JP6320962B2 - Speech recognition system, speech recognition method, program - Google Patents

Description

  The present invention relates to a voice recognition system, a voice recognition method, and a program including a client device, a plurality of voice recognition server devices, and a management unit.

  For example, Patent Literature 1 discloses a conventional server / client type speech recognition method. In the server / client type speech recognition method of Patent Document 1, a noise model is generated based on a signal of a non-voice section observed by a client device, and noise common to the client device and the speech recognition server device is determined by the noise model and the speech model. A superimposed speech model is generated, and the client device encodes the feature quantity of the input speech to be recognized based on the noise superimposed speech model and transmits it to the speech recognition server device. The speech recognition server device uses the code received from the client device. Based on the noise superimposed speech model, it is converted into a feature value. Since this method generates a noise-superimposed speech model based on a noise model each time, there is an advantage that it can cope with speech recognition processing under various noises.

  Further, in the speech recognition method of Patent Document 2, the similarity (referred to as noise similarity) between a signal in a noise interval (interval that is not a speech interval) at the time of speech input and a noise signal superimposed when creating a noise-superimposed speech model. Calculated and used as a noise-superimposed speech model whose similarity is a predetermined value or more as a probability model for speech recognition, or stored in advance if there is no noise-superimposed speech model whose similarity is a predetermined value or more A noise-superimposed speech model adapted to the noise signal is created using the model and the signal in the noise interval to obtain a stochastic model for speech recognition. This method has an advantage in that it is not necessary to request a high-level processing from the client device.

Japanese Patent No. 4769121 Japanese Patent No. 4242320

  In the method of Patent Document 1, if it takes time to generate a noise-superimposed speech model and use it for recognition after observing the noise model, the noise characteristics during noise model observation and the noise characteristics during recognition use The characteristics may be different and the speech recognition performance may be affected. In order to generate a noise-superimposed speech model at high speed for use in speech recognition, a high processing capability is required for the client device. Similarly, the voice recognition server apparatus also has a problem that the amount of operation temporarily increases in order to create a noise superimposed voice model.

  If the method of Patent Document 2 is applied to a server / client type speech recognition system that is used by a large number of users at the same time, one speech recognition server device can be used to deal with signals in the noise interval of all users. It is necessary to store a noise-superimposed speech model corresponding to various types of noise, and the model management becomes complicated. Alternatively, every time a signal in the noise interval is received from the client device, there is a problem that the amount of operation of the speech recognition server device temporarily increases in order to create a noise superimposed speech model.

  Therefore, an object of the present invention is to provide a voice recognition system that does not require advanced processing from a client device and can realize high performance at a low introduction cost.

  The speech recognition system of the present invention includes a client device, a plurality of speech recognition server devices, and a management unit.

  The client device includes a transmission unit. The transmission unit transmits the input acoustic signal or a signal derived from the acoustic signal to the voice recognition server device selected based on the sound collection condition. Each of the voice recognition server devices includes a setting storage unit and a utilization rate transmission unit. The setting storage unit stores in advance settings relating to voice recognition. The usage rate transmission unit transmits information on the usage rate, which is a rate at which the client device uses the own device as a transmission destination, to the management unit. The management unit transmits the threshold value of the sound collection condition updated based on the usage rate to the client device, and updates based on the usage rate for the voice recognition server device specified based on the usage rate. At least one of the operations for transmitting information on the set setting is executed.

  According to the voice recognition system of the present invention, high performance can be realized at a low introduction cost without requiring high-level processing from the client device.

1 is a block diagram illustrating a configuration of a voice recognition system according to Embodiment 1. FIG. 1 is a block diagram illustrating a configuration of a voice recognition server device according to a first embodiment. FIG. 3 is a block diagram illustrating a configuration of a management unit according to the first embodiment. FIG. 3 is a sequence diagram illustrating a voice recognition operation of the voice recognition system according to the first embodiment. FIG. 3 is a sequence diagram illustrating an information update operation of the voice recognition system according to the first embodiment. The figure which shows the ratio (utilization rate) of the use track record before the setting update in an application example. The figure which shows the ratio (utilization rate) of the use track record after the setting update in an application example. FIG. 3 is a block diagram illustrating a configuration of a voice recognition system according to a second embodiment. The block diagram which shows the structure of the speech recognition server apparatus of Example 2. FIG. FIG. 9 is a block diagram illustrating a configuration of a management unit according to the second embodiment. FIG. 9 is a sequence diagram illustrating an information update operation of the voice recognition system according to the second embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

  In the following description, a signal uttered as a speech recognition target is referred to as a sound signal, and a signal collected in a state where a sound signal and a background noise signal other than the sound signal are mixed is referred to as an acoustic signal.

  Hereinafter, an outline of the voice recognition system according to the first embodiment will be described. As described above, if the acoustic signal input to the client device is distributed and transmitted to a plurality of voice recognition server devices for each sound collection condition, and the voice recognition processing is executed in each voice recognition server device, the client device Therefore, it is possible to achieve high performance at a low introduction cost without requiring advanced processing. However, under what kind of sound collection conditions the client device uses this system is unknown at the planning stage of this system, so there is a possibility of traffic concentration and depopulation to a specific voice recognition server device. There is. In other words, if the frequency of use of the system under a certain sound collection condition is remarkably increased, the load is concentrated on the voice recognition server device selected based on the sound collection condition. Further, if the frequency of use of the system under a certain sound pickup condition is extremely low, the voice recognition server device selected based on the sound pickup condition is hardly used. If processing delay occurs due to local load concentration, it is disadvantageous for the user. Further, if there is a voice recognition server device that is hardly used, equipment is wasted, which is disadvantageous for the system operator. Or, since the usage tendency of the users of the system fluctuates after the start of operation of the system, and the above problem newly occurs, the system cannot follow the fluctuations. The disadvantage of the system operator.

