JP2008250059A - Voice recognition device, voice recognition system and voice recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device, a voice recognition system and a voice recognition method for correctly recognizing input voice including noise. <P>SOLUTION: The voice recognition system includes a server, and a plurality of operation noise models generated according to each of the plurality of operations, and a plurality of environment noise models generated according to each of a plurality of places, are stored in a database in the server. A plurality of repeaters are connected to the server, and a mobile terminal is connected in a communicable manner. The mobile terminal acquires input voice and detects a behavior of a testee, and transmits voice data regarding input voice and acceleration data regarding the behavior, to the repeater. The repeater attaches a relay ID to the voice data and the acceleration data and transmits them to the server. The server specifies the behavior of the testee, and estimates a location as a current position of the testee. Then, the voice of the testee included in the input speech is recognized by using the environment noise model according to the behavior, and the environment noise model according to the current position of the testee. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a speech recognition device, a speech recognition system, and a speech recognition method, and more particularly to a speech recognition device, a speech recognition system, and a speech recognition method that recognize a subject's speech from input speech including noise.

An example of a conventional speech recognition apparatus of this type is shown in Patent Document 1. In the technique of Patent Document 1, a signal in a speech section and a signal in a noise section in an input speech signal are discriminated, and a noise model is learned from the observed signal in the noise section. Then, a speech model without noise and a noise model prepared in advance are synthesized to generate a noise superimposed speech model. Further, the noise section signal is superimposed on a reference signal prepared in advance, and a long-time average of the characteristic parameters is obtained. These operations are executed before the signal of the voice section is input. When a signal in the speech interval is input, a long-time average of the feature parameter of this signal is obtained, and a difference from the long-term average of the feature parameter of the reference signal on which the signal in the noise interval is superimposed is obtained. This difference is added to the noise superimposed speech model to obtain a CMN-completed noise superimposed speech model. Then, a model matching likelihood between the CMN-completed noise superimposed speech model and the feature parameter of the speech section signal is calculated, and a recognition result is output.
JP 2006-145694 A

  However, in the technique of Patent Document 1, a noise model is learned from a noise section in an input speech signal. However, since it is difficult to accurately distinguish between a speech section and a noise section, the noise model may not be appropriately created. There is. In addition, when the noise interval is short, data used for noise model estimation is reduced, and the reliability of the noise model is lowered. Therefore, there is a possibility that speech recognition cannot be appropriately performed with the technique of Patent Document 1.

  Therefore, a main object of the present invention is to provide a novel speech recognition apparatus, speech recognition system, and speech recognition method.

  Another object of the present invention is to provide a speech recognition device, a speech recognition system, and a speech recognition method capable of accurately recognizing even an input speech including noise.

  The present invention employs the following configuration in order to solve the above problems. The reference numerals in parentheses, supplementary explanations, and the like indicate correspondence relationships with embodiments described later to help understanding of the present invention, and do not limit the present invention in any way.

  The invention according to claim 1 is an operation noise model storage means for storing a plurality of operation noise models created corresponding to each of a plurality of movements in association with each of the plurality of movements, and an input voice including a voice of a subject Input speech detection means for detecting motion, action specification means for specifying the action of the subject, action noise model reading means for reading the action noise model corresponding to the action specified by the action specification means from the action noise model storage means, and action noise model The speech recognition apparatus includes a recognition unit that recognizes the voice of the subject included in the input voice detected by the voice detection unit using the operation noise model read by the reading unit.

  According to the first aspect of the present invention, the speech recognition device (12, 18) includes the operation noise model storage means (48), the input speech detection means (38, S1), the action specifying means (S5), and the action noise model reading means (S9). ) And recognition means (S13). The operation noise model storage means stores a plurality of operation noise models created corresponding to each of the plurality of operations in association with each of the plurality of operations. For example, the motion noise model is a noise model for noise caused by motion such as a sound of clothing being rubbed or a sound of an instrument used in a predetermined motion when the subject performs a predetermined motion. The input voice detecting means detects input voice including the voice of the subject, and the action specifying means specifies the action of the subject. The operating noise model reading means reads an operating noise model corresponding to the motion of the subject from the operating noise model storage means. The recognition means performs noise suppression processing using the read operation noise model, and recognizes the subject's voice included in the input voice. As the noise suppression processing using the noise model, for example, noise suppression processing by a PMC (Parallel Model Combination) method or GMM (Gaussian Mixture Model) can be used.

  According to the first aspect of the present invention, an operation noise model corresponding to a plurality of operations is prepared in advance, and the operation noise model corresponding to the operation of the subject is selected to perform speech recognition. The resulting noise can be suppressed appropriately and speech recognition can be performed accurately.

  The invention of claim 2 is dependent on the invention of claim 1, and a place noise model storage that stores a plurality of place noise models created corresponding to each of the plurality of places in association with each of the plurality of places. A place noise model reading means for reading a place noise model corresponding to the place specified by the place specifying means, and a place noise model reading means for reading the place noise model corresponding to the place specified by the place specifying means, Using the motion noise model read by the motion noise model reading means and the location noise model read by the location noise model reading means, the speech of the subject included in the input speech detected by the speech detection means is recognized. To do.

