) recorded on the recording medium of the storage unit 92. As the recording medium, a hard disk, flexible disk, MO, MT, CD-ROM, DVD-ROM, or the like can be used.

  The host CPU 91 causes the voice recognition apparatus 100 to perform a voice reproduction operation by transmitting a voice reproduction start command and voice reproduction data to the voice recognition apparatus 100 in accordance with a scenario set in advance in the voice recognition control program. Further, the host CPU 91 causes the voice recognition apparatus 100 to perform a voice recognition operation by transmitting a voice recognition start command and option list designation information to the voice recognition apparatus 100.

  In the speech recognition apparatus 100, the command analysis unit 10 can analyze a command transmitted from the host CPU 91 and control the speech reproduction operation and the speech recognition operation independently or integrally according to the command. For example, in the voice recognition process, the number of options (words or sentences) that are candidates for voice recognition is limited by performing the voice reproduction operation and the voice recognition operation in accordance with a preset scenario in the voice recognition control program. It is possible to improve the recognition rate.

  The audio input unit 20 includes a microphone that converts audio into an electrical signal (audio signal), an amplifier that amplifies the audio signal output from the microphone, and a low-pass filter that limits the band of the amplified audio signal. . The A / D converter 30 samples the analog audio signal output from the audio input unit 20 and converts it into a digital audio signal (audio data). For example, the voice frequency band in the voice data is 12 kHz, and the number of bits is 16 bits.

  The speech recognition processing unit 40 includes a CPU and software, a digital circuit, or an analog circuit, and includes a signal processing unit 41, an acoustic model selection unit 42, and a coincidence detection unit 43. The memory 50 is constituted by, for example, a ROM (read only memory) or a flash memory, and includes an acoustic model storage unit 51 and an option list storage unit 52.

  The signal processing unit 41 extracts a frequency component of the audio signal by performing Fourier transform on the input audio signal, and generates a feature pattern representing the distribution state of the frequency component. The generated feature pattern is output to the coincidence detection unit 43. Further, the signal processing unit 41 activates the voice detection signal and outputs it to the coincidence detection unit 43 and the host CPU 91 when the level of the input voice signal exceeds a predetermined value. Thereby, the presence or absence of a request or answer from the user can be determined.

  Here, an example of a method for obtaining a feature pattern from an audio signal will be described. The signal processing unit 41 performs a filtering process on the input audio signal to emphasize high frequency components. Next, the signal processing unit 41 creates a plurality of frames by dividing the time-series audio signal at predetermined time intervals by applying a Hamming window to the audio waveform represented by the audio signal. Furthermore, the signal processing unit 41 extracts a plurality of frequency components by performing Fourier transform on the audio signal for each frame. Since each frequency component is a complex number, the signal processing unit 41 obtains an absolute value of each frequency component.

  The signal processing unit 41 integrates the absolute values of these frequency components over a frequency domain window determined based on the Mel scale (perceptual scale of pitch), thereby obtaining a number corresponding to the number of windows. Find the numerical value of. Further, the signal processing unit 41 takes the logarithm of those numerical values and performs a discrete cosine transform on the logarithmic values. Thereby, if there are 20 windows in the frequency domain, 20 numerical values are obtained.

  Of the numerical values obtained in this way, the lower ones (for example, 12) are called MFCC (Mel Frequency Cepstrum Coefficient). The signal processing unit 41 calculates a MFCC for each frame, and obtains a connected MFCC by connecting a plurality of MFCCs according to an HMM (Hidden Markov Model). This connected MFCC corresponds to a feature pattern and is represented as a point in a multidimensional space (for example, a 12-dimensional space).

  Here, “phoneme” means an element of a sound that is regarded as the same in a certain language. Below, the case where Japanese is used as a language is demonstrated. Japanese phonemes include “a”, “i”, “u”, “e”, “o” vowels, “k”, “s”, “t”, “n” and other consonants, The semi-vowels of “j” and “w” and the special mora of “N”, “Q”, and “H” are applicable.

  The acoustic model storage unit 51 is a part that stores a plurality of sets of acoustic models. The acoustic model here represents a distribution state of frequency components of phonemes used in a predetermined language, that is, a phoneme feature pattern. Different phonemes have different frequency components, resulting in different acoustic models. By collecting acoustic models for a plurality of phonemes under certain conditions, a set of acoustic models is obtained. Therefore, a plurality of sets of acoustic models can be configured by collecting acoustic models under a plurality of different conditions. For example, when a plurality of sets of acoustic models are configured on the condition of the types of terms, the acoustic model storage unit 51 uses a general-purpose set of acoustic models 0 and a specific term for a plurality of different types of terms. At least one set of acoustic models (in FIG. 1, a plurality of sets of acoustic models 1, 2,...) Are stored.

  The general-purpose set of acoustic models 0 includes a plurality of acoustic models obtained by collecting frequency component distribution states of a plurality of phonemes included in unspecified types of terms used in a predetermined language. Yes. Further, each set of acoustic models 1, 2,... For specific terms is obtained by collecting the distribution state of frequency components of a plurality of phonemes included in a specific type of term used in a predetermined language. A plurality of acoustic models. Thereby, speech recognition processing can be performed using the acoustic model set according to the type of term.

  For example, a set of acoustic models 1 for number recognition includes a plurality of acoustic models specially trained to recognize terms including numbers used in temperature setting, time setting, and the like. The set of acoustic models 2 for foreign words includes a plurality of acoustic models trained specifically for recognizing katakana terms in Japanese and terms including foreign words such as Japanese English.

  Here, each set of acoustic models is created in advance using a plurality of (for example, about 200) speakers uttered for a plurality of phonemes used in a predetermined language. In creating an acoustic model, an MFCC is obtained from a speech signal representing each phoneme. However, in the MFCC created using voices uttered by a large number of speakers, each numerical value varies.

  Therefore, the acoustic model for each phoneme has a spread including variation in the multidimensional space representing the MFCC. If the feature pattern generated from the audio signal input to the signal processing unit 41 is within the range of the acoustic model, it is determined that the feature pattern matches the acoustic model.

  The option list storage unit 52 stores an option list including a plurality of option data each representing a plurality of options and acoustic model specifying information for specifying a set of acoustic models. In FIG. 1, the option list storage unit 52 stores a plurality of option lists A, B, C,.