  Therefore, in this embodiment, the load on a specific voice recognition server device is monitored by monitoring the usage rate of the voice recognition server device and changing the setting of the voice recognition server device or the client device according to the usage rate. Disclosed is a speech recognition system that can be operated so as not to concentrate and can improve the performance (performance) of the entire speech recognition system.

  Hereinafter, the configuration of the speech recognition system according to the present embodiment will be described with reference to FIGS. 1, 2, and 3. FIG. 1 is a block diagram showing the configuration of the speech recognition system 1 of the present embodiment. FIG. 2 is a block diagram illustrating the configuration of the voice recognition server device 21-n according to the present embodiment. FIG. 3 is a block diagram illustrating the configuration of the management unit 30 according to the present embodiment.

  As shown in FIG. 1, the speech recognition system 1 of the present embodiment includes a client device 10 and a plurality of speech recognition server devices 21-1, ..., 21-n, ..., 21-N (N is N ≧ 2). An integer to be satisfied, n is an integer satisfying 1 ≦ n ≦ N), and the management unit 30. Although only one client device 10 is shown in FIG. 1, it is assumed that there are a plurality of client devices 10. When the voice recognition server devices 21-1,..., 21-n,..., 21-N are collectively called, they are referred to as a voice recognition server device group 20. It is assumed that the client device 10 and the voice recognition server device group 20 are connected to be communicable wirelessly or by wire via a network. The management unit 30 may be configured as a single hardware (device). When the management unit 30 is configured as a single piece of hardware (device), it may be called the management device 30. When the management unit 30 is configured as a single piece of hardware (device), the client device 10, the voice recognition server device group 20, and the management unit 30 (management device 30) are connected to be communicable wirelessly or wired via a network. It shall be. Further, the management unit 30 may be a configuration requirement in the client device 10 or a configuration requirement in any voice recognition server device in the voice recognition server device group 20.

  As illustrated in FIG. 1, the client device 10 includes a sound collection condition extraction unit 11, a threshold storage unit 111, a selection unit 12, a transmission destination storage unit 121, a signal processing unit 13, and a transmission unit 14. The receiving unit 15, the presenting unit 16, and the transmission destination changing unit 17 are included. As shown in FIG. 2, all the speech recognition server devices (represented by 21-n) included in the speech recognition server device group 20 include an acoustic signal receiving unit 21A, a speech recognizing unit 21B, and a recognition result transmitting unit. 21C, a utilization rate transmission unit 21D, a setting information reception unit 21E, a setting update unit 21F, and a setting storage unit 21G. As illustrated in FIG. 3, the management unit 30 (management device 30) includes a usage rate reception unit 30A, a setting information update unit 30B, a setting information transmission unit 30C, a transmission destination information update unit 30D, and transmission destination information transmission. A unit 30E, a destination storage unit 30G, and a setting storage unit 30F.

  Hereinafter, the speech recognition operation of this system will be described with reference to FIG. FIG. 4 is a sequence diagram showing the voice recognition operation of the voice recognition system 1 of the present embodiment. First, the sound collection condition extraction unit 11 extracts a sound collection condition of the input acoustic signal (S11). The selection unit 12 selects a voice recognition server device (for example, the voice recognition server device 21-1) as a transmission destination of the corresponding acoustic signal based on the extracted sound collection condition (S12). It is assumed that the relationship between the sound collection condition and the voice recognition server device as the transmission destination is stored in advance in the transmission destination storage unit 121.

<Sound collection conditions>
The sound collection conditions include, for example, a feature amount relating to the S / N ratio, which is a ratio of the size of the audio signal and the background noise signal, a feature amount relating to the distortion of the acoustic signal, a feature amount relating to the spectrum shape of the background noise signal, and a background noise signal. It is possible to set a condition based on a threshold value for at least one of the feature amounts related to the size of the feature amount. It is assumed that the threshold value is stored in advance in the threshold value storage unit 111.

  The background noise signal is a signal observed by a microphone for a certain time immediately before the voiced sound or the target sound is input. The magnitude of the background noise signal is an average value of the power spectrum of the background noise signal over a certain period of time. The spectrum shape of the background noise signal is a component of each band in the spectrum of the background noise signal and its time change. The S / N ratio between the audio signal and the background noise signal is the ratio of the size of the audio signal in the acoustic signal being input to the uttered speech (target sound) and the size of the background noise signal. As the audio signal, a power spectrum obtained by subtracting the average value of the power spectrum of the background noise signal for a certain time from the power spectrum of the acoustic signal for a certain time during the input of the uttered voice (target sound) can be used. The magnitude of the audio signal is an average value of the power spectrum of the audio signal for a certain time during the input of the uttered voice (target sound).

  The distortion of the acoustic signal means that the microphone element, the microphone amplifier circuit, and the A / D conversion are clipping because the input of the acoustic signal is too large. A section where the input signal level has an amplitude greater than or equal to a predetermined threshold is detected, and the time ratio is calculated. If this ratio is high, the strain is large, and if the ratio is small, the strain is small. If the amplitude does not exceed the threshold, no distortion can be achieved. The threshold is set according to the clipping level of the microphone element, circuit, and AD conversion.