  More specifically, the speech recognition apparatus further includes a place noise model storage means (48), a place specifying means (S7), and a place noise model reading means (S9). The location noise model storage means stores a plurality of noise models created using noise data corresponding to the noise collected at each of the plurality of locations in association with each of the plurality of locations. For example, noise such as talking voice is generated in a place (room) where many people gather, and the operation sound is generated in a room where electronic devices are placed. A noise model for such ambient or environmental noise is stored. The place specifying means specifies the place where the subject exists, that is, the current position. The place noise model reading means reads the place noise model corresponding to the specified current position from the place noise model storage means. Therefore, the recognition means suppresses noise using the operation noise model and the location noise model, and recognizes the voice of the subject included in the input voice.

  According to the invention of claim 2, since not only the operation noise but also the environmental noise is suppressed, the voice can be recognized more accurately.

  The invention of claim 3 is dependent on the invention of claim 1 or 2, and is estimated by an estimation means for estimating a signal-to-noise ratio of a speech signal corresponding to the input speech detected by the input speech detection means, and the estimation means. And adjusting means for adjusting the synthesis ratio of the noise model according to the signal-to-noise ratio.

  The invention of claim 3 further comprises estimation means (S3) and adjustment means (S11). The estimation means estimates the signal-to-noise ratio (SNR) of the input speech, that is, the ratio of the relative magnitudes of speech and noise. The adjusting means adjusts the synthesis ratio of the noise model (only the operating noise model, both the operating noise model and the environmental noise model) according to the signal-to-noise ratio. For example, when a noise superposition model is created by synthesizing a speech model and a noise model, the synthesis ratio is adjusted according to the estimated SNR.

  According to the invention of claim 3, since model synthesis is performed in consideration of the SNR of the input speech, speech recognition can be performed more accurately.

  The invention of claim 4 is dependent on the invention of claim 2 or 3, wherein the location specifying means detects identification information issued by a repeater installed in the environment, and the location of the repeater that issued the identification information is determined by the subject. Identified as the location of

  In the invention of claim 4, the repeater (16) is installed in the surrounding or environment, that is, a plurality of places, and the place specifying means (S5) is the identification information (repeater ID) issued by the repeater existing in the communicable range. ) Is detected. And the place where the repeater is installed is specified based on the repeater ID, and the specified place is estimated (specified) as the place (current position) where the subject exists.

  According to the fourth aspect of the present invention, the current position of the subject can be easily specified, and the voice can be accurately recognized using the place noise model corresponding to the specified place.

  The invention of claim 5 is a voice recognition system comprising a portable terminal and a voice recognition device, wherein the portable terminal is detected by an input voice detecting means for detecting an input voice including the voice of the subject, and an input voice detecting means. A transmission means for transmitting a voice signal for the input voice to the voice recognition apparatus, the voice recognition apparatus being created corresponding to each of a plurality of operations, a reception means for receiving the voice signal transmitted by the transmission means A motion noise model storage unit that stores a plurality of motion noise models in association with each of the plurality of motions, a motion identification unit that identifies a subject's motion, and a motion noise model corresponding to the motion identified by the motion identification unit Using the operation noise model read out from the noise model storage means, and the operation noise model read out by the operation noise model read out means, It comprises recognition means for recognizing speech of the subject included in the input voice detected by the voice detecting means, a speech recognition system.

  In the invention of claim 5, the voice recognition system (10) includes a portable terminal (18) and a voice recognition device (12). For example, the speech recognition apparatus recognizes and records speech content (business content in the embodiment) that a subject such as a nurse speaks during work. The portable terminal is carried or worn by the subject, and transmits a voice signal of the input voice including the voice of the subject detected by the input voice detection means (38) to the voice recognition device by the transmission means (28). The speech recognition apparatus includes a receiving means (S1), an operating noise model storage means (48), an action specifying means (S5), an operating noise model reading means (S9), and a recognizing means (S13). The receiving means receives an audio signal transmitted from the mobile terminal. The operation noise model storage means stores noise (such as a rubbing sound of clothes) caused by an operation for nursing work for each operation. The action specifying unit specifies the action of the subject. The operating noise model reading means reads an operating noise model corresponding to the motion of the subject from the operating noise model storage means. The recognizing unit performs noise suppression processing using an operation noise model corresponding to the motion of the subject, and recognizes the subject's voice included in the voice signal (the voice signal of the input voice) received by the receiving unit.

  According to the invention of claim 5, as in the invention of claim 1, a noise model corresponding to a plurality of places is prepared in advance, and a speech model is selected by selecting a noise model corresponding to the place where the subject exists. Therefore, noise can be appropriately suppressed and voice recognition can be performed accurately.

  The invention of claim 6 is dependent on the invention of claim 5, and the speech recognition apparatus stores a plurality of location noise models created corresponding to each of the plurality of locations in association with each of the plurality of locations. And a place noise model reading means for reading out a place noise model corresponding to the place specified by the place specifying means from the place noise model storing means. The recognition means is included in the input speech detected by the speech detection means using the motion noise model read by the motion noise model reading means and the location noise model read by the location noise model reading means. Recognize subject's voice.

  More specifically, the speech recognition apparatus further includes a place noise model storage means (48), a place specifying means (S7), and a place noise model reading means (S9). The location noise model storage means stores a plurality of noise models created using noise data corresponding to the noise collected at each of the plurality of locations in association with each of the plurality of locations. The place specifying means specifies the place where the subject exists, that is, the current position. The place noise model reading means reads the place noise model corresponding to the specified current position from the place noise model storage means. Therefore, the recognition means suppresses noise using the operation noise model and the location noise model, and recognizes the voice of the subject included in the input voice.

  According to the invention of claim 6, since not only the operation noise but also the environmental noise is suppressed, the voice can be recognized more accurately.