  For example, acoustic model specifying information for specifying a general-purpose set of acoustic models 0 is added to an option list including unspecified types of terms as options. In addition, in the option list including specific types of terms as options, acoustic model specifying information for specifying any one of a plurality of sets of acoustic models 1, 2,. Yes.

  When the command analysis unit 10 receives the voice recognition start command and the option list designation information from the host CPU 91, a plurality of option lists A, B, C,... Stored in the option list storage unit 52 according to the option list designation information are received. Specify one option list from

  The acoustic model selection unit 42 reads acoustic model specifying information included in the option list specified by the command analysis unit 10 from the option list storage unit 52, and sets a plurality of acoustic models stored in the acoustic model storage unit 51. A set of acoustic models specified by the acoustic model specifying information is selected.

  The coincidence detection unit 43 sequentially reads out a plurality of option data included in the option list specified by the command analysis unit 10 from the option list storage unit 52 and sets a set of acoustic models selected by the acoustic model selection unit 42. , An acoustic model (for example, an acoustic model for one syllable) corresponding to at least a part of the option data is read from the acoustic model storage unit 51.

  As a result, the coincidence detection unit 43 selects, by the acoustic model selection unit 42, a feature pattern generated from at least a part of the audio signal input to the signal processing unit 41 when the audio detection signal is activated. Speech recognition processing (pattern matching) in comparison with an acoustic model corresponding to at least a part of each option data included in the option list designated by the command analysis unit 10 in the set of acoustic models And output the speech recognition result.

  Here, the coincidence detection unit 43 reads out a plurality of acoustic models corresponding to a plurality of option data included in the option list from the set of acoustic models selected by the acoustic model selection unit 42, You may compare with a feature pattern and those acoustic models. Alternatively, the coincidence detection unit 43 reads out one acoustic model corresponding to one option data included in the option list from the set of acoustic models selected by the acoustic model selection unit 42, The operation may be repeated for a plurality of option data by comparing with the acoustic model.

  Further, when a match between the feature pattern and the acoustic model is detected, the match detection unit 43 may determine that the option corresponding to the acoustic model matches the input audio signal. In addition, when the coincidence between the feature pattern and the acoustic model is not detected, the coincidence detection unit 43 may output a speech recognition result indicating that speech recognition is impossible. Alternatively, the coincidence detection unit 43 obtains a distance between the position of the feature pattern in the multidimensional space representing the MFCC and the center of the acoustic model spread, and an option corresponding to the acoustic model closest to the position of the feature pattern is input. It may be determined that the audio signal matches the received audio signal.

  For example, the coincidence detection unit 43 uses a feature pattern generated from the first syllable of the input speech signal as an acoustic model corresponding to the first syllable of the options represented by each option data included in the option list. Compare with In the option list, when there is only one option having the head of the syllable in which the match is detected, the match detection unit 43 may determine that the option matches the input audio signal. On the other hand, in the option list, when there are a plurality of options having a syllable whose match is detected at the head, the match detection unit 43 determines the range of syllables from which a match should be detected until the option is narrowed down to one. You may enlarge.

  Here, the “syllable” means a set of sounds that are composed of one vowel as a main sound and that vowels alone or with one or more consonants before and after the vowel. Semi-vowels and special mora can also constitute syllables. That is, one syllable is composed of one or more phonemes. Japanese syllables include “a”, “i”, “u”, “e”, “o”, “ka”, “ki”, “ku”, “ke”, “ko”, etc. .

  For example, the acoustic model corresponding to the syllable “a” is an acoustic model representing the phoneme “a” constituting the syllable “a”. The acoustic model corresponding to the syllable “ka” includes the acoustic model representing the first phoneme “k” that constitutes the syllable “ka” and the second phoneme “a” that constitutes the syllable “ka”. It is a combination with an acoustic model that represents.

  When one syllable of the input speech signal is composed of one phoneme, if the phoneme match is detected, the syllable match is detected. On the other hand, when one syllable of the input speech signal is composed of a plurality of phonemes, if a coincidence of these phonemes is detected, a coincidence of syllables is detected.

  When a match as described above is detected between one option and the input audio signal, the match detection unit 43 selects an option for which a match is detected from among a plurality of options included in the option list. A speech recognition result including information to be specified (for example, option number or option data representing the option) is output. Thereby, the host CPU 91 can recognize an option corresponding to at least a part of the voice signal input to the voice recognition processing unit 40.

  Based on the voice recognition result, the host CPU 91 transmits voice reproduction data representing a question or a message together with a voice reproduction start command to the command analysis unit 10 according to a scenario set in advance in the voice recognition control program. Further, the host CPU 91 transmits new option list designation information to the command analysis unit 10 together with the voice recognition start command.

  The command analysis unit 10 supplies the audio reproduction data to the audio reproduction unit 60 according to the audio reproduction start command. The audio reproduction data may be audio data according to a predetermined compression format or text data. The audio reproduction unit 60 generates an output audio signal representing the audio to be output based on the audio reproduction data supplied from the command analysis unit 10.

  When the voice reproduction data is text data, the voice reproduction unit 60 uses a voice synthesis database including voice data representing voice waveforms for various phonemes to generate words or sentences represented by the text data. An output audio signal is synthesized by connecting audio data for a plurality of contained phonemes.

  The D / A converter 70 converts the digital output audio signal supplied from the audio reproduction unit 60 into an analog output audio signal. The audio output unit 80 includes a power amplifier that amplifies the analog output audio signal supplied from the D / A converter 70 and a speaker that emits audio in accordance with the power-amplified output audio signal.

  The speaker outputs a question or message represented by the sound reproduction data transmitted from the host CPU 91 as sound. This creates a situation where a user's answer to a question or message that is uttered based on audio playback data is predicted to be one of several options, and applies an option list that includes option data representing those options be able to.

<Modification of First Embodiment>
FIG. 2 is a block diagram showing a part of the configuration of an electronic apparatus equipped with a speech recognition apparatus according to a modification of the first embodiment of the present invention. In the modification of the first embodiment, the storage unit 92 of the control unit 200 stores a plurality of option lists A, B, C,..., And the option list sequentially selected from them is a voice recognition device. 100. Accordingly, in the speech recognition apparatus 100, memories 50a and 50b are used instead of the memory 50 shown in FIG. The other points are the same as in the first embodiment.

