JP2008145988A - Noise detecting device and noise detecting method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To discriminate the kind (sound source) of a noise. <P>SOLUTION: A final discriminator for discriminating binary values indicating whether data of a noise-superposed speech having a noise superposed in a speech section is a noise generated by a predetermined sound source is held for each predetermined sound source, the input data of the noise-superposed speech is discriminated by using held final discriminators by predetermined sound sources, and a final discriminator having the highest score of one of the binary values is decided according to results of the discrimination to detect the sound source of the noise in the data being the predetermined sound source that the decided final discriminator indicates. Further, a plurality of data including the data of the noise-superposed speech are held as data for learning, and boosting is used to derive final discriminators by the predetermined sound sources using the held data for learning. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a noise detection device and a noise detection method.

  Conventionally, when speech recognition technology is used, misrecognition is often caused by noise superimposed on speech. Focusing on this fact, many studies have been made to remove noise including spectral subtraction. Here, the noise removal will be described in detail. Noise removal is performed by first estimating the noise and then subtracting the estimated noise from the noise-superimposed speech (sound on which the noise is superimposed). Done. There are two methods for estimating noise, such as estimating noise from a noise-only section (non-speech section) immediately before utterance, and estimating noise by stochastically tracking information obtained from a noise-only section. Often used. For example, noise estimation based on minimum statistics is used (V. Stahl, A. Fischer, and R. Bippus, “Quantile based noise estimation for spectral subtruction and Wiener filtering”, Proc.ICASSP 2000, pp.1875- 1878, May 2000).

  By the way, noise estimation as a pre-stage of noise removal is considered to be a very effective technique for stationary noise and noise that changes slowly in time, and noise removal (suppression). It is expected that a high effect can be obtained. However, considering that voice recognition technology is used in a real environment such as in a house, some noises that occur suddenly during speech, such as phone call sounds (sudden noise), are also included. Not a few. For example, FIG. 18 shows a waveform in which a telephone call sound is superimposed in the voice. Thus, when noise suddenly occurs during speech, the speech recognition rate decreases even if the noise lasts only for a short time.

  For this reason, an effective method for such a sudden noise must be studied, but it is usually difficult to estimate the sudden noise using the method as described above. Also, HMM synthesis methods (Kazuhiro Miki, Takanobu Nishiura, Satoshi Nakamura, Kiyohiro Shikano, “Examination of environmental sound discrimination using HMM”, IEICE Speech Society, SP99-106, pp. 79-84 (1999) -12), Masaki Ida, Satoshi Nakamura, “Model-adapted HMM using noise DB for speech recognition under noise by multipath model with SN ratio”, IEICE Technical Report, Vol.101, No.522, pp .51-56,2001-12) it is considered to use, the use of the technique of HMM composition, unless Failure to identify whether advance what noise is superimposed on the voice, the number of combinations As a result, it takes time for speech recognition, so it cannot be said to be an appropriate method.

  For this reason, when it is intended to deal with sudden noise, it is considered that it is desirable not to estimate noise as a pre-stage of noise removal but to use a procedure for detecting noise. In order to detect this noise, it is conceivable to use a technique for detecting noise by examining the power of speech, a technique for detecting noise by AdaBoost (Adaboost), or the like. However, the method of detecting noise by examining the power of speech can detect to some extent if the SNR (Signal vs. Noise Ratio) is extremely bad as shown in the waveform of FIG. 19, when three types of noise (“spray sound”, “sound that breaks paper”, “phone call sound”) with an SNR of 5 dB are superimposed on the voice section. (“Spray sound” and “phone call sound” are completely superimposed on the voice section), and it is almost impossible to detect them.

  On the other hand, a technique for detecting noise by AdaBoost will be described. AdaBoost is a powerful technique for the binary discrimination problem, and is one of techniques called Boosting. Here, Boosting is a method of generating a final discriminator by weighted majority of a plurality of weak discriminators having low discrimination performance and outputting a discrimination result by the final discriminator. AdaBoost is often used as a method for detecting an object such as a face from image information because of its high accuracy and high speed (Paul Viola and Michael Jones: “Rapid Object Detection using a Boosted Cascade of Simple Features”. IEEECVPR, vol.1, pp.511-518, 2001.). Further, Non-Patent Document 1 and Non-Patent Document 2 disclose a technique for detecting a speech section that does not include noise using AdaBoost.