  The invention of claim 7 is dependent on the invention of claim 6, further comprising a plurality of repeaters arranged corresponding to each of the plurality of places and relaying communication between the portable terminal and the server, The audio signal transmitted from the portable terminal is received, the identification information added to the audio signal received by the receiving means is added to the received audio signal, and the identification information added to the audio signal received by the receiving means is transmitted to the server. Based on the above, the installation location of the repeater that transmitted the audio signal is specified as the location where the subject exists.

  The invention of claim 7 further includes a plurality of repeaters (16) arranged corresponding to each of the plurality of places. The repeater receives an audio signal for the input voice including the voice of the subject transmitted from the portable terminal (18) existing in a communicable range, and adds its own identification information (relay ID) to the received audio signal. In addition, it is transmitted to the voice recognition device (12). The location specifying means (S7) specifies the installation location of the repeater based on the repeater ID, and specifies the installation location as the location where the mobile terminal exists, that is, the current location of the subject.

  According to the seventh aspect of the present invention, the current position of the subject can be easily specified, and the place noise model corresponding to the current position of the subject can be selected and voice recognition can be performed accurately.

  The invention of claim 8 is a speech recognition method for a computer comprising an operation noise model storage means for storing a plurality of operation noise models created corresponding to each of a plurality of operations in association with each of the plurality of operations. (A) Detect input speech including subject's speech, (b) identify subject's motion, and (c) store motion noise model corresponding to motion identified in step (b) And (d) a speech recognition method for recognizing a subject's speech included in the input speech detected in step (a) using the operating noise model read out in step (c).

  In the invention of claim 8 as well, as in the invention of the speech recognition apparatus of claim 1, speech recognition can be performed accurately.

  According to the present invention, an operation noise model corresponding to a plurality of movements is prepared in advance, and voice recognition is performed using the movement noise model corresponding to the movement of the subject. Appropriate suppression is possible, and even speech that contains noise can be accurately recognized.

  The above object, other objects, features and advantages of the present invention will become more apparent from the following detailed description of embodiments with reference to the drawings.

  Referring to FIG. 1, a speech recognition system 10 according to an embodiment of the present invention includes a server 12 that also functions as a speech recognition device, and is applied to an organization such as a hospital. Content that is uttered during work (for example, business content) is recognized and recorded.

  The server 12 is connected to a plurality of repeaters 16 via a wired or wireless communication line (network) 14. Each of the plurality of repeaters 16 is disposed at a predetermined position such as a place where the nurse performs work or work, such as an entrance to a hospital room, a bed in the hospital room or the vicinity thereof, a corridor, and a nurse station. In addition, a mobile terminal 18 is connected to each of the plurality of repeaters 16 so that wireless communication is possible. The portable terminal 18 is possessed (attached) by the nurse, and data transmitted from the portable terminal 18 is transmitted to the server via the repeater 16 that exists in a wirelessly communicable range (for example, a radius of 1 to 3 meters). 12 is transmitted.

  Although one mobile terminal 18 is shown in FIG. 1, the voice recognition system 10 may include a plurality of mobile terminals 18, and each of the plurality of mobile terminals 18 is assigned to each of a plurality of nurses. It is done. The mobile terminal 18 may be directly connected to the network 14 by a wireless LAN or the like.

  FIG. 2 is a block diagram showing a specific configuration of the mobile terminal 18, and the mobile terminal 18 includes a CPU 20. The CPU 20 includes a memory 22, an encoder 24, a non-contact sensor 26, an interface 28, a timer 30, a DIP switch 32, a wireless transmitter 34, a wireless receiver 36, and a plurality of acceleration sensors 40a, 40b, 40c, 40d, 40e, 40f. (Hereinafter, these may be collectively referred to as “acceleration sensor 40”).

  The memory 22 functions as a work memory or a buffer memory and is used by the CPU 20. A headset microphone 38 is connected to the encoder 24, and the encoder 24 modulates an audio signal for input audio input from the headset microphone 38 into compressed audio data such as MP3. The compressed audio data is stored in the memory 22 in accordance with an instruction from the CPU 20. The compressed audio data stored in the memory 22 is transmitted to the server 12 via the repeater 16 and the network 14 at regular time intervals (for example, 10 seconds to 30 seconds) in accordance with an instruction from the CPU 20.

  The reason why the audio signal is compression-modulated is to reduce the capacity of the memory 22 and to reduce the amount of data to be transmitted to the server 12.

  The headset microphone 38 used in this embodiment has directivity. This is to improve the accuracy of speech recognition by detecting the input speech from which ambient noise (noise caused by the nurse's action and environmental noise) is removed as much as possible. The headset microphone 38 is used because the nurse's hands are often closed when the nurse performs the work, and the nurse has a hand other than the tool used for the work. This is to make things as small as possible. However, instead of the headset microphone 38, a pin microphone having directivity may be attached to the neck, for example.

  A pyroelectric sensor can be used as the non-contact sensor 26, and the CPU 20 turns on / off the headset microphone 38 in accordance with an input from the non-contact sensor 26. In this embodiment, when the nurse raises and lowers his / her hand twice in front of the non-contact sensor 26, that is, the pyroelectric sensor, the detection signal is input to the CPU 20, and the CPU 20 turns on the headset microphone 38 accordingly. Then, when the nurse raises and lowers his / her hand twice in front of the pyroelectric sensor, the headset microphone 38 is turned off. The reason why the headset microphone 38 can be turned on / off is to protect the privacy of the nurse. In other words, the headset microphone 38 is turned on when the business content needs to be recognized and recorded, and the headset microphone 38 is turned off when there is no need to record the business content such as a break time.