  The host CPU 91 of the control unit 200 transmits a voice playback start command and voice playback data to the voice recognition device 100 in accordance with a scenario set in advance in the voice recognition control program, thereby performing voice playback operation on the voice recognition device 100. Let it be done. Further, the host CPU 91 selects one option list from among a plurality of option lists A, B, C,... Stored in the storage unit 92, and receives the voice recognition start command and the selected option list. By transmitting to the voice recognition apparatus 100, the voice recognition apparatus 100 is caused to perform a voice recognition operation.

  In the speech recognition apparatus 100, the memory 50 a is configured by, for example, a ROM (read only memory) or a flash memory, and includes an acoustic model storage unit 51. Further, the memory 50b is constituted by, for example, a RAM (Random Access Memory) or the like, and includes an option list storage unit 52.

  When the command analysis unit 10 receives the voice recognition start command and the option list from the host CPU 91, the command analysis unit 10 stores the received option list in the option list storage unit 52 according to the voice recognition start command. FIG. 2 shows a state where the option list storage unit 52 stores the option list A.

  The acoustic model selection unit 42 reads out acoustic model specifying information included in the option list stored in the option list storage unit 52, and from among a plurality of sets of acoustic models stored in the acoustic model storage unit 51, A set of acoustic models specified by the acoustic model specifying information is selected.

  The coincidence detection unit 43 sequentially reads a plurality of option data included in the option list stored in the option list storage unit 52, and among the set of acoustic models selected by the acoustic model selection unit 42, An acoustic model (for example, an acoustic model for one syllable) corresponding to at least a part of the option data is read from the acoustic model storage unit 51.

  As a result, the coincidence detection unit 43 selects, by the acoustic model selection unit 42, a feature pattern generated from at least a part of the audio signal input to the signal processing unit 41 when the audio detection signal is activated. Among the set of acoustic models, the speech recognition process (pattern) is compared with the acoustic model corresponding to at least a part of each option data included in the option list stored in the option list storage unit 52. Matching) and output the speech recognition result.

  For example, the coincidence detection unit 43 uses a feature pattern generated from the first syllable of the input speech signal as an acoustic model corresponding to the first syllable of the options represented by each option data included in the option list. Compare with In the option list, when there is only one option having the head of the syllable in which the match is detected, the match detection unit 43 may determine that the option matches the input audio signal. On the other hand, in the option list, when there are a plurality of options having a syllable whose match is detected at the head, the match detection unit 43 determines the range of syllables from which a match should be detected until the option is narrowed down to one. You may enlarge.

<Voice Recognition Method According to First Embodiment>
Next, a speech recognition method according to the first embodiment of the present invention will be described with reference to FIGS. FIG. 3 is a flowchart showing a speech recognition method implemented by the speech recognition apparatus shown in FIG. 1 or 2.

  In step S11 of FIG. 3, the match detection unit 43 starts from an option list storage unit 52 that stores an option list that includes a plurality of option data each representing a plurality of options and acoustic model specifying information that specifies a set of acoustic models. The acoustic model selection unit 42 reads out acoustic model specifying information while reading out a plurality of option data.

  In step S12, the acoustic model selection unit 42 selects a sound from a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions. A set of acoustic models specified by the model specifying information is selected. On the other hand, in step S13, the signal processing unit 41 extracts the frequency component of the input audio signal and generates a feature pattern representing the distribution state of the frequency component.

  In step S <b> 14, the coincidence detection unit 43 selects a feature pattern generated from at least a part of the audio signal input to the signal processing unit 41 within the set of acoustic models selected by the acoustic model selection unit 42. A speech recognition process is performed in comparison with an acoustic model corresponding to at least a part of each option data read in step S11, and a speech recognition result is output. Thereby, one speech recognition operation is completed.

<Specific Example of First Embodiment>
Next, a specific example of the speech recognition operation performed by the speech recognition apparatus according to the first embodiment of the present invention will be described with reference to FIGS. In the following, the case where the electronic device shown in FIG. 1 is an air conditioner and the speech recognition apparatus is applied to the operation of the air conditioner will be described.

  FIG. 4 is a diagram illustrating an example of an option list stored in the option list storage unit illustrated in FIG. 4A shows the contents of the option list A, and FIG. 4B shows the contents of the option list B. The option data representing the options includes data representing Roman notation or Kana notation that can identify the phonemes included in the options.

  For example, when the air conditioner is turned on, the host CPU 91 of the control unit 200 transmits voice reproduction data representing a question “Do you want to cool or heat?” Together with the voice reproduction start command to the command analysis unit 10. Further, the host CPU 91 transmits option list designation information for designating the option list A to the command analysis unit 10 together with the voice recognition start command.

  The command analysis unit 10 of the speech recognition apparatus 100 specifies an option list A from among a plurality of option lists stored in the option list storage unit 52 according to the option list specifying information. The option list A includes option data representing a plurality of options related to the operation setting of the air conditioner corresponding to the option number, and an acoustic model number as the acoustic model specifying information.

  The acoustic model selection unit 42 is a general-purpose set of acoustics identified by the acoustic model number “0” included in the option list A from among a plurality of sets of acoustic models stored in the acoustic model storage unit 51. Select model 0.

  The coincidence detection unit 43 includes the phonemes “r · e” and “d · a” included in the first syllables “re” and “da” of the option 1 “cooling” and the option 2 “heating” in the option list A. The acoustic model corresponding to each is read out from the general-purpose set of acoustic models 0 stored in the acoustic model storage unit 51.

  On the other hand, the command analysis unit 10 supplies the audio reproduction data to the audio reproduction unit 60 in accordance with the audio reproduction start command. The audio reproduction unit 60 generates an output audio signal based on the audio reproduction data and supplies it to the D / A converter 70. The D / A converter 70 converts the digital output audio signal into an analog output audio signal, and supplies the analog output audio signal to the audio output unit 80. As a result, the voice output unit 80 issues a question “Do you want to use cooling or heating?”.

  In response to a question issued from the voice output unit 80, the user says “I will cool”. The signal processing unit 41 reads the phoneme “r · e · i · b · o · u. For each, a feature pattern representing a distribution state of frequency components is generated.

