JP5017246B2 - Dictionary learning apparatus and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a dictionary updating system which can update the dictionary for precisely identifying the scene. <P>SOLUTION: This updating system has a sound feature extracting section 11 which extracts the sound feature amount for identifying the scene by analyzing the sound signals, a relevant information acquisition section 21 which acquires the information other than the sound signals as the relevant information, a dictionary controller 61 which controls the dictionary of the sound feature amounts by correlating it with the relevant information, a similarity calculation section 41 which calculates the similarity by collating the sound feature amounts and the dictionary, a learning controller 31 which calculates the similarity of those having the common information from the dictionary and updates the dictionary when the similarity is larger than the first threshold, and a dictionary updating section 51 which updates the sound signal parameters used to collate the sound feature amounts to obtain the similarity when updating the dictionary. <P>COPYRIGHT: (C)2010,JPO&amp;INPIT

Description

  The present invention relates to a dictionary learning apparatus and method for adaptively learning a dictionary used for scene identification.

  With recent advances in various sensor technologies and semiconductor technologies, small devices equipped with various sensors have become widespread. For example, a recent mobile phone is equipped with a GPS (Global Positioning System), and the position of a user holding the mobile phone can be grasped. A service that uses this location information to guide directions to a destination has been realized. In addition, in a device equipped with an acceleration sensor, it is possible to grasp a user's movement status such as being stopped / running / on a vehicle. For example, by installing an acceleration sensor in a music player, it is possible to control to play music while running, but to reduce the volume when riding on a vehicle.

  If the user's situation can be ascertained in this way, it is possible to provide information and switch functions depending on the situation, but there is a limit to grasping the user's situation widely using only a single sensor. For example, the position can be grasped by using the GPS, but it is known that the accuracy of the indoor is lower than that of the suburbs, and the GPS cannot be applied to an application that switches functions and services depending on the indoor position. Although it is possible to grasp the state of the vehicle by using the acceleration sensor, it is not easy to grasp the type of the vehicle such as a bus or a train.

  To solve these problems, we have prepared a solution such as RFID (Radio Frequency IDentification) tag so that the indoor position can be grasped and information for grasping the detailed situation using the position as a key. Solutions such as searching can be considered. These are solutions by infrastructure development, but if the balance between the cost of infrastructure development and the functions provided is poor, it will not be a realistic solution.

  On the other hand, there is an approach to expand the range of situations that can be grasped by cooperating multiple sensors instead of infrastructure maintenance. For example, considering the above-mentioned vehicles such as buses and trains, the running sounds generated by the buses and trains are different, and by acquiring an acoustic signal from a microphone, it is possible to identify a situation that is difficult to grasp only by acceleration. In Patent Literature 1 and Non-Patent Literature 1, a dictionary for identification is learned from voice data collected in a plurality of situations including buses and trains, and the situation is grasped from the dictionary with the highest similarity by collating with the dictionary. The method is public. In this way, by coordinating the acceleration and the microphone, it is possible to grasp a situation such as “moving on a bus” which is difficult with a single sensor.

  There are also terminals that support factory patrols that check whether a plurality of machines installed in the factory are operating normally. Assuming an application that has a function to switch and provide items to be confirmed for each machine in front of the user's eyes, the GPS cannot grasp the detailed indoor position, but it can be driven by using a microphone. Can understand the difference between machines with different sounds.

  Thus, the situation which can be grasped spreads by cooperating the sensor and microphone which were conventionally used for the situation grasp. In addition, microphones are installed in many small devices such as mobile phones, and the introduction cost is low compared to infrastructure maintenance.

  However, the methods of Patent Document 1 and Non-Patent Document 1 have a problem that it is necessary to learn a dictionary to be referred to in order to appropriately determine a scene from a large amount of acoustic signals. For example, in order to identify a scene with high accuracy, it is necessary to collect learning data of at least several tens of minutes to several hours. In addition, when the microphone is continuously operated, as in the case of buses and trains, not only acoustic signals mixed from a single sound source that can be easily identified, but also multiple sound sources are mixed, or the device itself is moved. Since noise or the like due to the input is input, there is a problem that the scene may not be sufficiently identified only by the acoustic signal.

As a method for strengthening a dictionary in which a large amount of learning data cannot be prepared and sufficient performance cannot be obtained, online adaptation in which the dictionary is updated using data input at the time of scene identification can be considered. For example, in speaker verification by voice, an on-line adaptation method is adopted as a method for dealing with changes in voice quality over time and environmental changes (see Patent Document 2). However, in these methods, it is assumed that which dictionary is updated among the collation dictionaries prepared for each speaker is determined from information for specifying the speaker such as an ID. It has become. The input of the ID corresponds to designating the corresponding dictionary to the user as to what the situation to be updated is, and there is a problem that the user is troublesome. If a dictionary to be adapted is selected from all dictionaries, there is no need for the user to be troublesome, but there is a problem in that if the dictionary selection fails, incorrect adaptation is performed. Especially considering the situation of judging the difference in vehicle type and machine type given in the above example, there is a possibility that a similar dictionary will be constructed in the initial stage because there are similar environmental sounds in different places As a result, there is a problem that erroneous adaptation frequently occurs.
JP 2002-323371 A Japanese Patent No. 3849841 Peltonen, V. et.al "Computational auditory scene recognition" ICASSP2002

  Thus, in the past, even if an attempt was made to supplement a scene that could not be grasped only by other sensors using information obtained by analyzing acoustic information, the dictionary required to identify the scene did not match the environment. There was a problem that sufficient discrimination performance was not achieved. Further, when adapting a dictionary that does not match the environment, there is a problem in that a dictionary corresponding to a similar scene is erroneously updated unless a target dictionary is designated.

  In addition, when identifying scenes using only acoustic information or adapting a dictionary, there is a problem that sufficient performance is not achieved because there are mixed unreliable acoustic signals that are not suitable for identification and adaptation. There is a point.

  Therefore, the present invention has been made in view of the above problems, and it is an object of the present invention to provide a dictionary learning apparatus and method that can update a dictionary for accurately identifying a scene.

The present invention relates to an acoustic feature extraction unit that analyzes an input acoustic signal and extracts an acoustic feature quantity to be identified for identifying a scene, and to acquire information other than the input acoustic signal as related information An information acquisition unit, a dictionary management unit that manages a plurality of previously prepared acoustic feature quantity datums in association with the related information, and the identification target acoustic feature quantity and the plurality of dictionaries are similar to each other A similarity calculating unit for calculating the degree of each, and the related information acquired from the plurality of dictionaries is narrowed down to the common dictionary, the similarity of the narrowed down dictionary is acquired from the similarity calculating unit , A learning control unit that controls to update the narrowed-down dictionary when the similarity is larger than a first threshold, and updates the dictionary that has been narrowed down and the similarity is larger than the first threshold. Place to do , The acoustic signal parameter for determining the degree of similarity, a dictionary learning device characterized by comprising a, a dictionary updating unit that updates using acoustic features of the identification target.

  According to the present invention, it is possible to update the dictionary to improve the scene identification accuracy.