  The interface 28 is an interface such as a LAN (wireless LAN) adapter, whereby the portable terminal 18 is connected to the network 14. Therefore, the mobile terminal 18 can communicate with the server 12 via the network 14.

  The timer 30 is a circuit that measures the date and time, and the CPU 20 acquires time data from the timer 30. The DIP switch 32 is composed of, for example, 8 bits, and can set a numerical value between 0 and 255 by switching each bit on / off. This numerical value is nurse identification information (nurse ID), and a different value is set for each portable terminal 18. CPU20 attaches time data and nurse ID as a label to the audio | voice data to transmit, and transmits to the repeater 16. That is, voice data, time data, and data (numerical data) about the nurse ID are transmitted from the portable terminal 18 to the repeater 16.

  In this embodiment, the nurse ID is set using the DIP switch 32. However, the present invention should not be limited to this. For example, instead of the DIP switch 32, a ROM storing a nurse ID may be provided.

  The wireless transmitter 34 transmits the above-described voice data, time data, and data on the nurse ID (hereinafter, these may be referred to as transmission data) to the repeater 16 in accordance with an instruction from the CPU 20. The wireless receiver 36 receives a weak radio wave emitted from the repeater 16 existing in a wireless communicable range, demodulates the repeater ID, and processes data on the demodulated repeater ID by the CPU 20.

  Each of the acceleration sensors 40 is, for example, a multi-axis (3-axis) acceleration sensor, and is used to detect the operation of a nurse who carries or wears the portable terminal 18. In this embodiment, acceleration for each part of the head, both hands, waist (or abdomen) and both feet is detected, and for each motion stored in advance in the detected acceleration data and a motion DB 46 (see FIG. 5) described later. The motion data is compared with each other to identify (identify) one motion. For example, by executing the well-known DP matching, the degree of approximation (similarity) between the detected acceleration data and the acceleration data stored in the motion DB 46 can be easily obtained.

  The portable terminal 18 configured as described above is possessed or worn by a subject such as a nurse. For example, as shown in FIG. 3, circuit components other than the non-contact sensor 26, the headset microphone 38, and the acceleration sensor 40 are accommodated in a box (housing) 60, and the box 60 is accommodated in a front pocket of a nurse's lab coat or the like. Is done. The non-contact sensor 26 is housed in a pen-type case so as to be inserted into the breast pocket of the nurse's lab coat. In the drawings, the box 60 and the non-contact sensor 26 are shown outside each pocket for easy understanding. The headset microphone 38 is attached to the nurse's head.

  Further, as described above, the acceleration sensor 40 is attached or fixed to a predetermined part of the nurse. For example, as shown in FIG. 3, the acceleration sensor 40a is attached to the nurse's head, the acceleration sensor 40b is attached to the nurse's right wrist, and the acceleration sensor 40c is attached to the nurse's left wrist. 40d is attached to the waist (or abdomen) of the nurse, acceleration 40e is attached to the nurse's right ankle, and acceleration 40f is attached to the nurse's left ankle.

  Although not shown in FIG. 3, the non-contact sensor 26 is connected to the CPU 20 in the box 60 using a connection line, and the headset microphone 38 is electrically connected to the encoder 24 in the box 60 using a connection line. The acceleration sensor 40 is connected to the CPU 20 in the box 60 using a connection line. However, you may make it connect by short distance radio | wireless like Bluetooth (trademark), without using a connection line. That is, it is only necessary to be electrically connected.

  As described above, the speech recognition system 10 recognizes and records the content of a utterance by a nurse or the like during work. Further, when performing speech recognition, noise suppression processing is appropriately performed.

As the noise suppression processing, for example, a PMC (Parallel Model Combination) method can be used. In the PMC method, a noise superimposed speech model is estimated (created) by synthesizing a speech model and a noise model, and the noise superimposed speech model and input speech are collated. As a result, even input speech including noise can be recognized with high accuracy. This PMC method can deal with input speech including noise without requiring an actual noise superimposed noise model. In a commonly used estimation method called Log-Add approximation, the average vector μ _x of the noise superimposed speech model can be estimated as shown in Equation 1.

Here, μ _s and μ _n respectively represent average vectors of logarithmic spectral energy of the speech model and the noise model. “^” Means an estimated value. same as below.

  Regarding PMC method, “MJFGales,“ Model-Based Techniques for Noise Robust Speech Recognition, ”Ph.D Thesis, Cambridge University, 1995.” and “MJFGales, SJ Young,“ A fast and flexible implementation of parallel. The model combination, “Proc. of ICASSP, pp. 133-136, 1995.” is disclosed in detail.

  Also, noise suppression processing using GMM (Gaussian Mixture Model) can be used. Assuming that there is no correlation between speech and noise, the logarithm vectors of noise superimposed speech (input speech), clean speech and noise mel filter bank output in frame i are X (i), S (i) and Assuming N (i), it can be expressed as in Equation 2.

Here, g (S (i), N (i)) is a mismatch function. Further, the auxiliary function f _b for the b-th bank of the filter bank is defined by Equation 3.

Here, s _b and n _b are the mel filter bank output for clean speech and noise. By applying the first-order Taylor expansion to Equation 2, the mean and variance can be estimated. When clean speech is represented by a K-mixed Gaussian distribution model such as Equation 4 and a noise signal is represented by a single Gaussian distribution N (μ _n , Σ _n ), the mean and variance are as shown in Equation 5 and Equation 6, respectively. It can be estimated approximately.