  The coincidence detection unit 43 uses the characteristic pattern of the first phoneme “r” of the first syllable generated by the signal processing unit 41 as the first phoneme “r” of the first syllable of option 1 and option 2 and By comparing with the acoustic model of “d”, the coincidence of phoneme “r” is detected.

  When the phoneme in which the match is detected represents a consonant, the match detection unit 43 further compares the second phoneme of the first syllable. The coincidence detection unit 43 uses the characteristic pattern of the second phoneme “e” of the first syllable generated by the signal processing unit 41 as the second phoneme “e” of the first syllable of option 1 and option 2 and A phoneme “e” match is detected by comparison with the acoustic model “a”.

  Thereby, the coincidence of the first syllable “re” is detected. In this case, since there is only one option for which a match is detected, a speech recognition result is obtained here. Note that if there are a plurality of options for which a match is detected, it is not possible to recognize which one is applicable, so the match detection unit 43 generates a sound model corresponding to the phoneme included in the next syllable. Read out from a set of general-purpose acoustic models 0 stored in the model storage unit 51, and expand the range of syllables to be detected for coincidence.

  The coincidence detection unit 43 outputs a speech recognition result including information specifying option 1 “cooling” having the first syllable “re” from which coincidence is detected to the host CPU 91. The information specifying the option 1 “cooling” corresponds to, for example, the option number “1”, the romaji notation “reibou” of the phoneme included in the option, or a part “re” thereof.

  Thereby, the host CPU 91 of the control unit 200 can recognize the option 1 “cooling” corresponding to at least a part of the input audio signal. When the first speech recognition operation is thus completed, the host CPU 91 sets the operation of the air conditioner to “cooling”.

  Next, the host CPU 91 transmits voice reproduction data representing a question “How many times should the set temperature be?” To the command analysis unit 10 together with a voice reproduction start command. Further, the host CPU 91 transmits option list designation information for designating the option list B to the command analysis unit 10 together with the voice recognition start command.

  The command analysis unit 10 of the speech recognition apparatus 100 designates the option list B from among a plurality of option lists stored in the option list storage unit 52 according to the option list designation information. The option list B includes option data representing a plurality of options related to the temperature setting of the air conditioner corresponding to the option number, and an acoustic model number as the acoustic model specifying information.

  The acoustic model selection unit 42 is a set for recognizing a number specified by the acoustic model number “1” included in the option list B from among a plurality of sets of acoustic models stored in the acoustic model storage unit 51. The acoustic model 1 is selected.

  The coincidence detection unit 43 is included in the first syllables “ni”, “ni”,... Of option 1 “20 ° C.”, option 2 “21 ° C.”,. .. Are read out from the set of acoustic models 1 for number recognition stored in the acoustic model storage unit 51. The acoustic models corresponding to the phonemes “n · i”, “n · i”,. .

  On the other hand, the command analysis unit 10 supplies the audio reproduction data to the audio reproduction unit 60 in accordance with the audio reproduction start command. The audio reproduction unit 60 generates an output audio signal based on the audio reproduction data and supplies it to the D / A converter 70. The D / A converter 70 converts the digital output audio signal into an analog output audio signal, and supplies the analog output audio signal to the audio output unit 80. As a result, the voice output unit 80 issues a question “How many times should the set temperature be?”.

  In response to a question issued from the voice output unit 80, the user says “set to 27 ° C.” and the signal processing unit 41 reads the phoneme “n.i.j.u.u.n.a.n”. For each of “a...”, A feature pattern representing a distribution state of frequency components is generated.

  The coincidence detection unit 43 uses the feature pattern of the first syllable “n” of the first syllable generated by the signal processing unit 41 as the first phoneme of the first syllable of option 1, option 2,. The phoneme “n” matches are detected by comparing with the acoustic models “n”, “n”,.

  When the phoneme in which the match is detected represents a consonant, the match detection unit 43 further compares the second phoneme of the first syllable. The coincidence detection unit 43 uses the characteristic pattern of the second phoneme “i” of the first syllable generated by the signal processing unit 41 as the second phoneme of the first syllable of option 1, option 2,. The phoneme “i” matches are detected by comparing with the acoustic models “i”, “i”,.

  In this case, since there are a plurality of options in which the match of the first syllable “ni” is detected and it cannot be recognized which matches, the match detection unit 43 is included in the next syllable. For each phoneme, the corresponding acoustic model is read out of the set of acoustic models 1 for number recognition stored in the acoustic model storage unit 51, and the range of syllables whose coincidence is to be detected is expanded.

  The coincidence detection unit 43 outputs to the host CPU 91 a speech recognition result including information for specifying option 8 “27 ° C.” having a plurality of syllables “Japanese” whose coincidence is finally detected. The information specifying the option 8 “27 ° C.” corresponds to, for example, the option number “8”, the romanization of the phoneme included in the option “nijuunanado”, or a part thereof “nijuuna”.

  Thereby, the host CPU 91 of the control unit 200 can recognize the option 8 “27 ° C.” corresponding to at least a part of the input audio signal. When the second speech recognition operation is thus completed, the host CPU 91 sets the set temperature of the air conditioner to “27 ° C.”.

  Further, the host CPU 91 may continue the next voice recognition operation or may end a series of voice recognition operations. When the series of voice recognition operations is completed, the host CPU 91 transmits voice reproduction data representing a message “acknowledged” together with the voice reproduction start command to the command analysis unit 10.

  The command analysis unit 10 supplies the audio reproduction data to the audio reproduction unit 60 according to the audio reproduction start command. The audio reproduction unit 60 generates an output audio signal based on the audio reproduction data and supplies it to the D / A converter 70. The D / A converter 70 converts the digital output audio signal into an analog output audio signal, and supplies the analog output audio signal to the audio output unit 80. As a result, the message “I understand” is issued from the voice output unit 80.

  According to the first embodiment of the present invention, a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions are prepared, One set of acoustic models specified by the acoustic model specifying information included in the option list is selected from the plurality of sets of acoustic models. Thereby, since the speech recognition process is performed using the acoustic model set according to the conditions such as the contents of the plurality of option data included in the option list, the recognition rate in the speech recognition process can be improved. .