Kwon, O., Lee, T .: “Optimizing speech / non-speech classifier design using adaboost” Proc. IEEE ICASSP 2003, pp I-436-I-439.pp. Apr. 2003 Hiroyoshi Matsuda, Tetsuya Takiguchi, Yasuo Ariki: “Speech detection by Real Adaboost”, Acoustical Society of Japan 2006 Autumn Meeting, 2-P-12, PP.117-118, 2006-09.

  By the way, the above-described conventional technique has a problem that the type of noise (sound source) cannot be identified as described below. That is, in the method of detecting noise by AdaBoost, for example, the discriminator discriminates the binary value of “noise” or “not noise”, so that the type of noise (sound source) can be identified. Can not.

  Accordingly, the present invention has been made to solve the above-described problems of the prior art, and an object thereof is to provide a noise detection device and a noise detection method capable of identifying the type of noise (sound source). To do.

  In order to solve the above-described problems and achieve the object, the invention according to claim 1 is a binary signal indicating whether or not noise superimposed speech data in which noise is superimposed on a speech section is noise from a predetermined sound source. A final discriminator holding unit for holding a final discriminator for each predetermined sound source, and a final discriminator for each predetermined sound source in which the data of the input noise superimposed speech is held by the final discriminator holding unit The sound source of the noise present in the data is determined by determining the final classifier having the highest score identified as one of the binary values as a result of the identification. And detecting means for detecting that the sound source is a predetermined sound source indicated by the final discriminator.

  According to a second aspect of the present invention, in the above invention, learning data holding means for holding a plurality of data including noise superimposed voice data as learning data, and whether the data is noise caused by a predetermined sound source. The learning data held by the learning data holding means is predetermined by using boosting for deriving a final discriminator that has finished learning by learning a discriminator for discriminating the binary from the learning data. And a final classifier deriving unit for deriving a final classifier for each sound source.

  The invention according to claim 3 is characterized in that, in the above invention, the final discriminator deriving means derives the final discriminator using Adaboost as the boosting.

  According to a fourth aspect of the present invention, in the above invention, the detection means identifies the noise-superimposed speech data in units of frames, and the noise section of the data is a section divided by the frame. Is further detected.

  According to a fifth aspect of the present invention, in the above invention, the identification result identified by the final discriminator determined by the detection means is another frame in the continuous frames of the input data. When a frame with a result different from that is included, smoothing means for smoothing the frame with a different result is further provided.

  The invention according to claim 6 provides a final discriminator for each predetermined sound source for identifying a binary value indicating whether or not noise superimposed speech data in which noise is superimposed on a speech section is noise caused by a predetermined sound source. A final discriminator holding step that is held in the discriminator, and the input noise superimposed speech data is discriminated using each final discriminator for each of the predetermined sound sources held in the final discriminator holding step, and the result of the discrimination By determining the final discriminator having the highest score identified by any one of the two values, the noise source present in the data is a predetermined sound source indicated by the determined final discriminator. And a detecting step for detecting this.

  According to the first or sixth aspect of the present invention, the final discriminator for identifying the binary value of whether or not the noise-superimposed speech data in which noise is superimposed on the speech section is noise due to the predetermined sound source is provided as the predetermined sound source. Each of the stored noise superposed speech data is identified using each final discriminator for each predetermined sound source, and as a result of identification, the score identified as one of the two values By determining the final discriminator having the highest noise level, it is detected that the noise source existing in the data is the predetermined sound source indicated by the determined final discriminator. It becomes possible.

  According to the invention of claim 2, the discriminator that holds a plurality of data including noise superimposed voice data as learning data and discriminates whether or not the data is noise caused by a predetermined sound source. The type of noise (sound source) because the final classifier for each given sound source is derived from the stored learning data using boosting that derives the final classifier that has been learned by learning from the training data Can be properly identified.

  According to the invention of claim 3, since the noise detection device derives the final discriminator using Adaboost as boosting, it becomes possible to appropriately identify the type of noise (sound source).

  According to the invention of claim 4, the noise detection device further identifies that the noise-superimposed speech data is identified in units of frames and further detects that the noise section of the data is a section divided by frames. In addition to the above effect, it is also possible to detect a noise interval.

  According to the invention of claim 5, the noise detection apparatus is configured such that the identification result identified by the final classifier determined by the detection means is different from the other frames in the continuous frames of the input data. When frames with different results are included, smoothing is performed on the frames with different results, so that the noise type (sound source) can be accurately identified.