<Sound Collection Condition Extraction Unit 11 (S11), Selection Unit 12 (S12)>
Hereinafter, an example of the operation (S11, S12) of the sound collection condition extraction unit 11 and the selection unit 12 will be described. The sound collection condition extraction unit 11 extracts, for example, a feature amount representing the sound collection condition from the input sound signal, and the sound signal input according to the value of the feature amount is grouped (for example, a code representing the sound collection condition). Divide.

  Next, as shown in Table 1, the selection unit 12 transmits a corresponding acoustic signal based on the relationship between a group (a code representing a sound collection condition) and an index (a code representing a destination speech recognition server device). A previous voice recognition server device (for example, the voice recognition server device 21-1) is selected (S12).

  The feature amount x is, for example, the S / N ratio that is the ratio of the size of the audio signal included in the acoustic signal and the size of the background noise signal, the presence or absence of distortion of the acoustic signal, the frequency of distortion, the spectrum shape of the background noise signal, the background The magnitude of the noise signal, etc.

When the feature quantity x is an S / N ratio, for example, threshold values are set as θ ₁ = 0 dB, θ ₂ = 10 dB, θ ₃ = 20 dB, and so on. Are extracted as sound collection conditions, and the selection unit 12 selects index 2.

When the feature quantity x is a distortion of an acoustic signal, for example, the ratio of the time when the absolute value of the amplitude is 30000 or more in 0.5 seconds is a feature quantity x in a signal quantized with a bit depth of 16 bits. The threshold is set as θ ₁ = 0.8, etc., and if x = 0, the sound collection condition extraction unit 11 extracts group 1 as the sound collection condition, the selection unit 12 selects index 1, and if x = 0.9 The sound collection condition extraction unit 11 extracts group 2 as the sound collection condition, and the selection unit 12 selects index 2.

When the feature amount x is the spectrum shape of the background noise signal, for example, the size of the background noise signal is x ₁ , x ₂ ,..., X _m (m is an integer satisfying m ≧ 2) depending on the frequency band and duration. Separately evaluate. The sound collection condition extraction unit 11 extracts a group from the combination of evaluation results, and the selection unit 12 selects the index. As another method of using the spectrum shape of the background noise signal as a feature amount, it is also possible to store a plurality of types of background noise signal models and classify the input background noise signal into one of the models. it can. The multiple types of background noise signals are, for example, white noise, pink noise, burst noise, and the like. In this method, a group corresponding to each model is allocated, and a group is determined according to a model in which background noise signals of input signals are classified.

When the feature amount x is set to the size of the background noise signal, for example, θ ₁ = 40 dBA, θ ₂ = 55 dBA, θ ₃ = 70 dBA, etc. are set, and if the feature amount x = 50 dBA, the sound collection condition extraction unit 11 sets the group 2 Extracting is performed as a sound collection condition, and the selection unit 12 selects index 2. Here, dBA is a unit of dB value of noise level measured under frequency weighting characteristics (A characteristics) considering human hearing.

<Signal processing unit 13 (S13)>
The signal processing unit 13 performs signal processing on the corresponding acoustic signal when the extracted sound collection condition corresponds to a predetermined condition (S13). Specifically, the signal processing unit 13 is a speech recognition target in a speech recognition server device in which the S / N ratio and the size of the background noise signal are determined based on the sound collection conditions extracted by the sound collection condition extraction unit 11. The corresponding acoustic signal is subjected to signal processing so as to conform to the range of the feature amount assumed as. For example, in the sound pickup condition in the vicinity of S / N ratio = 1, that is, in the vicinity of 0 dB, the size of the audio signal and the size of the background noise signal are the same. Easy to invite. Accordingly, when the sound collection condition extraction unit 11 extracts a sound collection condition in the vicinity of S / N ratio = 0 dB, the signal processing unit 13 performs signal processing for suppressing the background noise signal with respect to the sound signal of the sound collection condition. Apply. Alternatively, for example, when the sound collection condition extraction unit 11 extracts a sound collection condition in the vicinity of S / N ratio = 100, that is, in the vicinity of 20 dB, according to the value of the S / N ratio, similar to the sound collection condition in the vicinity of 0 dB described above. Thus, the process of adaptively suppressing the background noise signal may be performed, or the process of suppressing may not be performed at all. Even in other sound collection conditions, the signal processing unit 13 adaptively performs signal processing on the acoustic signal based on the result extracted by the sound collection condition extraction unit 11.

Hereinafter, an example of the operation (S13) of the signal processing unit 13 will be described. In speech recognition, in many cases, input speech is corrected by signal processing as preprocessing. Examples of acoustic characteristics to be dealt with by preprocessing in speech recognition include additive noise and multiplicative noise. Additive noise is a signal that is additively observed with respect to a speech signal, such as noise ubiquitous in the speech input environment. On the other hand, multiplicative noise is noise (distortion) caused by acoustic characteristics such as microphone characteristics and spatial transfer characteristics, and is observed as a convolution operation with respect to the original speech waveform in the time waveform. It becomes a distortion. Examples of speech recognition processing that addresses additive noise include the noise suppression method based on the spectral subtraction method disclosed in paragraph [0005] of Reference Patent Document 1, or the Wiener method disclosed in paragraph [0007] of that document. There is a method of suppressing noise from speech with superimposed noise and applying it to speech recognition, such as a noise suppression method based on a filter method (hereinafter referred to as WF method).
(Reference Patent Document 1: Japanese Patent No. 4464797)

As an example of speech recognition processing that copes with multiplicative noise in addition to additive noise, a noise superimposed speech model is generated by superimposing a noise model on a speech model from which the influence of multiplicative noise has been removed as in Patent Document 1. Then, there is a method of updating the model based on the multiplicative feature amount. Alternatively, there is a method of generating a normalized noise superimposed speech model after normalizing a noise model based on the multiplicative noise feature quantity as in the invention of Reference Patent Document 2.
(Reference Patent Document 2: Japanese Patent No. 5200080)

  Noise suppression is typically considered as signal processing performed by the signal processing unit 13. As signal processing other than noise suppression, for example, AGC (Automatic Gain Control), CMN (Cepstrum Mean Normalization), and an equalizer may be used.