<Voice Recognition Device According to Second Embodiment>
FIG. 5 is a block diagram showing a part of the configuration of an electronic apparatus equipped with a speech recognition apparatus according to the second embodiment of the present invention. In the second embodiment, a set of acoustic models is selected from a plurality of sets of acoustic models based on the characteristics of voice quality in the input audio signal.

  Therefore, a voice quality determination unit 44 and a voice quality determination information storage unit 53 are added to the first embodiment shown in FIG. On the other hand, each option list stored in the option list storage unit 52 may not include acoustic model specifying information. The other points are the same as in the first embodiment.

  The acoustic model storage unit 51 illustrated in FIG. 5 stores a plurality of sets of acoustic models obtained by collecting frequency component distribution states of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities. . Thereby, speech recognition processing can be performed using the acoustic model set according to the type of voice quality.

  For example, the acoustic model storage unit 51 may store a plurality of sets of acoustic models generated on the basis of voice signals obtained by recording voices of a plurality of groups of speakers having different ages or genders. In FIG. 5, the acoustic model storage unit 51 includes one set of acoustic models 0 for children, one set of acoustic models 1 for women, one set of acoustic models 2 for men, and one set for elderly people. The acoustic model 3 is stored. Here, “female” means a woman in a certain range of ages, and “male” means a man in a certain range of ages.

  The voice quality determination information storage unit 53 stores voice quality determination information representing characteristics of voice quality used for determining the type of voice quality of the speaker. The voice quality determination information is, for example, information that specifies the range of the fundamental frequency of the voice or the range of the formant frequency (resonance frequency of the vocal tract) according to the age or sex of the speaker.

  In FIG. 5, the voice quality determination information storage unit 53 includes voice quality determination information that represents the characteristics of the voice quality of the child, voice quality determination information that represents the characteristics of the female voice quality, voice quality determination information that represents the characteristics of the male voice quality, Voice quality determination information representing the characteristics of the voice quality. Here, “female” means a woman in a certain range of ages, and “male” means a man in a certain range of ages.

  The voice quality determination unit 44 determines the type of voice quality in the voice signal based on the voice quality feature of the voice represented by the input voice signal. For example, the voice quality determination unit 44 extracts a fundamental frequency, a formant frequency, or the like from the input voice signal and compares it with the frequency range represented by the voice quality determination information stored in the voice quality determination information storage unit 53. The type of voice quality (child, female, male, or elderly person) having the frequency range closest to the extracted frequency is determined, and the determination result is output to the acoustic model selection unit 42.

  According to the determination result output from the voice quality determination unit 44, the acoustic model selection unit 42 selects the voice quality type determined by the voice quality determination unit 44 from among a plurality of sets of acoustic models stored in the acoustic model storage unit 51. A corresponding set of acoustic models is selected.

  The coincidence detection unit 43 sequentially reads out a plurality of option data included in the option list specified by the command analysis unit 10 from the option list storage unit 52 and sets a set of acoustic models selected by the acoustic model selection unit 42. , An acoustic model (for example, an acoustic model for one syllable) corresponding to at least a part of the option data is read from the acoustic model storage unit 51.

  As a result, the coincidence detection unit 43 selects, by the acoustic model selection unit 42, a feature pattern generated from at least a part of the audio signal input to the signal processing unit 41 when the audio detection signal is activated. Speech recognition processing (pattern matching) in comparison with an acoustic model corresponding to at least a part of each option data included in the option list designated by the command analysis unit 10 in the set of acoustic models And output the speech recognition result.

  For example, the coincidence detection unit 43 uses a feature pattern generated from the first syllable of the input speech signal as an acoustic model corresponding to the first syllable of the options represented by each option data included in the option list. Compare with In the option list, when there is only one option having the head of the syllable in which the match is detected, the match detection unit 43 may determine that the option matches the input audio signal. On the other hand, in the option list, when there are a plurality of options having a syllable whose match is detected at the head, the match detection unit 43 determines the range of syllables from which a match should be detected until the option is narrowed down to one. You may enlarge.

  Here, the option list storage unit 52 may store an option list further including control information for permitting or prohibiting the voice recognition processing for a specific voice quality type. The control information may include information indicating permission or prohibition of voice recognition processing for each type of voice quality. For example, the option list B (see FIG. 4B) regarding the temperature setting of the air conditioner may include information that prohibits the voice recognition process for the child and permits the voice recognition process for the rest.

  In that case, the coincidence detection unit 43 determines whether or not to start the speech recognition processing for the option list based on the type of voice quality determined by the voice quality determination unit 44 and the control information included in the option list. judge. Thereby, it is possible to control the security level in the operation of the electronic device by voice recognition. For example, by prohibiting children from setting the temperature of the air conditioner, it is possible to prevent dangerous operation of the air conditioner.

  Also in the second embodiment, as in the modification of the first embodiment, the storage unit 92 of the control unit 200 stores a plurality of option lists A, B, C,. 100, the memories 50a and 50b shown in FIG. 2 may be used instead of the memory 50 shown in FIG.

<Modification of Second Embodiment>
FIG. 6 is a block diagram showing a part of the configuration of a speech recognition apparatus according to a modification of the second embodiment of the present invention. In the modification of the second embodiment, an acoustic model that has been trained (speaker optimization processing) to be adapted to a specific speaker is used. For this purpose, an acoustic model training unit 45 and a voice quality determination information extraction unit 46 are added to the second embodiment. The other points are the same as in the second embodiment.

  The acoustic model storage unit 51 shown in FIG. 6 stores a set of initial acoustic models for unspecified speakers. When training is performed by a specific speaker, the acoustic model training unit 45 uses a speaker adaptation function based on the feature pattern output from the signal processing unit 41 and the phoneme string information supplied from the outside to perform training training. Train a set of initial acoustic models to adapt to the person's voice.

  The acoustic model training unit 45 stores a set of trained acoustic models in the acoustic model storage unit 51. Thereby, the acoustic model storage unit 51 stores a plurality of sets of acoustic models obtained by collecting frequency component distribution states of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities. .

  On the other hand, the voice quality determination information extraction unit 46 extracts voice quality determination information representing the characteristics of the voice quality used for determining the voice quality of the training speaker, and stores the voice quality determination information in the voice quality determination information storage unit 53. The voice quality determination information is, for example, information that specifies the fundamental frequency range or formant frequency range of the training speaker's voice.