  Exemplary embodiments of a noise detection device and a noise detection method according to the present invention will be described below in detail with reference to the accompanying drawings. In the following, the main terms used in the embodiment, the outline and features of the noise detection device according to the first embodiment, the configuration and processing procedure of the noise detection device according to the first embodiment, and the effects of the first embodiment will be described in order. Next, another embodiment will be described.

[Explanation of terms]
First, main terms used in the following examples will be described. “Noise” used in the following embodiments refers to “sound” that is different from “speech” to be recognized when using the speech recognition technology, and usually impedes the recognition of “speech” to be recognized. It is the “sound” that is considered to be. In the following, the “speech” to be recognized is broadly classified into two types, “speech segment” in which “speech” to be recognized exists and “non-speech segment” in which “speech” to be recognized does not exist. However, “noise” is superimposed on the “speech segment” (“speech” in which “speech” and “noise” to be recognized are superimposed), and only “noise” is included in the “non-speech segment”. Existence is defined as “noise superimposed speech”.

  By the way, if “noise” is included in the “speech” to be recognized, the speech recognition rate decreases. Therefore, speech recognition should be performed after removing (suppressing) “noise”. As a pre-stage of removal (suppression), it is necessary to detect “noise”. Moreover, it is desirable that the “noise” is detected after identifying the type (sound source) of “noise”.

  Here, the “sound source” will be specifically described. For example, “noise” includes “spray sound” (for example, “swoosh”), “sound to break paper” (for example, “buzzy” sound) Etc.) and “phone call sound” (for example, “Purururu” sound, etc.) are considered to be various kinds of “sound sources”. Since the difference between these “sound sources” appears as a waveform difference as shown in FIG. 19, identifying and detecting “sound source” when detecting “noise” eliminates “noise”. (Repression) will also help. In other words, by identifying the “sound source” of “noise” (by knowing what “noise” is mixed in “speech”), each noise source is pre-determined for noise reduction (suppression). The “noise” can be removed (suppressed) using the stored “noise” data. For this reason, it is important how the “noise detection apparatus” according to the present invention identifies the “sound source” of “noise” by any method.

[Outline and Features of Noise Detection Device According to Embodiment 1]
Next, the outline and features of the noise detection apparatus according to the first embodiment will be described with reference to FIG. FIG. 1 is a diagram for explaining the outline and features of the noise detection apparatus according to the first embodiment.

  As described above, the noise detection apparatus according to the first embodiment is mainly characterized by detecting “noise” from “speech” to be recognized, and mainly characterized by identifying the type of noise (sound source).

  This main feature will be briefly described. The noise detection apparatus according to the first embodiment holds data including noise superimposed speech as learning data in a learning data holding unit.

  For example, as shown in FIG. 1 (1), the noise detection apparatus according to the first embodiment holds one group composed of a plurality of data as learning data in the learning data holding unit. For each group, a label is assigned to each sound source such as “voice only”, “spray sound”, “paper breaking sound”, and “phone call sound”.

  These learning data are data for the noise detection apparatus according to the first embodiment to derive a final discriminator (a discriminator used for noise detection), and thus may be input in advance by a user of the noise detection apparatus. By doing so, the noise detection apparatus holds in advance.

  With such a configuration, the noise detection apparatus according to the first embodiment derives a final discriminator for each sound source using boosting. Here, boosting is an algorithm for deriving a final discriminator that has completed learning by causing a discriminator for discriminating binary values of whether or not the data is a predetermined sound source to be learned from learning data. That is. In the first embodiment, the final discriminator for each sound source is derived using AdaBoost (Adaboost) as boosting. The AdaBoost algorithm will be described in detail when the configuration of the noise detection apparatus is described.

  For example, as shown in (2) of FIG. 1, the noise detection device uses “AdaBoost” to obtain a “voice-only discriminator” (a discriminator that discriminates whether or not it is “voice only”) from learning data. By learning, the “speech-only final discriminator” is derived and held. Similarly, as shown in (2) of FIG. 1, the noise detection device uses “AdaBoost” to obtain “spray sound discriminator” (discriminator for discriminating whether it is “spray sound”) as learning data. By learning from, the "spray sound final discriminator" is derived and retained, and the "paper breaking sound discriminator" (discriminator discriminating whether or not it is a "paper breaking sound") is learned. By learning from the data, the “final discriminator of the sound that breaks the paper” is derived and retained, and the discriminator that discriminates whether or not it is a “telephone call tone discriminator” (“phone call tone”) ) Is learned from the learning data, so that the “final classifier of the telephone call tone” is derived and held.