<AGC>
Automatic Gain Control (AGC) detects the input signal level based on the short-time average power or short-time average amplitude of the input audio signal so that the difference between the input signal level and the optimum level (target value) is reduced. This is a process for adjusting the gain of the audio input stage. AGC has an effect of preventing the voice feature amount from becoming unclear due to the voice waveform after A / D conversion being too small or excessive. AGC is disclosed in paragraph [0001] of Reference Patent Document 3, for example.
(Reference Patent Document 3: Japanese Patent No. 3588555)

<CMN>
Cepstrum Mean Normalization (CMN) is a process of obtaining a long-term cepstrum average of an input speech signal in a cepstrum that is a feature amount of speech recognition, and subtracting the long-term cepstrum average from the cepstrum of the input speech of each frame. The CMN is used to reduce the influence of multiplicative distortion represented by the characteristics of the microphone, the position of the microphone, and the shape of the room. CMN is disclosed in paragraph [0010] of Reference Patent Document 1, for example.

  When the CMN is performed by the signal processing unit 13 of the client device 10, the client device 10 transmits to the speech recognition server device an MFCC (Mel Frequency Cepstrum after CMN application) as a signal derived from an acoustic signal for speech recognition. ) Is transmitted, it is possible to omit the cepstrum analysis process again by the speech recognition server device.

<Equalizer>
The equalizer is a process for adjusting the gain of the input audio signal for each frequency band. For example, if it is known in advance that the acoustic characteristics of a microphone for voice input are not flat, sound can be collected after improving the acoustic characteristics via an equalizer. The equalizer is disclosed in, for example, paragraphs [0010] and [0016] of Reference Patent Document 4.
(Reference Patent Document 4: Japanese Patent No. 2865268)

  Next, the transmission unit 14 transmits an acoustic signal or a signal derived from the acoustic signal to the speech recognition server device (speech recognition server device selected in step S12) corresponding to the extracted sound collection condition (S14). . At this time, the transmission unit 14 varies the transmission destination between the case where the signal processing of step S13 is not performed and the case where the signal processing is performed, and the acoustic signal which is not subjected to signal processing or the acoustic signal which is subjected to signal processing. A signal shall be transmitted. Moreover, you may determine the transmission destination in any one of different speech recognition server apparatuses only by whether the signal processing of step S13 was implemented irrespective of the speech recognition server apparatus selected by step S12. Note that the signal derived from the acoustic signal refers to a signal representing a feature amount of the acoustic signal, an acoustic signal subjected to the signal processing in step S13, and the like. In addition, when the transmission unit 14 transmits an acoustic signal or a signal derived from the acoustic signal, the sound recognition condition (group), the threshold value thereof, and information on the presence or absence of signal processing in the signal processing unit 13 are transmitted to the voice recognition server device You may send it. The voice recognition server device can record the sound pickup condition or signal processing condition for selecting the voice recognition server device from the sound pickup condition (group), its threshold value, and the presence or absence of signal processing. become.

  The acoustic signal receiving units 21A of the voice recognition server devices 21-1, ..., 21-n, ..., 21-N receive an acoustic signal or a signal derived from the acoustic signal from the client device 10 (S21A). The speech recognition unit 21B of the speech recognition server device (for example, the speech recognition server device 21-1) that has received the acoustic signal or the signal derived from the acoustic signal executes speech recognition processing (S21B).

<Voice recognition processing (S21B)>
The voice recognition process in step S21B is executed as follows, for example. The voice recognition unit 21B converts one sentence or one word utterance into a character string. The speech recognition unit 21B uses speech power and its variation, MFCC (Mel-Frequency Cepstrum Coefficient) and its dynamic variation as speech features. The speech recognition unit 21B searches for a word string using a statistical acoustic model or a language model.

  After executing the speech recognition process in step S21B, the recognition result transmitting unit 21C transmits the speech recognition result to the client device 10 (S21C). The receiving unit 15 of the client device 10 receives the voice recognition result (S15A). The presentation unit 16 of the client device 10 presents the received voice recognition result (S16).

  Hereinafter, the information update operation of the speech recognition system 1 of the present embodiment will be described with reference to FIG. FIG. 5 is a sequence diagram showing an information update operation of the speech recognition system 1 of the present embodiment. First, the utilization rate transmission unit 21D of all the voice recognition server devices transmits information on the utilization rate to the management unit 30 (S21D). For example, the usage rate of a certain voice recognition server device is obtained by accumulating the number of times the client device 10 has used the voice recognition server device as a transmission destination, the amount of data transmission, etc. for all the client devices 10 using the voice recognition server device. It is defined as the ratio divided by a predetermined time. Step S <b> 21 </ b> D may be periodically executed at a preset time, or may be executed every elapse of a preset predetermined time. The utilization rate transmitting unit 21D may send information on the utilization rate instead of the utilization rate itself. The information on the utilization rate is, for example, a data reception amount (transmission amount) per unit time for a sound signal input from each client device or data of a signal derived from the sound signal, or a recognition process performed by the voice recognition server device. Such as CPU time. The data reception amount (transmission amount) and CPU time per unit time are not information that represents the utilization rate by itself, but the management unit 30 also receives the same reception amount (transmission amount) and CPU from other voice recognition server devices. The utilization rate per unit time can be obtained by acquiring the time, and the management unit 30 summing up these reception amounts (transmission amounts) and CPU time. Accordingly, the amount of data received per unit time (transmission amount), CPU time, and the like are classified as information on the utilization rate. The utilization rate receiving unit 30A of the management unit 30 receives information on the utilization rate from the voice recognition server device group 20 (S30A). The setting information updating unit 30B of the management unit 30 updates information (setting information) related to the setting of the voice recognition server device (represented by 21-n in this case) specified based on the usage rate obtained from the usage rate information. (30B). It is assumed that the setting information before update is stored in the setting storage unit 30F of the management unit 30, and the setting information update unit 30B stores or overwrites the updated setting information in the setting storage unit 30F.