  A dictionary learning apparatus according to an embodiment of the present invention will be described with reference to the drawings.

(First embodiment)
A dictionary learning apparatus according to a first embodiment of the present invention will be described with reference to FIGS.

(1) Configuration of Dictionary Learning Device First, the configuration of the dictionary learning device will be described with reference to FIG. FIG. 1 is a schematic configuration diagram of a dictionary learning apparatus according to the present embodiment.

  As illustrated in FIG. 1, the dictionary learning device includes an acoustic feature extraction unit 11, a related information acquisition unit 21, a learning control unit 31, a similarity calculation unit 41, a dictionary update unit 51, and a dictionary management unit 61.

  This dictionary learning device can also be realized, for example, by using a general-purpose computer device as basic hardware. That is, the acoustic feature extraction unit 11, the related information acquisition unit 21, the learning control unit 31, the similarity calculation unit 41, the dictionary update unit 51, and the dictionary management unit 61 cause the processor mounted on the computer device to execute the program. Can be realized. At this time, the dictionary learning device may be realized by installing the above program in a computer device in advance, or may be stored in a storage medium such as a CD-ROM, or distributed through the network. Thus, this program may be realized by appropriately installing it in a computer device.

(1-1) Acoustic feature extraction unit 11
The acoustic feature extraction unit 11 analyzes the input acoustic signal, extracts an acoustic feature amount that emphasizes the difference in the scene from the acoustic signal for scene identification, and sends the extracted acoustic feature amount to the learning control unit 31. Output. Details of a method for extracting an acoustic feature amount from an acoustic signal for scene identification will be described later.

  The “scene” means a place or situation. The situation is, for example, changes accompanying changes in time, place, and surrounding environment. Moreover, there is no restriction | limiting in particular in the input method of the acoustic signal used as analysis object, For example, you may make it the structure which can be acquired in real time from the audio | voice input part provided with the microphone, amplifier, AD converter etc., for example. Further, a configuration may be adopted in which an acoustic signal stored in a storage device as a digital signal can be acquired by reading.

(1-2) Related information acquisition unit 21
The related information acquisition unit 21 acquires information of sensors other than acoustic signals such as an acceleration sensor and GPS, and outputs the acquired related information to the learning control unit 31.

  The related information to be output may be an acoustic feature quantity obtained by analyzing a raw data series obtained from a sensor or a scene estimation result estimated from an acoustic feature quantity.

  For example, in the case of an acceleration sensor capable of acquiring three-axis acceleration, by analyzing XYZ-axis acceleration data obtained for each operation cycle, an acoustic feature amount such as an absolute movement amount (G value) can be extracted. It is also possible to grasp the situation such as stationary, walking, and running from the XYZ-axis acceleration data and the absolute movement amount obtained for each operation cycle.

  Therefore, it is possible to output the above-described acoustic feature amount and determination result as related information. Details of the related information acquisition unit 21 will be described later.

(1-3) Dictionary management unit 61
The dictionary management unit 61 manages a plurality of determination dictionaries (hereinafter simply referred to as “dictionaries”) that are collated in order to identify a scene from the acoustic feature amount in association with the related information described above.

  As a dictionary, a mixed normal distribution (Gaussian Mixture Model, GMM) that statistically represents an appearance pattern of an acoustic signal observed in a scene to be identified can be considered as an example. A plurality of dictionaries corresponding to a scene to be identified are classified and managed for each related information obtained from the related information acquisition unit 21. Details will be described later.

(1-4) Similarity calculation unit 41
The similarity calculation unit 41 compares the acoustic feature quantity with a dictionary prepared in advance and calculates a “similarity score” representing the similarity. When the dictionary is the GMM described above, the acoustic signal parameters such as the average vector of the acoustic feature amount, the dispersion vector, and the weights of a plurality of vectors possessed by the GMM to express the appearance pattern of the acoustic signal are used. The likelihood that can be calculated by the following equation (1) can be used as the similarity score. The similarity calculation unit 41 calculates a likelihood obtained by converting the likelihood into a posteriori probability as the similarity. Details of the similarity score will be described later.

(1-5) Learning control unit 31
The learning control unit 31 determines whether or not to use the acoustic feature amount obtained from the acoustic feature extraction unit 11 for updating the dictionary managed by the dictionary management unit 61. When there is a dictionary to be updated, the learning control unit 31 notifies the dictionary update unit 51 of the dictionary to be updated and the obtained acoustic feature amount.

  The learning control unit 31 narrows down candidate dictionaries among the dictionaries in the dictionary management unit 61 based on the related information obtained from the related information acquisition unit 21 in order to determine the necessity of updating. The similarity with the acoustic feature amount for the narrowed-down dictionary is acquired by inquiring the similarity calculation unit 41. By comparing the acquired similarities, it is determined whether or not there is a dictionary to be updated, and if a dictionary to be updated is found, the dictionary update unit 51 is controlled to learn the dictionary. Details of the method for determining the necessity of update will be described later.

(1-6) Dictionary update unit 51
The dictionary update unit 51 updates the update target dictionary input from the learning control unit 31 by using the acoustic feature amount input from the learning control unit 31. More specifically, the acoustic signal used for collation with the acoustic feature quantity is updated in order to obtain the similarity. When the dictionary is the GMM described above, for example, the dictionary can be updated by changing the mean vector, the variance vector, and the weight vector, which are acoustic signal parameters, by the EM algorithm.

(1-7) Effect As described above, in this embodiment, when updating a dictionary used for scene identification, an operation is performed to narrow down a dictionary to be updated from related information. Thereby, the dictionary of a scene that emits a similar acoustic signal that originally exists in another situation is not erroneously updated, and the scene identification performance can be improved.

(2) Specific example Next, the specific example which applied the operation | movement of the dictionary learning apparatus of a present Example to the information provision apparatus is demonstrated. The dictionary learning device includes an acceleration sensor, a GPS, and a microphone. The dictionary learning device supports security guards who patrol the factory premises, identifies scenes for each patrol place, and checks the contents to be confirmed for each identified scene (patrol place). Consider a case where the present invention is applied to a provided information providing apparatus.

  That is, the information providing apparatus includes the dictionary learning apparatus and the scene identification apparatus of the present embodiment. The scene identification device identifies a scene in the factory based on the dictionary updated by the dictionary learning device of this embodiment. Then, the information providing device provides information based on the identified scene. FIG. 2 shows a schematic diagram of the site of a factory where guards go around with this information providing device in this embodiment. FIG. 3 shows an example of information displayed according to the front view of the information providing apparatus and the scene identification result.

  In the schematic diagram of FIG. 2, a plurality of machines are installed in the site of a factory where it is difficult to grasp an accurate position by GPS. When a security guard moves within the site with an information providing device equipped with an acceleration sensor, GPS, and microphone, the identification unit of the scene identification device is updated to the dictionary updated by the dictionary learning device of the present embodiment. Identify scenes in the factory based on. And the provision part of an information provision apparatus provides the information which a guard should confirm about the machine in front of the place where the guard is patroling corresponding to the identified scene. As shown in FIG. 3, the contents to be checked are, for example, confirmation of operation and confirmation of oil leakage in front of the machine 3 at the site A.