  When speech recognition is performed after training, the voice quality determination unit 44 determines whether the voice quality of the voice signal is the voice quality of the training speaker based on the voice quality characteristics of the input voice signal. For example, the voice quality determination unit 44 extracts a fundamental frequency, a formant frequency, or the like from the input voice signal and compares it with the frequency range represented by the voice quality determination information stored in the voice quality determination information storage unit 53. Then, it is determined whether or not the voice quality in the voice signal is the voice quality of the training speaker, and the determination result is output to the acoustic model selection unit 42.

  According to the determination result output from the voice quality determination unit 44, the acoustic model selection unit 42 selects the voice quality type determined by the voice quality determination unit 44 from among a plurality of sets of acoustic models stored in the acoustic model storage unit 51. A corresponding set of acoustic models is selected. That is, the acoustic model selection unit 42 selects a set of trained acoustic models stored in the acoustic model storage unit 51 when it is determined that the voice quality in the speech signal is the voice quality of the training speaker. On the other hand, the acoustic model selection unit 42 selects a set of initial acoustic models stored in the acoustic model storage unit 51 when it is determined that the voice quality in the voice signal is not the voice quality of the training speaker.

  The coincidence detection unit 43 sequentially reads out a plurality of option data included in the option list designated by the command analysis unit 10 (FIG. 5) from the option list storage unit 52 and 1 selected by the acoustic model selection unit 42. Among the set of acoustic models, an acoustic model (for example, an acoustic model for one syllable) corresponding to at least a part of the option data is read from the acoustic model storage unit 51.

  As a result, the coincidence detection unit 43 selects, by the acoustic model selection unit 42, a feature pattern generated from at least a part of the audio signal input to the signal processing unit 41 when the audio detection signal is activated. Speech recognition processing (pattern matching) in comparison with an acoustic model corresponding to at least a part of each option data included in the option list designated by the command analysis unit 10 in the set of acoustic models And output the speech recognition result.

  For example, the coincidence detection unit 43 uses a feature pattern generated from the first syllable of the input speech signal as an acoustic model corresponding to the first syllable of the options represented by each option data included in the option list. Compare with In the option list, when there is only one option having the head of the syllable in which the match is detected, the match detection unit 43 may determine that the option matches the input audio signal. On the other hand, in the option list, when there are a plurality of options having a syllable whose match is detected at the head, the match detection unit 43 determines the range of syllables from which a match should be detected until the option is narrowed down to one. You may enlarge.

<Voice Recognition Method According to Second Embodiment>
Next, a speech recognition method according to the second embodiment of the present invention will be described with reference to FIGS. FIG. 7 is a flowchart showing a speech recognition method implemented by the speech recognition apparatus shown in FIG.

  In step S21 of FIG. 7, the match detection unit 43 reads a plurality of option data from the option list storage unit 52 that stores an option list including a plurality of option data each representing a plurality of options.

  In step S22, the signal processing unit 41 extracts the frequency component of the input audio signal and generates a feature pattern representing the distribution state of the frequency component. In step S23, the voice quality determination unit 44 determines the type of voice quality in the input audio signal.

  In step S24, the acoustic model selection unit 42 performs a step from among a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities. A set of acoustic models corresponding to the type of voice quality determined in S23 is selected.

  In step S <b> 25, the coincidence detection unit 43 uses the feature pattern generated from at least a part of the audio signal input to the signal processing unit 41, among the set of acoustic models selected by the acoustic model selection unit 42. A speech recognition process is performed in comparison with an acoustic model corresponding to at least a part of each option data read in step S21, and a speech recognition result is output. Thereby, one speech recognition operation is completed.

  According to the second embodiment of the present invention, a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities are prepared, One set of acoustic models corresponding to the type of voice quality determined based on the input voice signal is selected from the plurality of sets of acoustic models. Thereby, since the speech recognition process is performed using the acoustic model set according to the type of voice quality of the speaker, the recognition rate in the speech recognition process can be improved.

<Voice Recognition Device According to Third Embodiment>
FIG. 8 is a block diagram showing a part of the configuration of an electronic device equipped with a speech recognition apparatus according to the third embodiment of the present invention. In the third embodiment, feature patterns generated from input speech signals are compared with a plurality of sets of acoustic models, and a speech recognition result with the highest recognition accuracy is output as a final speech recognition result. The

  For this purpose, the coincidence detection unit 43 in the first embodiment shown in FIG. 1 includes a plurality of parts (in FIG. 8, four parts 43a to 43d are shown), and a recognition accuracy judgment unit 47 is added. On the other hand, each option list stored in the option list storage unit 52 may not include acoustic model specifying information. The other points are the same as in the first embodiment.

  The acoustic model storage unit 51 illustrated in FIG. 8 stores a plurality of sets of acoustic models obtained by collecting frequency component distribution states of a plurality of phonemes used in a predetermined language under a plurality of different conditions. .

  For example, as in the first embodiment, the acoustic model storage unit 51 is obtained by collecting distribution states of frequency components of a plurality of phonemes included in an unspecified type term used in a predetermined language. A set of acoustic models and at least one set of acoustic models obtained by collecting the distribution states of frequency components of a plurality of phonemes included in a specific type of term used in a predetermined language. May be. In that case, speech recognition processing can be performed using an acoustic model set according to the type of term.

  Alternatively, as in the second embodiment, the acoustic model storage unit 51 collects a plurality of sets of frequency components obtained by collecting frequency component distribution states of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities. An acoustic model may be stored. In that case, the speech recognition process can be performed using an acoustic model set according to the type of voice quality.

  In FIG. 8, the acoustic model storage unit 51 includes one set of acoustic models 0 for children, one set of acoustic models 1 for women, one set of acoustic models 2 for men, and one set for elderly people. The acoustic model 3 is stored. Here, “female” means a woman in a certain range of ages, and “male” means a man in a certain range of ages.

  The coincidence detection units 43 a to 43 d sequentially read out a plurality of option data included in the option list specified by the command analysis unit 10 from the option list storage unit 52 and also store a plurality of sets stored in the acoustic model storage unit 51. Among the acoustic models, an acoustic model (for example, an acoustic model for one syllable) corresponding to at least a part of the option data is read out.