<Setting>
Here, the setting is information stored in the setting storage unit 21G of each voice recognition server device, and an acoustic model used for voice recognition, a language model, the acoustic model, and an operation related to a recognition operation using the language model. This refers to settings, other parameters used for voice recognition, and other settings determined in advance for voice recognition. It is assumed that each setting is adjusted to have high recognition performance specialized for a certain sound pickup condition, and each voice recognition server device has a different setting from each other or shares the same setting in several units. It shall be. The setting storage unit 21G can store a plurality of settings. In this case, it is assumed that each voice recognition server device has previously selected one of the settings stored in the setting storage unit 21G as an active setting.

  The setting information update unit 30B described above may update the setting information for the voice recognition server device having a low usage rate so as to be the same as the setting of the voice recognition server device having a high usage rate. This increases the number of voice recognition server devices that share settings with the voice recognition server device on which the load is concentrated, thus reducing the load concentration. Further, the setting information update unit 30B described above may update the setting information for the voice recognition server device having a high usage rate so as to be the same as the setting of the voice recognition server device having a low usage rate. Thereby, the traffic with respect to the applicable speech recognition server apparatus can be reduced temporarily. However, in this case, since the factor that concentrates the load is not removed, a measure to alleviate the concentration of the load is required by updating the setting information of other voice recognition server devices.

  Next, the setting information transmission unit 30C of the management unit 30 transmits the setting information described above to a voice recognition server device (one or more or more than one) having a high usage rate or a low usage rate (S30C). . For example, when the setting information is updated for the voice recognition server device having a low usage rate in step S30B, the setting information transmitting unit 30C transmits the setting information to the voice recognition server device having the low usage rate (S30C ). On the other hand, when the setting information is updated for the voice recognition server device having a high usage rate in step S30B, the setting information transmitting unit 30C transmits the setting information to the voice recognition server device having the high usage rate ( S30C).

  The setting information transmission unit 30C may transmit the setting information described above to any one (or both) of the voice recognition server devices corresponding to the highest usage rate and the lowest usage rate (S30C). Typically, the setting information transmitting unit 30C updates the setting information updated so that the setting is shared with the voice recognition server device having the highest usage rate with respect to the voice recognition server device having the lowest usage rate in step S30B. Can be considered. Alternatively, the setting information transmission unit 30C is set to the voice recognition server device having the lowest usage rate or the average usage rate with respect to the voice recognition server device having the highest usage rate and concentrated load in step S30B. It is conceivable to transmit the setting information updated so as to be shared. In this case, the purpose is to avoid a temporary increase in traffic to the voice recognition server device.

  The setting information receiving unit 21E of the voice recognition server device 21-n receives setting information from the management unit 30 (S21E). As described above, when a plurality of settings related to voice recognition are stored in advance in the setting storage unit 21G of the voice recognition server device 21-n, the setting update unit 21F stores the plurality of settings stored based on the received setting information. The setting is updated by selecting one of the settings (as an active setting) (S21F).

  Next, when the setting information transmission unit 30C transmits the setting information, the transmission destination information update unit 30D of the management unit 30 updates the transmission destination information along with this (S30D). As described above, the transmission destination information is information related to the relationship between the sound collection condition and the voice recognition server device that is the transmission destination. The transmission destination information may be referred to as information that links the sound collection condition and the transmission destination speech recognition server apparatus. It is assumed that the transmission destination information before update is stored in the transmission destination information storage unit 30G of the management unit 30, and the transmission destination information update unit 30D stores the updated transmission destination information in the transmission destination information storage unit 30G. Or memorize by overwriting. The transmission destination information transmission unit 30E of the management unit 30 transmits the updated transmission destination information to the client device 10 (S30E).

  The receiving unit 15 of the client device 10 receives the transmission destination information from the management unit 30 (S15B). In the client device 10, transmission destination information before update is stored in the transmission destination storage unit 121. The transmission destination changing unit 17 changes the stored transmission destination information based on the transmission destination information received from the management unit 30 (S17). The selection unit 12 selects a voice recognition server device that is a transmission destination of the corresponding acoustic signal based on the sound pickup condition of the input acoustic signal and the transmission destination information changed in step S17 (S12).