(2-1) Dictionary management unit 61
A specific example of the dictionary management unit 61 will be described.

  In the factory patrol in the present embodiment, it is possible to adjust the machine such that the guards patrol the factory in advance and record the sound of the site before the actual production. It is difficult to build a high-performance dictionary because it is not realistic for guards to make long visits in advance. However, the dictionary learning device can hold a dictionary as a seed linked to related information such as location and situation.

  Therefore, the dictionary for each machine is associated with the movement information obtained from the location and the acceleration by the preliminary tour. However, the machine-specific dictionary is constructed and prepared, although it does not show high accuracy.

  FIG. 4 shows a specific example of the dictionary constructed in the preliminary patrol stage. As shown in FIG. 4, the dictionary is prepared for each machine to be visited, and is classified and held for each place and movement state that can be acquired from related information.

(2-2) Description of Acoustic Feature Extraction Unit 11 A specific example of the acoustic feature extraction unit 11 will be described.

  The acoustic feature extraction unit 11 extracts an acoustic feature amount from the input acoustic signal. As an acoustic feature quantity for judging the situation from the acoustic signal, a case where (1) power fluctuation or (2) zero crossing number fluctuation is used is considered.

“Power fluctuation” is an acoustic feature amount focusing on short-term fluctuations in energy, and can be obtained as LSTERb in the following equation (2).

  As shown in the equation (2), the acoustic signal x (t) is cut out by a fixed time length N using a window function called a Hamming window w (t), and then the energy STE in units of frames is calculated. Compared with the average energy AveSTE obtained by averaging this energy STE in B frames, the energy STE can be calculated as a ratio including frames having energy values smaller than 0.5 times.

The “zero crossing variation” is an acoustic feature amount focusing on short-term variation in the number of zero crossings, and can be obtained as HZCRRb in the following equation (3).

  As shown in the equation (3), after the acoustic signal x (t) is cut out by a fixed time length N, the rate at which the sign of the signal changes in the frame is calculated as the zero crossing rate ZCR. Compared with the average zero-crossing change rate AveZCR obtained by averaging this zero-crossing rate in B frames, it can be calculated as a ratio including frames having a value larger than 1.5 times.

  Note that the above-described two acoustic feature quantities are examples of acoustic feature quantities used for identifying a scene, and other known acoustic feature quantities may be used.

(2-3) Description of Related Information Acquisition Unit 21 A specific example of the related information acquisition unit 21 will be described with reference to FIG. FIG. 6 is a flowchart illustrating the detailed operation of the related information acquisition unit 21.

  The related information acquisition unit 21 grasps the position and movement status from the GPS and the acceleration sensor, and outputs the information as related information.

  First, in step S2101, the related information acquisition unit 21 acquires X-axis, Y-axis, and Z-axis acceleration values (Gx, Gy, Gz) from an acceleration sensor that can acquire 3-axis acceleration data. The acceleration data can be acquired for each operation cycle of the sensor, and for example, acceleration data can be acquired every 10 ms.

Next, in step S2102, the related information acquisition unit 21 acquires acoustic feature quantities related to acceleration from the acceleration values (Gx, Gy, Gz) acquired for each operation cycle. As the acoustic feature amount, for example, a G value expressed by the following equation (4), a movement amount of each axis, and the like are calculated.

  Next, in step S2103, the related information acquisition unit 21 determines the movement status from the movement amount of each axis and the acoustic feature amount of the G value. In the present embodiment, consider an example in which two situations are determined, which are moving or stationary. Here, a threshold value is set for the movement amount and G value of each axis, and the movement status of whether it is stationary or moving is determined by comparison with the threshold value. When considering various situations such as “walking”, “jogging”, “running”, “vehicle” as well as the two situations of stationary or moving, it is not a simple threshold process but a statistical corresponding to the situation. A dictionary may be prepared, and a situation with a high score may be determined as a determination result by collation with the dictionary. In this case, a statistical model such as the GMM or HMM described above can be used as the dictionary.

  Next, in step S2104, the related information acquisition unit 21 proceeds to step S2105 in order to acquire position information using GPS when the estimated movement state changes or moves a certain amount. on the other hand. If there is no change, the process returns to step S2101.

  Next, in step S <b> 2105, the related information acquisition unit 21 converts the position information to be acquired into position information such as the site A and the site C by clustering the acquired position information into meaningful granularities such as the site. Note that, as in the present embodiment, in a place where GPS does not reach such as indoors, it is determined that the situation is the same as before, and the position information may be complemented from the situation immediately before.

  Next, in step S2106, the related information acquisition unit 21 outputs the movement status and position information determined in step S2103 to the learning control unit 31 as related information.

  Then, the same processing is repeated until all data is processed.

(3) Flow of updating the dictionary The flow of updating the dictionary while the factory learning is actually performed while the dictionary learning device manages the dictionary for each machine in the dictionary management unit 61 will be described with reference to the flowchart of FIG. To do.

  First, in step S0101, the acoustic feature extraction unit 11 extracts an acoustic feature amount from an input acoustic signal.

  Next, in step S0102, the related information acquisition unit 21 acquires related information from the GPS and the acceleration sensor.

  Next, in step S0103, when the learning control unit 31 acquires the acoustic feature amount and the related information, the dictionary that is a candidate to be updated using the related information from the dictionary for each machine managed by the dictionary management unit 61. Narrow down. For example, a case where the situation is currently stopped from the site A and the situation is known from the related information will be considered below. Then, the learning control unit 31 determines that the dictionary corresponding to the machine 1, the machine 2, and the machine 3 to which the same label is assigned in the dictionary management unit 61 is a dictionary to be updated, and proceeds to step S0104. Note that a portion surrounded by a frame in the flowchart of FIG. 5 corresponds to the operation of the learning control unit 31.

Next, in step S0104, the learning control unit 31 compares the three narrowed dictionaries with the input acoustic feature amount to calculate a similarity score. The similarity score is acquired by making an inquiry to the similarity calculation unit 41. For example, when the dictionary is constructed by a mixed normal distribution, the similarity score can be calculated by the following equation (5).

  Here, it is assumed that the similarity scores corresponding to the machine 1, the machine 2, and the machine 3 are score1, score2, and score3, respectively, and score1> score2 and score3.

  Next, in step S0105, the learning control unit 31 determines whether there is a dictionary to be adapted based on the obtained similarity score. For example, a method of determining as follows by setting the threshold value of the similarity score as α and β can be considered.

score1 ≧ α... Judged as learned.
α> score1 ≧ β... Adapting dictionary 1
score1 <β... Judged as an unknown situation.

  That is, the learning control unit 31 performs processing by determining that the dictionary managed by the dictionary management unit 61 is sufficiently learned and does not need to be adapted when score1 exceeds the threshold α in accordance with the above-described determination criteria. Finish.

  If the score 1 is less than the threshold value β, the learning control unit 31 determines that the sound does not match any sound of the machine 1, the machine 2, and the machine 3, and ends the process.

  Further, when α> score1 ≧ β, the learning control unit 31 determines that an adaptation target has been found, and proceeds to step S0106.