In practice, there are often few data when estimating the variance of noise. In addition, since the improvement in speech recognition accuracy by variance estimation is small compared to the improvement in speech recognition accuracy by average estimation, the variance of the speech model is expressed as Σ _{x, k} (b, b) ≈Σ _{s, k} ( b, b). Then, the clean voice can be expressed as in Expression 7.

  As described above, in the noise suppression processing by the GMM, only the speech is estimated from the input speech for each analysis frame using the noise model and the speech model. Then, the estimated speech and the speech model are collated. As a result, even input speech including noise can be recognized with high accuracy.

  For noise suppression processing by GMM, see “JCSegura, A. de la Torre, MCBenitez, AMPeinado,“ Model-based compensation of the additive noise for continuous speech recognition. Experiments using AURORA II database and tasks, ”Proc. of Eurospeech '01, vol. I, pp. 221-224, 2001 "for reference.

  In the noise suppression processing as described above, since the accuracy of speech recognition varies greatly depending on the noise model used, what kind of noise model is used becomes a problem. Here, as described above, the server 12 recognizes the voice of a subject such as a nurse. The sound of using or setting or installing a sphygmomanometer or infusion syringe is generated. Noise (operation noise) associated with the operation of such nursing work is input from the headset microphone 38 together with the voice of the nurse.

  Further, not only such operation noise but also noise from the surroundings (environment) is input from the headset microphone 38. For example, in the waiting room in the hospital, the patient's voice or TV (or radio) sound is generated, and in the nurse's station (nurse station), the voice of another nurse, the sound of a nurse call or other nursing care is generated. Noise is generated due to the movement of the teacher. Noise caused by such an environment (environmental noise) is also input from the headset microphone 38 together with the voice of the nurse.

  Here, a specific example of the average spectrum of the environmental noise recorded at a plurality of locations in the hospital is shown in FIG. The environmental noise is recorded for about 10 minutes in each of “Laundry room side”, “Wardway side corridor”, “Elevator hall”, “In the nurse station”, and “Staircase”. The microphone used was a small condenser microphone 4060 made by DPA, and MICROTRACK 24/96 made by M-AUDIO was used as a recording device. The average power spectrum of the environmental noise at each location is recorded at a sampling frequency of 48 kHz and 16 bits, then down-sampled to 16 kHz, short-time Fourier transformed with an analysis window length of 20 ms, and all frames of the recorded data Obtained by averaging with

  As shown in FIG. 4, the environmental noise “beside the washing machine” mainly includes the operation sound of the washing machine. In addition, the “corridor beside the hospital room”, “elevator hall”, and “stairs” were basically quiet except for occasional conversational voices, and the environmental noise was small. In addition, the environmental noise “in the nurse station” includes the operation sound generated by the equipment in addition to the occasional conversation voice, and a peak was observed in the vicinity of 500 Hz.

  In addition, although illustration is abbreviate | omitted, about operation noise, the noise which generate | occur | produces when a nurse actually performs nursing work is recorded for every nursing work.

  As can be seen from FIG. 4, the ambient environmental noise differs depending on the location. Therefore, when noise suppression processing is performed for speech recognition, noise suppression processing can be performed with higher accuracy by using a noise model specialized for the location. Can be performed. The same applies to the operation noise, and it is conceivable to use a noise model specialized for nursing work.

  Therefore, in this embodiment, the modeling of the operating noise detected for each of a plurality of nursing tasks in advance and the modeling of the environmental noise observed for each of a plurality of locations in advance are performed. The noise model is stored in a memory (in this embodiment, a database), and noise suppression processing is executed using them.

  Specifically, as shown in the block diagram of FIG. 5, a plurality of databases (DB), that is, a nurse DB 42, a repeater DB 44, an operation DB 46, a noise model DB 48, and a voice model DB 50 are connected to the server 12. . These DBs 42-50 are used to perform speech recognition by performing noise suppression processing on the input speech including the nurse's speech (speech).

  The nurse DB 42 stores information for identifying a nurse such as a nurse name in association with nurse identification information (nurse ID). The repeater DB 44 stores the location where the repeater 16 is installed in association with the identification information (repeater ID) of the repeater 16. Therefore, the server 12 can specify the nurse or the nurse name from the nurse ID, and can specify the place where the repeater 16 is installed from the repeater ID.

  Here, referring to FIG. 1, as described above, repeater 16 receives transmission data transmitted from portable terminal 18 that exists in a wireless communicable range. Then, the received transmission data is appended with its own repeater ID and transmitted to the server 12 via the network 14. That is, the nurse ID and the repeater ID are added to the voice data received by the server 12. Therefore, the server 12 can estimate (specify) the nurse who has input the voice corresponding to the received voice data and the input location (current position).

  In addition, since the repeater 16 and the portable terminal 18 can communicate with each other, the portable terminal 18 can also detect identification information (repeater ID) of the repeater 16 that exists in a range where wireless communication is possible. Therefore, instead of transmitting the transmission data with the relay ID added from the repeater 16 to the server 12, it is also possible to transmit the transmission data with the relay ID added from the portable terminal 18 to the server 12. it can. That is, the repeater ID is acquired from the repeater 16 within the communicable range of the portable terminal 18, the data about the repeater ID is attached to the transmission data, and the portable terminal 18 is connected via the interface 28 and the network 14. Can also be sent to the server 12.