<Application example of this system>
Hereinafter, application examples of this system will be described. First, a threshold for extracting a sound collection condition is determined by prior learning. Similarly, the parameters of the acoustic model of the speech recognition server device group 20 are determined. Specifically, the server administrator learns a plurality of recognition operation settings including an acoustic model in advance, and stores the learning result in the speech recognition server device group 20. In this example, it is assumed that ten speech recognition server devices are prepared, and five types of recognition operation settings (settings A, B, C) corresponding to the sound pickup conditions are set for two of the ten speech recognition server devices. , D, E)). The two speech recognition server devices in which the setting A is stored are referred to as speech recognition server devices 21-1 and 21-2. Similarly, the speech recognition server devices that hold the settings B, C, D, and E for each of the two units are designated as speech recognition server devices 21-3 and 21-4, 21-5 and 21-6, 21-7 and 21-. 8, 21-9 and 21-10. When these voice recognition server devices are collectively called, they are called the voice recognition server device group 20 as described above.

  The voice recognition server device group 20 holds operation settings specialized for a voice input of a specific voice level, noise level, and noise steadiness. For example, the speech recognition server devices 21-1 and 21-2 hold the acoustic model A created using speech mixed at a low noise level and operation settings (setting A) suitable for the acoustic model A. The speech recognition server devices 21-3 and 21-4 hold the acoustic model B created using speech mixed with a medium level of noise and operation settings (setting B) suitable for the acoustic model B. The speech recognition server devices 21-5 and 21-6 hold the acoustic model C created using speech mixed with non-stationary noise and operation settings (setting C) suitable for the acoustic model C. The speech recognition server devices 21-7 and 21-8 hold the acoustic model D created using speech mixed with a high level of noise and operation settings suitable for the acoustic model D (setting D). The speech recognition server devices 21-9 and 21-10 hold an acoustic model E created using speech with a low signal level and operation settings (setting E) suitable for the acoustic model E. The table below shows the relationship between the voice recognition server device and the settings held in these servers.

  The sound collection condition extraction unit 11 of the client device 10 extracts the sound collection condition of the acoustic signal from the voice level, noise level, noise continuity, and the like calculated from the input acoustic signal (S11). The selection unit 12 selects a speech recognition server device that holds the optimal speech recognition operation setting based on the extracted sound collection conditions (S12). The signal processing unit 13 suppresses a noise component mixed in the acoustic signal according to the extracted sound collection condition (S13). The transmission unit 14 transmits the acoustic signal in which noise is suppressed to the voice recognition server device selected in step S12 (S14).

  When the acoustic signal receiving unit 21A of the voice recognition server device group 20 receives the acoustic signal from the client device 10 (S21A), for example, a communication log as shown in the table below is stored in a predetermined storage area of the own device.

  The voice recognition unit 21B executes voice recognition processing according to any active setting (any one of settings A, B, C, D, and E) held in the setting storage unit 21G (S21B). The recognition result transmission unit 21C transmits the voice recognition result to the client device 10 (S21C).

  The utilization rate transmission unit 21D of the voice recognition server device group 20 transmits the communication log to the management unit 30, for example, once a week. The management unit 30 totals the data sizes of the acoustic signals received from the voice recognition server device group 20. For example, the received data size is 10 GB for the entire speech recognition server device group 20, 5 GB for the speech recognition server devices 21-1 and 21-2, 2 GB for the speech recognition server devices 21-3 and 21-4, and the speech recognition server device 21. -5, 21-6 is 1.5 GB, the speech recognition server device 21-7, 21-8 is 1 GB, and the speech recognition server device 21-9, 21-10 is 0.5 GB. FIG. 6 is a diagram illustrating a ratio (utilization rate) of a usage record before setting update in the application example.

The number N _A of speech recognition server apparatuses that retain the setting A after the setting update can be calculated by, for example, Expression (1).
N _A = └N _all・ D _A / D _all + 0.5┘ (1)
N _all is the total number of voice recognition server devices, D _all is the total data size received by all the voice recognition server devices, D _A is the data size received by the voice recognition server devices 21-1 and 21-2 holding the setting A Is the sum of └ ┘ is a floor function, and └x + 0.5┘ means rounding x. The numbers N _B , N _C , and N _D of the speech recognition server apparatuses that retain the settings B, C, and D after the setting update are obtained by rewriting the subscript A in the formula (1) to B, C, and D. The number N _E of speech recognition server apparatuses that retain the setting E after the setting update can be calculated by, for example, Expression (2).
N _E = N _all- (N _A + N _B + N _C + N _D ) (2)

In the example of FIG. 6, N _A = 5, N _B = 2, N _C = 2, N _D = 1 and N _E = 0, and the usage rate of the voice recognition server device after the setting update is expressed as shown in FIG. The The setting information update unit 30B of the management unit 30 uses the setting information of the voice recognition server devices 21-8, 21-9, and 21-10 with low usage rates as the voice recognition server devices 21-1 and 21-2 with high usage rates. (S30B). The setting information transmission unit 30C of the management unit 30 transmits the setting information described above to the voice recognition server devices 21-8, 21-9, and 21-10 having a low usage rate (S30C).

  The setting information receiving unit 21E of the voice recognition server devices 21-8, 21-9, and 21-10 receives the setting information from the management unit 30 (S21E). As described above, when a plurality of settings relating to voice recognition are stored in advance in the setting storage unit 21G of the voice recognition server devices 21-8, 21-9, and 21-10, the setting update unit 21F receives the received setting information. The setting is updated by selecting one of the plurality of settings stored based on the setting (setting A in this example) as an active setting (S21F).

  The transmission destination information transmission unit 30E of the management unit 30 transmits the transmission destination information updated with the above setting update to the client device 10 (S30E). When there are two or more voice recognition server devices with the same setting as in the above application example, any IP address of the voice recognition server devices having the same setting as the transmission destination by the terminal ID of the client device 10 May be changed to be a transmission destination, and the same client device 10 may transmit data to the same voice recognition server device. In addition, since there may be 0 voice recognition server devices that retain the setting after the update as in the setting E in the above example, similar setting information is designated in advance as an alternative. For example, the voice recognition server device of setting D is designated in advance as an alternative to the voice recognition server device of setting E, and the IP address of the voice recognition server device of setting D is associated.