  As a result, the coincidence detection units 43a to 43d use the acoustic model storage unit 51 to generate a feature pattern generated from at least a part of the audio signal input to the signal processing unit 41 when the audio detection signal is activated. Among a plurality of sets of acoustic models stored in the voice recognition processing (comparison with an acoustic model corresponding to at least a part of each option data included in the option list specified by the command analysis unit 10 ( Pattern matching), and outputs a plurality of speech recognition results and a plurality of recognition accuracy corresponding to a plurality of sets of acoustic models.

  For example, the coincidence detection units 43a to 43d correspond to the feature pattern generated from the head syllable of the input audio signal to the head syllable of the option represented by each option data included in the option list. Compare with acoustic model. In the option list, when there is only one option having a syllable whose head is detected as a match, the match detection units 43a to 43d may determine that the option matches the input audio signal. . On the other hand, in the option list, when there are a plurality of options having a syllable whose match is detected at the head, the match detection units 43a to 43d select a syllable whose match should be detected until the options are narrowed down to one. The range may be expanded.

  Here, the coincidence detection units 43a to 43d obtain a distance D between the position of the feature pattern in the multidimensional space representing the MFCC and the center of the spread of the acoustic model, and the difference between the predetermined value E and the distance D (E -D) may be used as the recognition accuracy. Here, the predetermined value E represents the maximum value of the recognition accuracy when the distance D is zero. When the feature pattern and the acoustic model are compared for a plurality (N) of phonemes before the speech recognition result is obtained, the coincidence detection units 43a to 43d determine the distance obtained for each phoneme. The average value ΣD (i) / N of D (i) may be obtained, and the difference (E−ΣD (i) / N) between the predetermined value E and the average value ΣD (i) / N may be used as the recognition accuracy.

  In the example illustrated in FIG. 8, the coincidence detection unit 43 a compares the feature pattern generated from at least a part of the audio signal with the acoustic model included in the one acoustic model 0 for the child, and performs speech recognition processing. And the first speech recognition result and the first recognition accuracy are output. The coincidence detection unit 43b performs a speech recognition process by comparing the feature pattern generated from at least a part of the speech signal with the acoustic model included in the pair of acoustic models 1 for women, and performs the second speech The recognition result and the second recognition accuracy are output.

  Further, the coincidence detection unit 43c performs a speech recognition process by comparing the feature pattern generated from at least a part of the speech signal with the acoustic model included in the set of acoustic models 2 for men. The voice recognition result and the third recognition accuracy are output. The coincidence detection unit 43d performs a speech recognition process by comparing the feature pattern generated from at least part of the speech signal with the acoustic model included in the set of acoustic models 3 for the elderly, and performs the fourth speech The recognition result and the fourth recognition accuracy are output.

  The recognition accuracy determination unit 47 outputs a speech recognition result having the highest recognition accuracy among the plurality of recognition accuracy output from the coincidence detection units 43a to 43d as a final speech recognition result. For example, when the speaker is a child, the first recognition accuracy is generally higher than the second to fourth recognition accuracy. In this case, the recognition accuracy determination unit 47 performs the first recognition accuracy determination. Is output as the final speech recognition result.

  Also in the third embodiment, as in the modification of the first embodiment, the storage unit 92 of the control unit 200 stores a plurality of option lists A, B, C,. 100, the memory 50a and 50b shown in FIG. 2 may be used to store the acoustic model storage 51 and option list storage 52 shown in FIG.

<Voice Recognition Method According to Third Embodiment>
Next, a speech recognition method according to the third embodiment of the present invention will be described with reference to FIGS. FIG. 9 is a flowchart showing a speech recognition method performed by the speech recognition apparatus shown in FIG.

  In step S31 of FIG. 9, the match detection unit 43 reads a plurality of option data from the option list storage unit 52 that stores an option list including a plurality of option data each representing a plurality of options. In step S32, the signal processing unit 41 extracts the frequency component of the input audio signal, and generates a feature pattern representing the distribution state of the frequency component.

  In step S <b> 33, the coincidence detection units 43 a to 43 d convert the feature pattern generated from at least a part of the audio signal input to the signal processing unit 41 to the distribution state of the frequency components of a plurality of phonemes used in a predetermined language. Among a plurality of sets of acoustic models obtained by collecting under a plurality of different conditions, speech recognition processing is performed in comparison with an acoustic model corresponding to at least a part of each option data read in step S31. And output a plurality of speech recognition results and a plurality of recognition accuracy corresponding to a plurality of sets of acoustic models.

  In step S <b> 34, the recognition accuracy determination unit 47 outputs, as a final speech recognition result, a speech recognition result with the highest recognition accuracy among the plurality of recognition accuracy output from the match detection units 43 a to 43 d. . Thereby, one speech recognition operation is completed.

  According to the third embodiment of the present invention, a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions are prepared, By simultaneously performing speech recognition processing using a plurality of sets of acoustic models, a plurality of speech recognition results and a plurality of recognition accuracy corresponding to the plurality of sets of acoustic models can be obtained. Furthermore, the speech recognition result that provides the highest recognition accuracy among the plurality of recognition accuracy is output as the final speech recognition result. As a result, a speech recognition result with the highest recognition accuracy can be obtained in speech recognition processing performed simultaneously using a plurality of sets of acoustic models set according to a plurality of different conditions. The rate can be improved.

  If the recognition accuracy of a predetermined set of acoustic models is often higher than the recognition accuracy of another set of acoustic models during repeated speech recognition processing, the speech is obtained using only the predetermined set of acoustic models. When the recognition processing is performed and the recognition accuracy in the predetermined set of acoustic models is equal to or lower than a preset threshold value, the recognition processing may be changed to the voice recognition processing using a plurality of sets of acoustic models again.

  In the above embodiment, a specific example in which the present invention is applied to an air conditioner has been described. However, the present invention is not limited to this embodiment, and can be applied to general electronic devices. Many modifications are possible within the technical idea of the present invention by those having ordinary knowledge in the field.