  As described above, the learning control unit 31 operates so as to determine a dictionary to be adapted based on the similarity score after narrowing down candidates of the dictionary to be adapted.

  Next, in step S0106, the dictionary update unit 51 updates the dictionary to be updated input from the learning control unit 31 using the acoustic feature amount input from the learning control unit 31, and then updates the existing dictionary. It operates to replace the dictionary of the management unit 61. When the dictionary is the GMM described above, the dictionary can be updated by using an existing algorithm such as the EM algorithm. By this operation, the dictionary corresponding to the machine 1 is learned from a larger number of acoustic signals than before, so that the scene identification performance is improved.

(4) Effects As described above, according to the present embodiment, when updating a dictionary used for scene identification, an operation is performed to narrow down a dictionary to be updated from related information.

  For example, considering the case where the machine 1 and the machine 4 make very similar sounds in the present embodiment, the dictionary of the machine 4 using the acoustic signal generated from the machine 1 in the case where the narrowing down by the related information is not performed. There is a risk of accidental adaptation. However, in this embodiment, such misadaptation can be avoided, so that the scene identification performance can be improved, and scenes that are difficult to identify with other sensors can be coordinated with a microphone with high accuracy. Can be identified.

(Second embodiment)
A dictionary learning apparatus according to a second embodiment of the present invention will be described with reference to FIGS.

  In the present embodiment, a dictionary construction unit 71 for adding a new dictionary is added to the dictionary management unit 61. This embodiment is different from the first embodiment in that the learning control unit 32 determines whether to update an existing dictionary in the dictionary management unit 61 or to add a new dictionary.

  In the following description, the same components as those in the first embodiment are given the same operation, and the description thereof is omitted.

(1) Configuration of Dictionary Learning Device First, the configuration of the dictionary learning device will be described with reference to FIG. FIG. 7 is a schematic configuration diagram of the dictionary learning apparatus according to the present embodiment.

  As illustrated in FIG. 7, the dictionary learning device includes an acoustic feature extraction unit 11, a related information acquisition unit 21, a learning control unit 32, a similarity calculation unit 41, a dictionary update unit 51, a dictionary management unit 62, and a dictionary construction unit 71. .

(1-1) Learning control unit 32
The learning control unit 32 determines whether or not to use the acoustic feature amount obtained from the acoustic feature extraction unit 11 for updating the dictionary managed by the dictionary management unit 61. If there is a dictionary to be updated, the dictionary updating unit 51 is notified of the dictionary to be updated and the obtained acoustic feature amount.

  In this embodiment, the learning control unit 32 determines whether or not a new dictionary should be constructed when there is no dictionary to be updated. When it is determined that a new dictionary is to be constructed, the learning control unit 32 notifies the dictionary construction unit 71 to construct a new dictionary, and if necessary, the obtained acoustic feature amount is stored in the dictionary construction unit 71. Notify

  Details of a method for determining the necessity of updating and a method for determining whether or not to add a new dictionary will be described later.

(1-2) Dictionary construction unit 71
The dictionary construction unit 71 constructs a new dictionary using the acoustic feature amount input from the learning control unit 31. When the new dictionary is the GMM described above, the dictionary construction unit 71 can construct a new dictionary using, for example, the EM algorithm.

  The newly constructed new dictionary is added and managed by the dictionary management unit 62 in association with the related information obtained in synchronization with the acoustic feature used for learning.

  As described above, in this embodiment, when an unknown situation is encountered, a new dictionary is learned from the obtained acoustic feature amount, so that a scene can be identified even for the unknown situation. .

(2) Specific Example The operation of the dictionary learning apparatus according to this embodiment will be specifically described. The dictionary learning device includes an acceleration sensor, a GPS, and a microphone, and considers a case where the dictionary learning device is applied to an information providing device that provides necessary information after grasping a situation where a user is placed in daily life. FIG. 8 shows an example of an information providing apparatus for life support information displayed according to the situation grasped in FIG.

  In the schematic diagram of FIG. 8, the commuting route from home to the office is displayed, and there are a plurality of routes (the same location may be moved by a bus or a vehicle) or similar vehicles at different locations. It is assumed that you take a bus.

  A user commute with an information providing device including an acceleration sensor, a GPS, and a microphone. This information providing apparatus has the dictionary learning apparatus and the scene identification apparatus of the present embodiment. Then, it is assumed that the scene identification device determines the scene and provides information suitable for the situation on the screen of the information providing device. For example, as shown in FIG. 9, when the user approaches the company, the schedule of the day is displayed, and when waiting for the bus, the approach information of the bus is displayed.

  In the above example, it can be grasped that the vehicle is moving toward the company using existing sensors, that is, GPS and acceleration sensors, but it is difficult to grasp different vehicles (buses / cars) through the same route. , Collaboration with microphone is required.

(3) Flow of New Construction Next, the flow of processing of the dictionary learning apparatus when updating a dictionary for scene identification while performing daily life support and updating a new dictionary as needed is shown in FIG. This will be described with reference to the flowchart of FIG. For simplification of processing, the dictionary management unit 61 that manages the dictionary starts the description from the state shown in FIG.

  First, in step S0201, the acoustic feature extraction unit 11 extracts an acoustic feature amount from the input acoustic signal. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0202, the related information acquisition unit 21 acquires related information from the GPS and the acceleration sensor. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0203, when the learning control unit 32 acquires the acoustic feature amount and the related information, the learning control unit 32 narrows down the dictionary that is a candidate to be updated using the related information from the dictionaries managed by the dictionary management unit 62. For example, when it is known from the related information that “currently moving from place C to place D”, the learning control unit 32 selects the dictionary CD corresponding to the train with the same label in the dictionary management unit 62. Then, the dictionary is determined to be updated, and the process proceeds to step S0204. Note that a portion surrounded by a frame in the flowchart of FIG. 11 corresponds to the operation of the learning control unit 32.

  Next, in step S0204, the learning control unit 32 performs collation between the narrowed-down dictionary CD and the input acoustic feature amount, and calculates a similarity score. Since the similarity score calculation method is the same as that of the first embodiment, the description thereof is omitted. Here, it is assumed that the similarity score corresponding to the train is scoreCD (> α).

  Next, in step S0205, the learning control unit 32 first determines the presence / absence of a dictionary to be adapted based on the obtained similarity score and the necessity / unnecessity of constructing a new dictionary. For example, a method can be considered in which the similarity score threshold is set as α and β and the following determination is made.

scoreCD ≧ α... Judged as learned.
α> scoreCD ≧ β: Adapt the dictionary,
scoreCD <β... A new dictionary is constructed based on an unknown situation.

  That is, when scoreCD is larger than the threshold value α, the learning control unit 32 determines that the dictionary CD managed by the dictionary management unit 62 is sufficiently learned and does not need to be adapted, and ends the process.

  If the scoreCD is less than the threshold β, the learning control unit 32 determines that the unknown situation does not match the existing dictionary CD, determines that a new dictionary is to be constructed, and proceeds to step S2107.