  Returning to FIG. 5, the operation DB 46 stores acceleration data of the acceleration sensor 40 when executing the corresponding nursing service corresponding to the name or identification information of the nurse's operation (nursing service). For example, the acceleration data stored in the motion DB 46 can be obtained by averaging a plurality of acceleration data detected when a plurality of (many) nurses perform nursing work (motion). However, since there are individual differences in actions or actions performed by humans as in nursing work, acceleration data may be acquired for each nurse and recorded in the action DB 46. In this way, it is considered that the specific accuracy of the operation (nursing work) of the nurse is improved.

  The noise model DB 48 stores an operation noise model and an environmental noise model. Specifically, an operating noise model is created (estimated) based on operating noise data (operating noise data) that records operating noise that occurs when multiple nursing tasks are performed. Correspondingly stored. The environmental noise model is created (estimated) based on environmental noise data (environmental noise data) recorded for each of a plurality of locations, and stored in association with each location.

  In this embodiment, as will be described later, since the nurse's current position is estimated based on the installation position of the repeater 16, an environmental noise model for each place where the repeater 16 is installed is stored in the noise model DB 48. Suppose you are.

  The speech model DB 50 stores a speech model (speech data) created based only on speech that does not include noise. For example, voice data input by a nurse using the voice recognition system 10 in a place without noise is stored (recorded). When a plurality of nurses use the voice recognition system 10, a voice model is stored for each nurse, and when performing voice recognition, a voice model corresponding to each nurse is used. Thus, speech recognition can be executed more appropriately.

  An example will be shown in which such a speech recognition system 10 is used to record the work content spoken by a nurse during work. For example, if a nurse at the nurse station heads to the patient A drip, the nurse turns on the headset microphone 38 and speaks, “I will go to patient A ’s drip”, and then the headset. The microphone 38 is turned off. Then, the digital signal and the compression (modulation) process are performed on the voice signal for the input voice including the utterance (voice) of the nurse in the portable terminal 18, and the compressed voice data corresponding to the voice signal for the input voice is obtained. Generated. Also, acceleration data from the acceleration sensor 40 is detected from when the headset microphone 38 is turned on to when it is turned off. For example, acceleration data regarding an operation for preparing an instrument or medicine necessary for infusion is detected.

  The transmission data in which the time data and the nurse ID are added to the compressed voice data and the acceleration data is transmitted to the repeater 16 in the communicable range. In this case, it is transmitted to the repeater 16 arranged at the nurse station. Then, the transmission data to which the repeater ID of the repeater 16 is added in the repeater 16 is transmitted to the server 12 via the network 14.

  When the server 12 receives the transmission data to which the repeater ID is added, the time when the voice corresponding to the voice data is input is specified based on the time data, the nurse DB 48 is referred to, and the nurse ID The nurse who inputs the voice corresponding to the voice data is identified. In the server 12, the operation DB 46 is referred to, and the operation of the nurse (nursing service) is specified based on the acceleration data included in the transmission data. For example, “operation for preparing instruments and medicines necessary for infusion” is specified. Further, in the server 12, the repeater DB 50 is referred to, the repeater 16 that has transmitted (relayed) the transmission data from the repeater ID is specified, and the place where the repeater 16 is installed (here, the nurse station) is a nursing care. It is specified as the teacher's current position. Furthermore, the server 12 estimates the signal-to-noise ratio (SNR) of the voice data.

  When the nurse's action is specified, the server 12 reads out an operation noise model corresponding to the action from the noise model DB 48. When the current position of the nurse is specified, the server 12 reads an environmental noise model corresponding to the current position of the nurse from the noise model DB 48. Then, the server 12 performs reverberation suppression processing on the audio signal corresponding to the audio data (the audio signal of the input audio) using the read operation noise model and environmental noise model, and is included in the input audio. The nurse's voice is recognized.

  When performing speech recognition, noise suppression processing using a noise model such as the above-described PMC method or GMM is performed. At this time, if model synthesis is performed in consideration of the SNR of the input speech, speech recognition can be executed with higher accuracy.

  In this way, the voice of the nurse, that is, the voice “I will go to the drip of patient A”, is recognized by using the noise model of the noise caused by the movement and location when the voice is emitted. The Then, the recognition result is output as a text sentence and stored in a memory in the server 12, for example. Further, since the server 12 can specify the time when the voice was emitted and the nurse (nurse name) who issued the voice, the voice recognition result is the time when the voice was emitted and the voice. Memorized with nurse name.

  Note that text data such as nurse's utterance contents (ie, nurse's work contents) stored in the server 12 is appropriately checked by accessing the server 12 from a computer such as a personal computer assigned to the nurse. And acquisition is possible.

  Although detailed explanation is omitted, for example, when a nurse moves into a hospital room and utters and records the utterance in the hospital room as in the above example, the location (in this case) An operation noise model and an environmental noise model corresponding to the location are selected (read out) corresponding to the operation in the room), and voice recognition is performed accurately using them.

  Below, the process which the voice recognition system 10 recognizes a nurse's voice from the input voice containing noise is demonstrated using a flowchart. Specifically, the server 12 executes the entire process according to the flowchart shown in FIG. Note that, here, the case where the noise suppression process is executed using the PMC method will be described, but the noise suppression process can also be executed by the GMM.