  According to the voice recognition system 1 of the present embodiment, the management unit 30 monitors the usage rate of the voice recognition server device, and changes the setting of the voice recognition server device according to the usage rate, thereby enabling a specific voice recognition server. The operation can be performed so that the load is not concentrated on the apparatus, and the utilization performance (performance) of the entire speech recognition system 1 can be improved.

  Hereinafter, the speech recognition system according to the second embodiment that achieves the same effect as that of the first embodiment by changing the threshold value for extracting the sound pickup conditions set in the client device will be described. First, the configuration of the speech recognition system according to the present embodiment will be described with reference to FIGS. FIG. 8 is a block diagram showing the configuration of the voice recognition system 4 of the present embodiment. FIG. 9 is a block diagram illustrating a configuration of the voice recognition server device 51-n according to the present embodiment. FIG. 10 is a block diagram illustrating a configuration of the management unit 60 of the present embodiment.

  As shown in FIG. 8, the speech recognition system 4 of this embodiment includes a client device 40, a plurality of speech recognition server devices 51-1,..., 51-n, 51 -N, and a management unit 60. . Although only one client device 40 is illustrated in FIG. 8, it is assumed that there are a plurality of client devices 40 as in the first embodiment. When the voice recognition server devices 51-1,..., 51-n,..., 51-N are collectively called, they are referred to as a voice recognition server device group 50. As in the first embodiment, the client device 40 and the voice recognition server device group 50 are connected to be communicable wirelessly or wired via a network. The management unit 60 may be configured as a single hardware (device), and may be referred to as a management device 60. In this case, the client device 40, the voice recognition server device group 50, and the management unit 60 (management device 60) are connected to be communicable wirelessly or wired via a network. As in the first embodiment, the management unit 60 may be a configuration requirement in the client device 40 or a configuration requirement in any of the voice recognition server devices in the voice recognition server device group 50.

  As illustrated in FIG. 8, the client device 40 includes a sound collection condition extraction unit 11, a threshold storage unit 111, a selection unit 12, a transmission destination storage unit 121, a signal processing unit 13, and a transmission unit 14. , Including the receiving unit 15, the presenting unit 16, and the threshold value changing unit 47, except that the transmission destination changing unit 17 is changed to the threshold value changing unit 47. is there. As shown in FIG. 9, the voice recognition server device 51-n (represented) includes an acoustic signal receiving unit 21A, a voice recognition unit 21B, a recognition result transmission unit 21C, a utilization rate transmission unit 21D, and a setting storage. It is the same as the speech recognition server device 21-n of the first embodiment except that the setting information receiving unit 21E and the setting update unit 21F are not included. As shown in FIG. 10, management unit 60 (management device 60) includes a utilization rate receiving unit 30A, a threshold update unit 60B, a threshold transmission unit 60C, and a threshold storage unit 60D. The utilization rate receiving unit 30A of the present embodiment is the same as the utilization rate receiving unit 30A of the first embodiment. Note that the voice recognition operation (S11 to S14, S21A to S21C, S15A, and S16) of the voice recognition system 4 of the present embodiment is exactly the same as the voice recognition operation of the first embodiment, and thus the description thereof is omitted.

Hereinafter, the information update operation of the speech recognition system 4 of the present embodiment will be described with reference to FIG. FIG. 11 is a sequence diagram showing an information update operation of the voice recognition system 4 of the present embodiment. Steps S21D and S30A are executed in the same manner as in the first embodiment. Next, the threshold update unit 60B of the management unit 60 updates the threshold value of the sound collection condition based on the above-described usage rate (S60B). This threshold value is a threshold value used for extraction of the sound pickup condition, and corresponds to, for example, the aforementioned θ ₁ , θ _{2, and the} like. The threshold update unit 60B updates the threshold according to the same policy as in the first embodiment. In other words, the load concentration is alleviated by adjusting the above-described threshold value so that the utilization rate of the voice recognition server device having a low utilization rate is increased. Alternatively, the traffic to the corresponding voice recognition server device is temporarily reduced by adjusting the threshold value so that the usage rate of the voice recognition server device having a high usage rate becomes low. The threshold before update is stored in the threshold storage unit 60D of the management unit 60, and the threshold update unit 60B stores the updated threshold in the threshold storage unit 60D. Or memorize by overwriting.

  Next, the threshold transmission unit 60C of the management unit 60 transmits the threshold updated in step S60B to the client device 40 (S60C). The receiving unit 15 of the client device 40 receives the threshold value from the management unit 60 (S15C). The threshold value changing unit 47 of the client device 40 changes the threshold value stored in the threshold value storage unit 111 based on the received threshold value (S47). The sound collection condition extraction unit 11 extracts a sound collection condition of the input acoustic signal using the changed threshold value (S11).

  According to the voice recognition system 4 of the present embodiment, the threshold for extracting the sound pickup conditions set in the client device 40 is changed so that the load is not concentrated on a specific voice recognition server device. Therefore, the utilization performance (performance) of the entire speech recognition system 4 can be improved.