  The learning control unit 32 determines that an adaptation target has been found when threshold α> scoreCD ≧ threshold β, and proceeds to step S2106. In this case, since it is determined that the train dictionary CD has already been sufficiently learned from the scoreCD, the processing ends.

  Here, for example, a situation is considered in which there is no time from the railway station B to the company A, and a taxi is used instead of a normal bus.

  In step S0203, the learning control unit 32 knows from the related information that it is currently “moving from location B to location A”, and the learning control unit 32 is given the same label in the dictionary management unit 62. The dictionary BA corresponding to the bus is determined as the dictionary to be updated, and the process proceeds to step S0204.

  In step S0204, the learning control unit 32 performs collation between the narrowed-down dictionary BA and the input acoustic feature amount, and calculates a similarity score. Since the similarity score calculation method is the same as that of the first embodiment, the description thereof is omitted. Here, it is assumed that the similarity score corresponding to the train is scoreAB (<β). If it is determined according to the above-described determination criteria, it can be determined that the current situation is an unknown situation different from the existing dictionary BA. In this case, the process proceeds to step S0207 and operates to construct a new dictionary.

  In step S 0207, the dictionary construction unit 71 operates to construct a new dictionary using the acoustic feature amount input from the learning control unit 31 and add the dictionary to the dictionary management unit 62.

  When the new dictionary is the GMM described above, a new dictionary can be constructed by using an existing algorithm such as the EM algorithm, and a new scene can be identified.

(4) Effect As described above, in this embodiment, when an unknown situation is encountered, the dictionary is operated to learn a new dictionary based on the obtained acoustic feature amount. Can be identified.

(5) Modification Example In the above embodiment, the dictionary construction unit 71 constructs a new dictionary only from the acoustic signals recorded when an unknown situation is detected. However, the present embodiment is not limited to this.

  For example, when the dictionary construction unit 71 constructs a new dictionary, a general dictionary as a seed is prepared in advance, and a general dictionary having a high similarity score with an acoustic signal obtained from an unknown situation is adopted as an initial dictionary. It is also possible to do.

  Specifically, a general-purpose dictionary corresponding to situations that may occur in daily life (bus, train, car, garbage, premises, etc.) is prepared. Then, the dictionary construction unit 71 operates to select from general-purpose dictionaries of this type as the initial dictionary. Such general-purpose dictionaries do not perfectly match the actual environment, but they can be more accurate than constructing from a few acoustic signals, or use labels attached to the initial seed dictionary There is an advantage that can be.

  As a method for selecting an initial dictionary from a general-purpose dictionary as a seed, a similarity score may be calculated in the same manner as in step S0204 for selecting an applicable dictionary, and a dictionary that maximizes the similarity score may be selected.

(Third embodiment)
A dictionary learning apparatus according to a third embodiment of the present invention will be described with reference to FIGS.

  In the present embodiment, a confidence interval estimation unit 81 that can be used for dictionary learning and adaptation, that is, to estimate a reliable interval for identifying a scene is added. That is, the present embodiment is different from the first embodiment in that the learning control unit 33 narrows down the acoustic feature amount section used when updating the dictionary based on the result of the confidence section estimation unit 81.

  In the following description, the same components as those in the first embodiment are given the same operation, and the description thereof is omitted.

(1) Configuration of Dictionary Learning Device First, the configuration of the dictionary learning device will be described with reference to FIG. FIG. 13 is a schematic configuration diagram of the dictionary learning apparatus according to the present embodiment.

  As illustrated in FIG. 13, the dictionary learning device includes an acoustic feature extraction unit 12, a related information acquisition unit 21, a learning control unit 33, a similarity calculation unit 41, a dictionary update unit 51, a dictionary management unit 62, and a confidence interval estimation unit 81. Have

(1-1) Confidence interval estimation unit 81
The confidence interval estimation unit 81 estimates a reliable interval that can be used for dictionary learning or adaptation from the acoustic signal based on the acoustic feature amount obtained from the acoustic feature extraction unit 11, and learns a reliable interval. Notify the controller 33. Details of the reliable section estimation method will be described later.

(1-2) Learning control unit 33
The learning control unit 33 determines whether or not to use the acoustic feature amount obtained from the acoustic feature extraction unit 11 for updating the dictionary managed by the dictionary management unit 62. When there is a dictionary to be updated, the learning control unit 33 notifies the dictionary updating unit 51 of the dictionary to be updated and the obtained acoustic feature amount. At this time, in the present embodiment, based on the information of the confidence interval obtained from the confidence interval estimation unit 81, not all the acoustic feature amounts are notified to the dictionary update unit 51, but only a reliable interval is notified. This is different from the first embodiment.

  As described above, in this embodiment, when updating a dictionary, not all acoustic signals are used for updating, but only a learning section that contributes to performance improvement can be used for dictionary adaptation, thereby improving scene identification performance. be able to.

(2) Specific Example The operation of the dictionary learning device will be specifically described. That is, consider a case where the dictionary learning device is provided with an acceleration sensor, a GPS, and a microphone, and is applied to an information providing device that provides necessary information after grasping a situation where a user is placed in daily life.

  As in the second embodiment, the assumed situation is that when a user commute with an information providing device equipped with an acceleration sensor, a GPS, and a microphone, the scene identification device determines the location and situation, A case is assumed where information suitable for the situation is provided on the screen of the information providing apparatus.

  Here, consider an example in which an acoustic signal as shown in FIG. 15 is obtained when “moving from place A to place B (by bus)”. As shown in FIG. 15, even when traveling on the same bus, when the bus stops due to a signal or the like, or when an acoustic signal different from the running sound of the bus such as an announcement sound or door opening / closing sound is mixed Can be considered.

  If scene identification or dictionary update is performed on all acoustic signals including such mixed sounds, there is a high risk of erroneous identification and adaptation. Therefore, in the present embodiment, the confidence interval is estimated so that the dictionary can be robustly constructed even in such a situation.

(3) Overall Flow The processing flow of the dictionary learning apparatus will be described using the flowchart of FIG. The dictionary learning apparatus estimates a reliable section estimated to contribute to performance improvement while performing daily life support, and updates the dictionary using only the acoustic signal included in the corresponding section.

  For simplification of description, the dictionary management unit 62 that manages the dictionary starts the description from the state shown in FIG.

  First, in step S0301, the acoustic feature extraction unit 12 extracts an acoustic feature amount from the input acoustic signal. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0302, the related information acquisition unit 21 acquires related information from the GPS and the acceleration sensor. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0303, the confidence interval estimation unit 81 compares the acoustic feature values of a plurality of frames to obtain a change amount of the acoustic feature value, and estimates a zone with a small change in acoustic feature value as a reliable zone. FIG. 16 shows an outline of an algorithm for estimating the confidence interval based on the change amount of the acoustic feature amount. As shown in FIG. 16, the acoustic change amount is calculated by comparing the acoustic feature amount extracted in units of frames between a plurality of frames. As a method for calculating the amount of change, for example, a norm that is an absolute value of a difference between feature vectors can be used. By setting a threshold value for determining that the change amount is large and unreliable with respect to the change amount, it is possible to determine that a section of the acoustic feature amount that does not exceed the threshold value is a reliable section.