  As shown in FIG. 6, when the server 12 starts the entire process, in step S <b> 1, an input voice is acquired. That is, the voice data about the input voice transmitted from the operation terminal 18 worn by the nurse through the repeater 16 is acquired. Specifically, the server 12 receives the transmission data to which the repeater ID is attached. This transmission data includes voice data, acceleration data, nurse ID, and time data. In the subsequent step S3, the SNR of the acquired input speech is estimated. That is, the ratio of the relative sizes of speech and noise is estimated.

  In subsequent step S5, the operation of the nurse is specified. Specifically, the server 12 compares the acceleration data included in the transmission data with the acceleration data for each operation stored in the operation DB 46, and determines the operation corresponding to the most approximate (similar) acceleration data as a nurse. Specified as the action.

  Although illustration is omitted, if all the approximations of the acceleration data included in the transmission data and the acceleration data stored in the motion DB 46 are less than a certain threshold, it is determined that there is no corresponding motion. It is. This is because an operation is not always accompanied when a voice is input.

  Subsequently, in step S7, the current position of the nurse is specified. Specifically, the server 12 refers to the repeater DB 44 and identifies the place where the repeater 16 described corresponding to the repeater ID acquired in step S1 is arranged as the current position of the nurse. .

  Next, in step S9, a noise model is determined from the identified nurse action and the current position. That is, the server 12 reads out from the noise model DB 48 an operation noise model corresponding to the nurse's operation and an environmental noise model corresponding to the nurse's current position. Further, in step S11, a noise superimposed speech model is created. That is, the server 12 synthesizes the noise model determined in step S9 and the speech model stored in the speech model DB 50 to create a noise superimposed speech model. When creating a noise-superimposed speech model, the synthesis ratio is adjusted based on the SNR of the input speech estimated in step S3.

  In subsequent step S13, voice recognition is executed. That is, the noise superposition model and the input voice are collated, and the nurse's voice is recognized from the input voice. In step S15, the recognition result is output. For example, the server 12 outputs the recognition result to the internal memory, records it as text data, and ends the entire process.

  According to this embodiment, an operation noise model corresponding to each of a plurality of movements and an environmental noise model corresponding to each of a plurality of places are stored in advance, and noise corresponding to the movement of the subject and the place where it exists Since speech recognition is performed using the model, noise included in the input speech can be appropriately suppressed, and speech recognition can be performed accurately even for input speech including noise.

  Moreover, according to the above-mentioned Example, since the installation position of the repeater 16 which relayed the transmission data from the portable terminal 18 is specified as a place where the subject exists, the current position of the subject can be easily specified.

  However, the method for specifying the current position of the nurse (subject) is not limited to this, and the current position of the subject may be specified using an appropriate method. For example, when the speech recognition system 10 is applied to a subject existing outdoors, the current position of the subject can be detected using a known GPS.

  Further, as described in the above embodiment, when a subject exists indoors, the current position of the subject can be specified by a passage sensor or the like. For example, a tag (such as a wireless tag or an infrared LED tag) that transmits identification information to a subject is attached, and a tag reader that receives identification information from the tag is provided at an appropriate location such as a hospital room entrance or a corridor ceiling. Like that. In such a case, the server 12 detects the current position of each subject by managing the entrance / exit of the subject's hospital room, and manages information on the current position in association with the identification information of the subject. Then, when a voice is input by the subject, the current position of the subject may be acquired.

  Furthermore, the movement of the subject can be estimated based on the current position of the subject. For example, if a place (room) is fixedly determined for each nursing task (work) performed by the nurse, the nursing task (operation) can be specified from the current position of the nurse. Therefore, in such a case, the acceleration sensor 40 shown in the above embodiment can be omitted. For this reason, the capacity of the transmission data can be reduced.

  In such a case, the location DB 52 is further connected to the server 12 as shown in FIG. The location DB 52 stores names or identification information of corresponding nurse operations (nursing duties) in association with names or identification information of a plurality of locations.

  Specifically, the server 12 executes the entire process according to the flowchart shown in FIG. However, since it is substantially the same as the processing described with reference to FIG. 6 in the above-described embodiment, different contents will be described and redundant description will be omitted. Further, as described above, in order to specify the operation of the nurse from the current position of the nurse, the process of step S5 shown in FIG. 6 is omitted.

  As shown in FIG. 8, when the server 12 specifies the nurse's current position in step S7, the server 12 refers to the location DB 52 in step S21 and stores the nurse's operation stored in correspondence with the specified current position. Identify nursing work.

  Although illustration is omitted, when the place DB 52 is provided, it is possible to specify the place from the operation of the nurse, contrary to the above. In such a case, the repeater 16 simply relays transmission data from the portable terminal 18. That is, the repeater 16 does not need to add the repeater ID. Further, since the current position of the nurse is not estimated (specified) based on the repeater ID, the repeater DB 44 can be deleted. As the operation, when the server 12 identifies the operation of the nurse in step S5 shown in FIG. 6, the server 12 refers to the location DB instead of the processing in step S7, and determines the location of the nurse, that is, the current position from the operation. Identify. That is, the process of the server 12 is the same as the process shown in FIG. 6 except that the process of step S7 is different.

  In the above-described embodiment, the server 12 executes the voice recognition process. However, the present invention is not limited to this, and the CPU 20 of the mobile terminal 18 can also execute the voice recognition process. That is, the portable terminal 18 can function as a voice recognition device. In this case, the mobile terminal 18 may have the database (42-50) shown in the above embodiment. Further, when the mobile terminal 18 executes the voice recognition process, a database connected to the network 14 or an external computer (for example, the server 12) may be referred to as appropriate.