  Next, the noise detection apparatus according to the first embodiment identifies input noise-superimposed speech data in units of data frames by a final discriminator for each stored sound source. Here, the data frame is a group of data divided by a certain time (for example, “20 ms”) and indicates a data section.

  For example, as shown in (3) of FIG. 1, the noise detection device converts the input data into “final discriminator of only sound”, “final discriminator of spray sound”, and “final discrimination of sound that breaks paper”. Using a final discriminator such as “final discriminator” or “final discriminator of telephone call tone”, the section of “frame No. 100” of data is identified. Then, for example, as shown in (3) of FIG. 1, the “frame No. 100” of the data is changed to any one of the binary values as a result of identification using the “sound-only final discriminator”. It is identified with a score of “−0.3”. Similarly, as a result of the identification using the “spray sound final discriminator”, one of the two values is identified with a score of “−0.1”, and “the final discriminator of the sound that breaks the paper” As a result of identification using, one of the two values is identified with a score of “−0.2”, and as a result of identification using the “final discriminator of phone call tone” Any one of the values (here, “phone call tone”) is identified with a score of “0.5”.

  Subsequently, the noise detection apparatus according to the first embodiment determines the final discriminator having the highest score identified by any one of the two values. For example, as shown in (4) of FIG. 1, the noise detection device is configured such that the final discriminator identified with the highest score of “0.5” as one of the two values is “phone call sound. It is determined that it is a final discriminator.

  Then, the noise detection device detects that the noise section of the data is a section divided by this frame, and that the noise source included in the data is the sound source indicated by the determined final discriminator. For example, as shown in (5) of FIG. 1, the noise detection apparatus is a section in which the noise section of the data is divided by “frame No. 100” and the noise source in the data is “telephone”. Is detected.

  For this reason, according to the noise detection apparatus according to the first embodiment, it is possible to identify the type of noise (sound source).

  In the first embodiment, the noise detection apparatus holds the learning data in advance, derives and holds the final discriminator from the learning data, and then inputs the noise superimposed speech input using the held final discriminator. Although the example of performing the data detection process has been described, the present invention is not limited to this. For example, the processing for holding the learning data and deriving the final discriminator from the learning data is performed by another device different from the noise detection device, and the noise detection device according to the present invention is derived by another device. The present invention can be similarly applied to a case where the final discriminator is held and the detection processing of the noise superimposed speech data input using the held final discriminator is performed.

  Also, in the first embodiment, a method is described in which the noise detection device identifies input noise-superimposed speech data in units of frames and detects that the data noise section is a section divided by frames. However, the present invention is not limited to this, and the present invention can be applied in the same manner even when the noise detection device does not rely on a method for identifying input noise superimposed speech data in units of frames. it can.

  In the first embodiment, the noise detection device holds one group, and for this one group, “voice only”, “spray sound”, “paper breaking sound”, “phone call sound” However, the present invention is not limited to this example. For example, when another sound source is selected as the sound source, or when a plurality of groups are held in duplicate instead of giving a label for each sound source to one group (a label is assigned to each sound source). The present invention can be similarly applied. That is, the sound source may be appropriately selected according to the environment where the noise detection apparatus according to the present invention is used, and any form of holding may be used.

[Configuration of Noise Detection Apparatus According to First Embodiment]
Next, the noise detection apparatus according to the first embodiment will be described with reference to FIGS. FIG. 2 is a block diagram illustrating the configuration of the noise detection apparatus according to the first embodiment, FIG. 3 is a diagram for explaining the output unit, and FIG. 4 is a diagram for explaining the learning data holding unit. FIG. 5 is a diagram for explaining the final discriminator holding unit, FIG. 6 is a diagram for explaining the detection result storage unit, and FIG. 7 and FIG. FIG. 9 and FIG. 10 are diagrams for explaining section sound source detection processing (Multi-class AdaBoost).

  As illustrated in FIG. 2, the noise detection apparatus 10 according to the first embodiment includes an input unit 11, an output unit 12, an input / output control I / F unit 13, a storage unit 20, and a control unit 30.