  And the learning control part 33 will judge whether the acoustic feature-value to process now is produced | generated from the confidence area, if an acoustic feature-value, related information, and the information of a confidence area are acquired. If it is not included in the confidence interval, the adaptive processing is not performed. For example, in a section where a door opening / closing sound is mixed in a stop section of a bus, it is determined that a section with a small amount of acoustic change is short and unreliable, and thus the processing ends without determining the necessity of adaptation. If the confidence interval estimation unit 81 determines that the interval is reliable, the process proceeds to step S0304. 14 corresponds to the operation of the learning control unit 33.

  Next, in step S0304, the learning control unit 33 narrows down the dictionary that is a candidate for updating using the related information from the dictionaries managed by the dictionary management unit 62. For example, when it is known from the related information that “currently moving from place B to place A”, the learning control unit 33 uses the dictionary management unit 62 to provide a dictionary corresponding to the bus and car with the same label. Is determined as the dictionary to be updated, and the process proceeds to step S0305.

  Next, in step S0305, the learning control unit 33 compares the two narrowed dictionaries with the input acoustic feature amount to calculate a similarity score.

  Next, in step S0306, the learning control unit 33 determines the presence / absence of a dictionary to be adapted based on the obtained similarity score. The method for calculating the similarity score and the method for determining the necessity of adaptation are the same as in the first embodiment, and are omitted. However, if adaptation is necessary, the dictionary is adapted to be adapted in step S0307.

(4) Effect According to the present embodiment, among the acoustic signals acquired from the microphone, a reliable section suitable for learning is determined as a reliable section and control is performed so that only a reliable section is used for learning. Yes.

  Therefore, according to the present embodiment, only the learning section that contributes to the performance improvement can be used for adaptation of the dictionary, and the scene identification performance can be improved.

(5) Modification Example In the above embodiment. The simplest confidence interval estimation method was explained. However, the present invention is not limited to this, and instead of the norm of two feature vectors, for example, the accumulated value of norms between a plurality of feature vector groups is used as an acoustic change amount, or it is removed when a reliable interval is a predetermined time or less. As a method for estimating a reliable section from an acoustic signal, various modifications can be considered.

  Further, not only the comparison of input acoustic feature amounts but also the features of unreliable acoustic signals are learned in advance and stored as a dictionary (hereinafter referred to as an inapplicable dictionary), and the similarity score with the inapplicable dictionary is If it is high, there is a possibility that the confidence interval is not used. For example, since the voices of people in the surroundings are often mixed with environmental sounds in daily life, a speech dictionary learned from human voices is prepared, and the similarity score with the speech dictionary is a threshold value. It can be considered that the above section is not used for dictionary construction because it is judged that human voice is mixed.

(Fourth embodiment)
A dictionary learning apparatus according to a fourth embodiment of the present invention will be described with reference to FIGS.

  In the present embodiment, a confidence interval estimation unit 82 for estimating a reliable interval that can be used for dictionary learning and adaptation is added. In the present embodiment, the configuration in which the learning control unit 33 narrows down the acoustic feature amount section used when updating the dictionary based on the result of the confidence section estimation unit 82 is different from the first embodiment.

  In the following description, the same components as those in the first embodiment are given the same operation, and the description thereof is omitted.

(1) Configuration of Dictionary Learning Device First, the configuration of the dictionary learning device will be described with reference to FIG. FIG. 17 is a schematic configuration diagram of the dictionary learning device according to the present embodiment.

  As illustrated in FIG. 17, the dictionary learning device includes an acoustic feature extraction unit 11, a related information acquisition unit 22, a learning control unit 33, a similarity calculation unit 41, a dictionary update unit 51, a dictionary management unit 62, and a confidence interval estimation unit 82. have.

(1-1) Confidence interval estimation unit 82
The confidence interval estimation unit 82 estimates a reliable interval that can be used for learning and adaptation of the dictionary from the acoustic signal based on the related information obtained from the related information acquisition unit 21, and performs learning control on the reliable interval. Notify unit 33. Details of the reliable section estimation method will be described later.

(1-2) Learning control unit 33
The learning control unit 33 determines whether or not the acoustic feature amount obtained from the acoustic feature extraction unit 11 is used for updating the dictionary managed by the dictionary management unit 62. When there is a dictionary to be updated, the learning control unit 33 notifies the dictionary updating unit 51 of the dictionary to be updated and the obtained acoustic feature amount. At this time, in this embodiment, based on the information of the confidence interval obtained from the confidence interval estimation unit 82, not all of the acoustic feature quantities are notified to the dictionary update unit 51, but only the reliable interval is notified. This is different from the first embodiment.

  As described above, in this embodiment, when updating a dictionary, not all acoustic signals are used for updating, but only a learning section that contributes to performance improvement can be used for dictionary adaptation, thereby improving scene identification performance. be able to.

(2) Specific Example The operation of the dictionary learning apparatus according to this embodiment will be specifically described.

  The dictionary learning device includes an acceleration sensor, a GPS, and a microphone, and considers a case where the dictionary learning device is applied to an information providing device that provides necessary information after grasping a situation where a user is placed in daily life.

  As in the second embodiment, the assumed situation is that when a user commute with an information providing device equipped with an acceleration sensor, a GPS, and a microphone, the scene identification device determines the location and situation, It is assumed that information suitable for the situation is provided on the screen of the information providing apparatus.

  Here, consider an example in which an acoustic signal as shown in FIG. 15 is obtained when “moving from place A to place B (by bus)”.

  As shown in FIG. 15, even when traveling on the same bus, when the bus stops due to a signal or the like, or when an acoustic signal different from the running sound of the bus such as an announcement sound or door opening / closing sound is mixed Can be considered. Such mixing of different acoustic signals may occur in conjunction with movement conditions. For example, when determining the type of vehicle from an acoustic signal, an acoustic signal recorded in a state of moving more stably than in a stationary state is useful. Based on this, in this embodiment, the confidence interval is estimated based on the change in the related information.

(3) Overall Flow The processing flow of the dictionary learning apparatus will be described using the flowchart of FIG.

  The dictionary learning apparatus estimates a reliable section estimated to contribute to performance improvement while performing daily life support, and updates the dictionary using only the acoustic signal included in the corresponding section. For simplification of description, the dictionary management unit 62 that manages the dictionary starts the description from the state shown in FIG.

  First, in step S0401, the acoustic feature extraction unit 11 extracts an acoustic feature amount from the input acoustic signal. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0402, the related information acquisition unit 22 acquires related information from the GPS and the acceleration sensor. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0403, the confidence interval estimation unit 82 estimates the confidence interval based on the related information. Considering the example of FIG. 18, for example, by setting a section in which the same movement state continues for a certain time or more as a confidence section, a bus stop time or the like having a short duration is not a confidence section, and only the travel time is learned. 31 can be processed. If the reliable interval estimation unit 82 determines that the interval is reliable, the process advances to step S0404.

  Next, in step S0404, the learning control unit 33 narrows down the dictionary that is a candidate for updating using the related information from the dictionaries managed by the dictionary management unit 62. For example, when it is known from the related information that “currently moving from place B to place A”, the learning control unit 33 uses the dictionary management unit 62 to provide a dictionary corresponding to the bus and car with the same label. Is determined as a dictionary to be updated, and the process advances to step S0405.