  DESCRIPTION OF SYMBOLS 100 ... Voice recognition apparatus, 10 ... Command analysis part, 20 ... Voice input part, 30 ... A / D converter, 40 ... Voice recognition process part, 41 ... Signal processing part, 42 ... Acoustic model selection part, 43, 43a- 43d ... coincidence detection unit, 44 ... voice quality determination unit, 45 ... acoustic model training unit, 46 ... voice quality determination information extraction unit, 47 ... recognition accuracy determination unit, 50, 50a, 50b ... memory, 51 ... acoustic model storage unit, 52 ... Option list storage unit, 53 ... Voice quality determination information storage unit, 60 ... Audio reproduction unit, 70 ... D / A converter, 80 ... Audio output unit, 200 ... Control unit, 91 ... Host CPU, 92 ... Storage unit

Claims (11)

An acoustic model storage unit that stores a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions;
An option list storage unit for storing an option list including option data representing options and acoustic model specifying information for specifying one set of acoustic models from the plurality of sets of acoustic models;
An acoustic model selection unit that selects the set of acoustic models identified by the acoustic model identification information;
A signal processing unit that extracts a frequency component of the input audio signal and generates a feature pattern representing a distribution state of the frequency component of the audio signal;
The feature pattern generated from at least a part of the speech signal is compared with an acoustic model corresponding to at least a part of the option data in the selected set of acoustic models, and speech recognition processing is performed. A coincidence detection unit that outputs a speech recognition result;
A speech recognition apparatus comprising:   The plurality of sets of acoustic models are used in the predetermined language and a set of acoustic models obtained by collecting the distribution states of the frequency components of phonemes included in unspecified types of terms used in the predetermined language. The speech recognition apparatus according to claim 1, further comprising: a set of acoustic models obtained by collecting distribution states of frequency components of phonemes included in a specific type of term. An acoustic model storage unit for storing a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities;
An option list storage unit that stores an option list including option data representing options;
A signal processing unit that extracts a frequency component of the input audio signal and generates a feature pattern representing a distribution state of the frequency component of the audio signal;
A voice quality determination unit that determines the type of voice quality in the audio signal;
An acoustic model selection unit that selects one set of acoustic models corresponding to the type of voice quality determined by the voice quality determination unit from the plurality of sets of acoustic models;
The feature pattern generated from at least a part of the speech signal is compared with an acoustic model corresponding to at least a part of the option data in the selected set of acoustic models, and speech recognition processing is performed. A coincidence detection unit that outputs a speech recognition result;
A speech recognition apparatus comprising: The option list storage unit stores an option list further including control information for permitting or prohibiting voice recognition processing for the specific voice quality type,
The coincidence detection unit determines whether to start voice recognition processing for the option list based on the voice quality type determined by the voice quality determination unit and the control information.
The speech recognition apparatus according to claim 3. An acoustic model storage unit that stores a plurality of sets of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions;
An option list storage unit that stores an option list including option data representing options;
A signal processing unit that extracts a frequency component of the input audio signal and generates a feature pattern representing a distribution state of the frequency component of the audio signal;
The feature pattern generated from at least a part of the speech signal is compared with an acoustic model corresponding to at least a part of the option data in the plurality of sets of acoustic models, and a speech recognition process is performed. A coincidence detector that outputs a plurality of speech recognition results and a plurality of recognition accuracy corresponding to a set of acoustic models;
A recognition accuracy determination unit that outputs a speech recognition result having the highest recognition accuracy among the plurality of recognition accuracy as a final speech recognition result;
A speech recognition apparatus comprising:   The plurality of sets of acoustic models are used in the predetermined language and a set of acoustic models obtained by collecting the distribution states of the frequency components of phonemes included in unspecified types of terms used in the predetermined language. 6. A speech recognition apparatus according to claim 5, further comprising: a set of acoustic models obtained by collecting distribution states of frequency components of phonemes included in a specific type of term.   The plurality of sets of acoustic models are composed of a plurality of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in the predetermined language for a plurality of different types of voice qualities. The speech recognition apparatus according to claim 5.   The electronic device which comprises the speech recognition apparatus of any one of Claims 1-7. A step (a) of reading out the option data and the acoustic model specifying information from an option list storage unit that stores an option list including option data representing options and acoustic model specifying information for specifying a set of acoustic models;
A set of acoustic models specified by the acoustic model identification information from a plurality of acoustic models obtained by collecting distribution states of frequency components of a plurality of phonemes used in a predetermined language under a plurality of different conditions Selecting an acoustic model (b);
Extracting a frequency component of the input voice signal and generating a feature pattern representing a distribution state of the frequency component of the voice signal;
An acoustic model corresponding to at least a part of the option data read out in step (a), among the selected set of acoustic models, the feature pattern generated from at least a part of the audio signal. A step (d) of performing speech recognition processing and outputting a speech recognition result in comparison with
A speech recognition method comprising: A step (a) of reading out the option data from an option list storage unit that stores an option list including option data representing options;
(B) extracting a frequency component of the input audio signal and generating a feature pattern representing a distribution state of the frequency component of the audio signal;
Determining the type of voice quality in the audio signal (c);
From the plurality of sets of acoustic models obtained by collecting the distribution states of frequency components of a plurality of phonemes used in a predetermined language for a plurality of different types of voice qualities, the types of voice qualities determined in step (c) Selecting a corresponding set of acoustic models (d);
An acoustic model corresponding to at least a part of the option data read out in step (a), among the selected set of acoustic models, the feature pattern generated from at least a part of the audio signal. A step (e) of performing speech recognition processing and outputting a speech recognition result in comparison with
A speech recognition method comprising: A step (a) of reading out the option data from an option list storage unit that stores an option list including option data representing options;
(B) extracting a frequency component of the input audio signal and generating a feature pattern representing a distribution state of the frequency component of the audio signal;
A plurality of sets of acoustic models obtained by collecting the feature patterns generated from at least a part of the speech signal under a plurality of different conditions of frequency component distribution states of a plurality of phonemes used in a predetermined language And performing speech recognition processing in comparison with an acoustic model corresponding to at least a part of the option data read in step (a), and a plurality of speech recognition results corresponding to the plurality of sets of acoustic models; Outputting a plurality of recognition accuracy (c);
A step (d) of outputting a speech recognition result having the highest recognition accuracy among the plurality of recognition accuracy as a final speech recognition result;
A speech recognition method comprising:

Images