<Supplementary note>
The apparatus of the present invention includes, for example, a single hardware entity as an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity. Can be connected to a communication unit, a CPU (Central Processing Unit, may include a cache memory or a register), a RAM or ROM that is a memory, an external storage device that is a hard disk, and an input unit, an output unit, or a communication unit thereof , A CPU, a RAM, a ROM, and a bus connected so that data can be exchanged between the external storage devices. If necessary, the hardware entity may be provided with a device (drive) that can read and write a recording medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device of the hardware entity stores a program necessary for realizing the above functions and data necessary for processing the program (not limited to the external storage device, for example, reading a program) It may be stored in a ROM that is a dedicated storage device). Data obtained by the processing of these programs is appropriately stored in a RAM or an external storage device.

  In the hardware entity, each program stored in an external storage device (or ROM or the like) and data necessary for processing each program are read into a memory as necessary, and are interpreted and executed by a CPU as appropriate. . As a result, the CPU realizes a predetermined function (respective component requirements expressed as the above-described unit, unit, etc.).

  The present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the above embodiment may be executed not only in time series according to the order of description but also in parallel or individually as required by the processing capability of the apparatus that executes the processing. .

  As described above, when the processing functions in the hardware entity (the apparatus of the present invention) described in the above embodiments are realized by a computer, the processing contents of the functions that the hardware entity should have are described by a program. Then, by executing this program on a computer, the processing functions in the hardware entity are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, a hardware entity is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

Claims (8)

A voice recognition system including a client device, a plurality of voice recognition server devices, and a management unit,
The client device is
A transmission unit that transmits an input acoustic signal or a signal derived from the acoustic signal to a voice recognition server device selected based on the sound collection condition;
Each of the voice recognition server devices
A setting storage unit that stores in advance settings relating to voice recognition;
A utilization rate transmitting unit that transmits information on the utilization rate, which is a rate at which the client device uses the device as a transmission destination, to the management unit;
The management unit
An operation of transmitting to the client device the threshold value used for extraction of the sound collection condition, the threshold value updated so as to reduce the bias of the utilization rate between the voice recognition server devices. For the voice recognition server device in which the usage rate is biased in comparison with other voice recognition server devices, the information regarding the setting updated so that the bias in the usage rate is reduced as a first operation The transmission destination information updated to correspond to the updated information relating to the setting, the transmission destination information being information relating to the relationship between the sound collection condition and the voice recognition server device that is the transmission destination. information, wherein when the operation to be transmitted to the client device and a second operation, the speech recognition system to perform at least one of operation of said second operation and the first operation The speech recognition system according to claim 1,
In the setting storage unit, a plurality of settings relating to voice recognition are stored in advance,
Each of the voice recognition server devices
A setting update unit that selects one of the plurality of stored settings based on setting information that is information related to the setting updated based on the utilization rate and updates the setting;
The management unit
The setting for one or more voice recognition server devices corresponding to a high usage rate compared to the usage rate of other voice recognition server devices or a low usage rate compared to the usage rate of other voice recognition server devices. A speech recognition system including a setting information transmission unit for transmitting information. The speech recognition system according to claim 1,
In the setting storage unit, a plurality of settings relating to voice recognition are stored in advance,
Each of the voice recognition server devices
A setting update unit that selects one of the plurality of stored settings based on setting information that is information related to the setting updated based on the utilization rate and updates the setting;
The management unit
For one or more speech recognition server devices corresponding to the highest utilization rate compared to the utilization rate of other speech recognition server devices or the lowest utilization rate compared to the utilization rate of other speech recognition server devices A speech recognition system including a setting information transmission unit for transmitting the setting information. The speech recognition system according to claim 2 or 3 ,
The management unit
A transmission destination information update unit that updates transmission destination information that is information relating to a relationship between a sound collection condition and a voice recognition server device that is a transmission destination when the setting information transmission unit transmits the setting information;
A transmission destination information transmission unit that transmits the updated transmission destination information to the client device;
The client device is
A destination storage unit for storing the destination information;
A transmission destination changing unit that changes the stored transmission destination information based on the transmission destination information received from the management unit;
A speech recognition system including a selection unit that selects a speech recognition server device that is a transmission destination of a corresponding acoustic signal based on the input sound collection condition of the input acoustic signal and the changed transmission destination information. The speech recognition system according to claim 1,
The client device is
A sound collection condition extraction unit that extracts sound collection conditions of the input acoustic signal;
A speech recognition system including a threshold value changing unit that changes a threshold value used for extraction of the sound pickup condition based on the threshold value of the sound pickup condition updated based on the utilization rate. A voice recognition method executed by a client device, a plurality of voice recognition server devices, and a management unit,
The client device is
Executing the step of transmitting an input acoustic signal or a signal derived from the acoustic signal to a voice recognition server device selected based on the sound collection condition;
Each of the voice recognition server devices
Storing in advance settings relating to voice recognition;
Executing the step of transmitting information relating to a utilization rate, which is a rate at which the client device has used its own device as a transmission destination, to the management unit;
The management unit
An operation of transmitting to the client device the threshold value used for extraction of the sound collection condition, the threshold value updated so as to reduce the bias of the utilization rate between the voice recognition server devices. As a first step, the information regarding the setting updated so that the bias of the utilization rate is reduced is compared with the other voice recognition server devices for the voice recognition server device in which the utilization rate is biased. The transmission destination information updated to correspond to the updated information relating to the setting, the transmission destination information being information relating to the relationship between the sound collection condition and the voice recognition server device that is the transmission destination. information, if the operation to be transmitted to the client apparatus as a second step, performing at least one of steps of said first step and said second step Speech recognition method that.   A program for causing a computer to function as a voice recognition server device included in the voice recognition system according to any one of claims 1 to 3. A program for causing a computer to function as a client device included in the voice recognition system according to any one of claims 4 and 5.

Images