  Next, in step S0405, the learning control unit 33 compares the two narrowed dictionaries with the input acoustic feature amount to calculate a similarity score.

  Next, in step S0406, the learning control unit 33 further determines the presence or absence of a dictionary to be adapted based on the obtained similarity score. Since the method for calculating the similarity score and the method for determining the necessity for adaptation are the same as those in the first embodiment, a description thereof is omitted.

  Next, in step S0407, the dictionary update unit 51 updates the adapted dictionary.

(4) Effect According to the present embodiment, among the acoustic signals acquired from the microphone, a reliable section suitable for learning is determined as a reliable section and only a reliable section is used for learning.

  Therefore, according to the present embodiment, only the confidence interval that contributes to the performance improvement can be used for adaptation of the dictionary, and the scene identification performance can be improved.

(Fifth embodiment)
A dictionary learning apparatus according to a fifth embodiment of the present invention will be described with reference to FIGS. 2 to 4, 19, and 20.

  In this embodiment, an abnormality notification unit 91 is added for notifying an abnormal state when an unknown state is detected. In the present embodiment, the configuration in which the learning control unit 34 determines whether an existing dictionary in the dictionary management unit 62 is updated or not is different from the first embodiment.

  In the following description, the same components as those in the first embodiment are given the same operation, and the description thereof is omitted.

(1) Configuration of Dictionary Learning Device First, the configuration of the dictionary learning device will be described with reference to FIG. FIG. 19 is a schematic configuration diagram of the dictionary learning apparatus according to the present embodiment.

  As illustrated in FIG. 19, the dictionary learning device includes an acoustic feature extraction unit 11, a related information acquisition unit 21, a learning control unit 34, a similarity calculation unit 41, a dictionary update unit 51, a dictionary management unit 62, and an abnormality notification unit 91. Have.

(1-1) Learning control unit 34
The learning control unit 34 determines whether or not the acoustic feature amount obtained from the acoustic feature extraction unit 11 is used for updating the dictionary managed by the dictionary management unit 62, and when there is a dictionary to be updated, The dictionary update unit 51 is notified of the dictionary to be updated and the obtained acoustic feature amount.

  Further, in this embodiment, the learning control unit 34 determines whether or not it is in an unknown state when there is no dictionary to be updated. If it is determined that the state is unknown, the abnormality notification unit 91 is notified that the state is abnormal. Details of the method for determining the necessity of updating and the method for determining an abnormal state will be described later.

(1-2) Abnormal state notification unit 91
The abnormal state notification unit 91 notifies that the current state is an abnormal state based on the determination result of the learning control unit 31. As a notification method, for example, various methods such as a method of displaying a warning on the screen of the information providing apparatus and a sound of a warning sound are conceivable.

  As described above, according to the present embodiment, when an unknown situation is encountered, the user operates to notify the user of the unknown state, so that the user can easily grasp the abnormal state.

(2) Specific Example The operation of the dictionary learning apparatus will be described with reference to the first embodiment.

  That is, also in this specific example, the information providing apparatus has the dictionary learning apparatus and the scene identification apparatus of the present embodiment. The scene identification device identifies a scene in the factory based on the dictionary updated by the dictionary learning device of this embodiment.

  When a security guard moves within the site with an information providing device equipped with an acceleration sensor, GPS, and microphone, the identification unit of the scene identification device is updated to the dictionary updated by the dictionary learning device of the present embodiment. Identify scenes in the factory based on. And the provision part of an information provision apparatus provides the information which a guard should confirm about the machine in front of the place where the guard is patroling corresponding to the identified scene.

(3) Flow of abnormality detection The flow of processing of the dictionary learning apparatus will be described with reference to the flowchart of FIG.

  The dictionary learning device updates a dictionary for scene identification while performing a factory tour, and notifies an abnormality when an unknown state is encountered. As shown in FIG. 4, in the dictionary management unit 62, a dictionary is prepared for each machine to be visited, and this dictionary is classified and held for each location and movement situation that can be acquired from related information.

  First, in step S0501, the acoustic feature extraction unit 11 extracts an acoustic feature amount from the input acoustic signal. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0502, the related information acquisition unit 21 acquires related information from the GPS and the acceleration sensor. Since this process is the same as that of the first embodiment, a description thereof will be omitted.

  Next, in step S0503, when the learning control unit 34 acquires the acoustic feature amount and the related information, the learning control unit 34 narrows down the dictionary that is a candidate to be updated using the related information from the dictionaries managed by the dictionary management unit 62. For example, when it is known from the related information that the user is currently “still at place A”, the learning control unit 34 corresponds to the machine 1, the machine 2, and the machine 3 to which the same label is assigned in the dictionary management unit 62. The dictionary to be updated is determined as the dictionary to be updated, and the process proceeds to step S0504. A portion surrounded by a frame in the flowchart of FIG. 20 corresponds to the operation of the learning control unit 34.

  Next, in step S0504, the learning control unit 34 compares the three narrowed dictionaries with the input acoustic feature amount, and calculates a similarity score. Since the similarity score calculation method is the same as that of the first embodiment, the description thereof is omitted. Here, it is assumed that the similarity score of the machine 1 having the largest value is score1 (<β).

  In step S0505, the learning control unit 34 first determines the presence / absence of a dictionary to be applied and the presence / absence of an abnormality based on the obtained similarity score. For example, a method can be considered in which the threshold value of the similarity score is set as α and β and the following determination is made.

score1 ≧ α... Judged as learned.
α> score1 ≧ β ... Adapt the dictionary,
score1 <β... An unknown state is determined (abnormal state).

  That is, when the score 1 is larger than the threshold value α, the learning control unit 34 determines that the dictionary managed by the dictionary management unit 62 is sufficiently learned and does not need to be adapted, and ends the process.

  If the score 1 is less than the threshold β, the learning control unit 34 determines that the situation is an unknown situation that does not match the existing dictionary 1, and proceeds to step S5107.

  Furthermore, when threshold value α> score1 ≧ threshold value β, learning control unit 34 determines that an adaptation target has been found, and proceeds to step S5106.

  Next, in step S0206, the dictionary update unit 51 updates the adapted dictionary.

  Next, in step S0207, since the current situation can be determined to be an unknown situation different from that of the existing dictionary, the abnormal state notification unit 91 notifies the abnormal state.

(4) Effect As described above, according to the present embodiment, the user operates to notify the user of an unknown state, so that the user can easily grasp the abnormal state.

(Example of change)
The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the gist thereof.

  For example, in each of the above-described embodiments, in order to efficiently generate related information, an example in which GPS and an acceleration sensor are linked is shown. However, the related information acquisition unit 21 is not limited to this, and a plurality of sensors may be operated in parallel.

  In addition to the GPS and the acceleration sensor, related information may be acquired from a clock, an illuminometer, an electrostatic sensor, a temperature sensor, and a humidity sensor.