  Furthermore, in the above-described embodiment, only the case where both the operating noise and the environmental noise are suppressed has been described. However, even if either one is suppressed, the accuracy of speech recognition is higher than when both are not suppressed. Can be improved. For example, when only operating noise is suppressed, the process of step S7 in FIG. 6 is deleted, and only the operating noise model is determined in step S9. In such a case, it is not necessary to construct an environmental noise model in the noise model DB 48. Further, when only the environmental noise is suppressed, the processing in step S5 in FIG. 6 is deleted, and in step S9, only the environmental noise model is determined. In such a case, it is not necessary to construct an operating noise model in the noise model DB 48.

  Furthermore, in the above-described embodiment, only the case where the voice recognition system 10 is applied to an organization such as a hospital and recognizes the voice of a subject such as a nurse has been described. However, the application need not be limited to this. Absent. For example, the voice of a doctor may be recognized, or the voice of a worker in the factory may be recognized by applying to a factory.

FIG. 1 is an illustrative view showing one example of a voice recognition system of the present invention. FIG. 2 is an illustrative view showing an electrical configuration of the portable terminal of FIG. FIG. 3 is an illustrative view showing a state in which a subject (nurse) wears the portable terminal of FIG. FIG. 4 is a graph showing an average power spectrum of environmental noise observed at each of a plurality of locations. FIG. 5 is a block diagram showing a database connected to the server of FIG. FIG. 6 is a flowchart showing the overall processing of the server of FIG. FIG. 7 is a block diagram illustrating a database connected to a server according to another embodiment. FIG. 8 is a flowchart showing the overall processing of the server according to another embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Voice recognition apparatus 12 ... Server 14 ... Network 16 ... Repeater 18 ... Portable terminal 38 ... Headset microphone 40 (40a, 40b, 40c, 40d, 40e, 40f) ... Acceleration sensor

Claims (8)

An operation noise model storage means for storing a plurality of operation noise models created corresponding to each of the plurality of operations in association with each of the plurality of operations;
Input voice detection means for detecting input voice including the voice of the subject;
Action specifying means for specifying the action of the subject;
The operation noise model reading means for reading out the operation noise model corresponding to the operation specified by the operation specifying means from the operation noise model storage means, and the operation noise model read out by the operation noise model reading means, A speech recognition apparatus comprising recognition means for recognizing the voice of the subject included in the input speech detected by the speech detection means. A place noise model storage means for storing a plurality of place noise models created corresponding to each of the plurality of places in association with each of the plurality of places;
A place identifying means for identifying the place where the subject exists, and a place noise model reading means for reading a place noise model corresponding to the place specified by the place specifying means from the place noise model storage means,
The recognizing unit uses the operation noise model read by the operation noise model reading unit and the place noise model read by the place noise model reading unit to detect the input speech detected by the sound detection unit. The speech recognition apparatus according to claim 1, wherein the speech of the subject included in the subject is recognized. Estimating means for estimating a signal-to-noise ratio of a speech signal with respect to input speech detected by the input speech detecting means; and adjusting means for adjusting a synthesis ratio of a noise model according to the signal-to-noise ratio estimated by the estimating means The speech recognition apparatus according to claim 1, further comprising:   The said place specific means detects the identification information which the repeater installed in the environment emits, and specifies the installation place of the repeater which issued the said identification information as a place where the said test subject exists. Voice recognition device. A speech recognition system comprising a mobile terminal and a speech recognition device,
The portable terminal is
Input voice detection means for detecting input voice including the voice of the subject, and transmission means for sending a voice signal for the input voice detected by the input voice detection means to the voice recognition device,
The voice recognition device
Receiving means for receiving the audio signal transmitted by the transmitting means;
An operation noise model storage means for storing a plurality of operation noise models created corresponding to each of the plurality of operations in association with each of the plurality of operations;
Action specifying means for specifying the action of the subject;
The operation noise model reading means for reading out the operation noise model corresponding to the operation specified by the operation specifying means from the operation noise model storage means, and the operation noise model read out by the operation noise model reading means, A speech recognition system comprising recognition means for recognizing the voice of the subject included in the input speech detected by the speech detection means. The voice recognition device
A place noise model storage means for storing a plurality of place noise models created corresponding to each of the plurality of places in association with each of the plurality of places;
A place identifying means for identifying the place where the subject exists, and a place noise model reading means for reading a place noise model corresponding to the place specified by the place specifying means from the place noise model storage means,
The recognizing unit uses the operation noise model read by the operation noise model reading unit and the place noise model read by the place noise model reading unit to detect the input speech detected by the sound detection unit. The speech recognition system according to claim 5, wherein the speech of the subject included in the subject is recognized. A plurality of repeaters that are arranged corresponding to each of the plurality of places and relay communication between the mobile terminal and the voice recognition device;
The repeater receives the audio signal transmitted from the mobile terminal, adds its own identification information to the received audio signal, and transmits it to the audio recognition device,
The location specifying means specifies the installation location of the repeater that has transmitted the audio signal as the location where the subject exists based on the identification information added to the audio signal received by the receiving means. The speech recognition system described. A computer speech recognition method comprising an operation noise model storage means for storing a plurality of operation noise models created corresponding to each of a plurality of operations in association with each of the plurality of operations,
(a) Detect input speech including subject's speech,
(b) identify the subject's movement;
(c) reading out an operating noise model corresponding to the operation specified in step (b) from the operating noise model storage means; and
(d) A voice recognition method for recognizing the voice of the subject included in the input voice detected in step (a) using the operating noise model read out in step (c).