It is a figure showing the schematic structural example of the dictionary learning apparatus of 1st Example of this invention. It is a figure showing an example of the patrol place in the factory patrol assistance application of 1st Example. It is a figure which shows an example of the information provided with the factory patrol assistance application of a 1st Example. It is a figure showing an example of the dictionary which the dictionary management part of a 1st Example manages. It is a flowchart which shows the flow of the dictionary learning of a 1st Example. It is a flowchart which shows operation | movement of the related information acquisition part of a 1st Example. It is a figure showing the example of schematic structure of the dictionary learning apparatus of 2nd Example. It is a figure showing an example of the condition which wants to grasp | ascertain with the daily life support application of 2nd Example. It is a figure which shows an example of the information provided with the daily life assistance application of a 2nd Example. It is a figure showing an example of the dictionary before the update which the dictionary management part of a 2nd Example manages. It is a flowchart which shows the flow of the dictionary learning of a 2nd Example. It is a figure showing an example of the dictionary after the update which the dictionary management part of a 2nd Example manages. It is a figure showing the example of schematic structure of the dictionary learning apparatus of 3rd Example. It is a flowchart which shows the flow of the dictionary learning of a 3rd Example. It is a figure which shows an example of the acoustic signal which requires the reliability estimation of a 3rd Example. It is the figure which showed the outline | summary of the algorithm of the confidence interval estimation part of the 3rd Example. It is a figure showing the example of schematic structure of the dictionary learning apparatus of a 4th Example. It is a flowchart which shows the flow of the dictionary learning of a 4th Example. It is a figure showing the example of schematic structure of the dictionary learning apparatus of a 5th Example. It is a flowchart which shows the flow of the dictionary learning of a 5th Example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Acoustic feature extraction part 21 ... Related information acquisition part 31 ... Learning control part 41 ... Similarity calculation part 51 ... Dictionary update part 61 ... Dictionary management part

Claims (15)

An acoustic feature extraction unit that analyzes an input acoustic signal and extracts an acoustic feature quantity to be identified for identifying a scene; and
A related information acquisition unit that acquires information other than the input acoustic signal as related information;
A dictionary management unit that manages a plurality of dictionaries of acoustic features prepared in advance in association with the related information;
A similarity calculation section for calculating the degree of similarity respectively by matching the acoustic features and the plurality of dictionaries of the identification target,
The narrowing to a plurality of the dictionaries the acquired related information is shared among the dictionary, obtains the similarity of the narrowed down dictionary from the similarity calculating unit, when said similarity is larger than the first threshold value A learning control unit that controls to update the refined dictionary;
When the dictionary is updated and the dictionary is updated with the similarity greater than the first threshold , an acoustic signal parameter for obtaining the similarity is updated using the acoustic feature quantity of the identification target. A dictionary update unit;
A dictionary learning device comprising: Using the acoustic feature amount of the identification target or the related information, further comprising a confidence interval estimation unit for estimating a reliable interval for identifying the scene;
The learning control unit performs control using only the acoustic feature amount of the identification target included in the estimated confidence interval.
The dictionary learning device according to claim 1. The dictionary is constructed as a mixed normal distribution of the acoustic features prepared in advance .
The similarity calculation unit calculates a posterior probability of the likelihood from the mixed normal distribution as the similarity.
The dictionary learning device according to claim 1. The learning control unit controls not to update the dictionary when the similarity is greater than a second threshold;
The dictionary learning device according to claim 1. In addition to the dictionary managed by the dictionary management unit, further comprising a dictionary construction unit to construct a new dictionary,
The learning control unit constructs the new dictionary or updates the dictionary using the similarity.
The dictionary learning device according to claim 1. The dictionary construction unit
A plurality of general-purpose dictionaries for selecting as an initial dictionary of the new dictionary are held in advance,
The general dictionary having the highest similarity with the acoustic feature quantity to be identified is set as an initial dictionary of the new dictionary.
The dictionary learning apparatus according to claim 5. It further includes an abnormality notification unit for notifying an abnormal state,
The learning control unit controls whether to update the dictionary or to notify the abnormal state based on the similarity.
The dictionary learning device according to claim 1. The related information acquisition unit acquires the related information from any of a GPS, an acceleration sensor, an electrostatic sensor, a temperature sensor, a humidity sensor, or a watch.
The dictionary learning device according to claim 1. The confidence interval estimation unit includes:
Calculating an acoustic variation score representing local variation of the acoustic signal from the acoustic feature quantity of the identification target ;
A section in which the acoustic fluctuation score is a certain value or less is determined as the confidence section.
The dictionary learning device according to claim 2. The confidence interval estimation unit includes:
A non-applicable dictionary for detecting a section of the acoustic signal that is not suitable for the determination;
Calculating the similarity between the acoustic feature quantity to be identified and the non-applicable dictionary;
Only the section where the similarity is smaller than a third threshold is calculated as the confidence section.
The dictionary learning device according to claim 2. The confidence interval estimation unit includes:
Calculating an interval in which the related information does not change as the confidence interval;
The dictionary learning device according to claim 2. The dictionary learning device according to claim 1;
An identification unit for identifying the scene using the dictionary of the dictionary learning device;
A scene identification device comprising: The dictionary learning device according to claim 1;
An identification unit for identifying the scene using the dictionary of the dictionary learning device;
A providing unit for providing information corresponding to the identified scene;
An information providing apparatus comprising: An acoustic feature extraction step of analyzing the input acoustic signal and extracting an acoustic feature quantity to be identified for identifying a scene;
A related information acquisition step of acquiring information other than the input acoustic signal as related information;
A dictionary management step for managing a plurality of previously prepared dictionaries of acoustic features in association with the related information;
A similarity calculation step of calculating the similarity each collating the acoustic features and the plurality of dictionaries of the identification target,
The narrowing to a plurality of the dictionaries the acquired related information is shared among the dictionary, obtains the similarity of the narrowed down dictionary from the results of the similarity calculation step, the degree of similarity is larger than the first threshold value A learning control step for controlling to update the narrowed-down dictionary,
When the dictionary is updated and the dictionary is updated with the similarity greater than the first threshold , an acoustic signal parameter for obtaining the similarity is updated using the acoustic feature quantity of the identification target. A dictionary update step;
A dictionary learning method characterized by comprising: On the computer,
An acoustic feature extraction function for analyzing an input acoustic signal and extracting an acoustic feature quantity to be identified for identifying a scene;
A related information acquisition function for acquiring information other than the input acoustic signal as related information;
A dictionary management function for managing a plurality of dictionaries of acoustic features prepared in advance in association with the related information;
A similarity calculation function of calculating a similarity each collating the acoustic features and the plurality of dictionaries of the identification target,
The narrowing to a plurality of the dictionaries the acquired related information is shared among the dictionary, obtains the similarity of the narrowed down dictionary from the results of the similarity calculation function, the similarity is larger than the first threshold value A learning control function that controls to update the narrowed-down dictionary,
When the dictionary is updated and the dictionary is updated with the similarity greater than the first threshold , an acoustic signal parameter for obtaining the similarity is updated using the acoustic feature quantity of the identification target. Dictionary update function,
Dictionary learning program for realizing.