  The input unit 11 inputs data used for various types of processing by the control unit 30 and operation instructions for performing various types of processing using a microphone, a keyboard, a mouse, or the like. Specifically, when the input unit 11 inputs learning data (a plurality of data including noise-superimposed speech data) with a microphone, the input data is held by a learning data holding unit 21 described later. Further, when the input unit 11 inputs data to be detected by the section sound source detection unit 32 described later with a microphone, the input data is held in the input data temporary storage unit 23 described later.

  In the first embodiment, the method of inputting the learning data and the detection target data to the noise detection device 10 by the input unit 11 as a microphone has been described. However, the present invention is not limited to this. For example, an audio file obtained by converting learning data into electronic data or an audio file obtained by converting data to be detected into electronic data to the input unit 11 as an external storage device or communication unit is input to the noise detection device 10. The present invention can be similarly applied.

  The output unit 12 outputs the results of various processes by the control unit 30 and operation instructions for performing various processes to a display or a printer. Specifically, the output unit 12 outputs the detection result stored in the detection result storage unit 24 described later to a display or a printer.

  For example, the output unit 12 outputs a detection result as shown in FIG. 3 to the display. 3 will be described in detail. The upper half of FIG. 3 is data to be detected, and is a diagram illustrating a waveform of noise superimposed speech that is present with noise superimposed on a speech section. The lower half of FIG. 3 shows the result of the noise detection apparatus 10 according to the present invention detecting the noise section and identifying the noise type (sound source) from the noise superimposed speech shown in the upper half. It is a thing. That is, in the waveform of the noise-superimposed speech exemplified in FIG. 3, noise having “spray sound” as a sound source, noise having “paper breaking sound” as a sound source, and “phone sound” It can be seen that noise having a “call sound” as a sound source was superimposed on the voice.

  In the example of FIG. 3, the method in which the output unit 12 outputs the detection result using a waveform or the like has been described. However, the present invention is not limited to this, and the detection result may be text (for example, “frame No. 100”). The sound source of the noise that exists in the section separated by "" is a phone call tone ", etc.) or other output that eliminates (suppresses) noise instead of outputting it to a display or printer Any of them may be output to the apparatus.

  The input / output control I / F unit 13 controls data transfer among the input unit 11, the output unit 12, the storage unit 20, and the control unit 30.

  The storage unit 20 stores data used for various controls in the control unit 30, and particularly as closely related to the present invention, as shown in FIG. 2, a learning data holding unit 21 and a final discriminator holding unit Unit 22, input data temporary storage unit 23, and detection result storage unit 24. The learning data holding unit 21 corresponds to the “learning data holding unit” described in the claims, and the final discriminator holding unit 22 is the “final discriminator holding unit” described in the claims. Corresponding to

  The learning data holding unit 21 holds a plurality of data including noise superimposed voice data as learning data. Specifically, the learning data holding unit 21 holds the learning data input by the input unit 11, and the held learning data is used for processing by the final discriminator derivation unit 31 described later.

  Here, the learning data held by the learning data holding unit 21 is data for the final discriminator deriving unit 31 to derive the final discriminator (the discriminator used for noise detection). The learning data holding unit 21 holds in advance by being input in advance by ten users. In the first embodiment, the case where the noise detection apparatus 10 holds the learning data in the learning data holding unit 21 in advance has been described. However, the present invention is not limited to this, and the learning data may be held. As for the process of deriving the final discriminator from the learning data, the present invention can be similarly applied to a case where another device different from the noise detection device 10 performs. In this case, the noise detection apparatus 10 may not include the learning data holding unit 21.

  For example, the learning data holding unit 21 holds learning data as shown in FIG. 4. That is, the learning data holding unit 21 includes a “voice only” learning data group, a “spray sound” learning data group, a “paper breaking sound” learning data group, As a group of “call tone” learning data, a common group composed of a plurality of voice data (illustrated by waveforms in FIG. 4) is held. This group of learning data is expressed as a “feature vector” in the AdaBoost theory.

  In addition, the learning data holding unit 21 determines whether the sound is “only sound or other”, “whether it is a spray sound or other than that”, “whether it is a sound that breaks paper”, for each sound source. , “Is it other than that”, “whether it is a phone call sound, or other than that”, and the like are stored in association with each voice data. Such information is expressed as “label” in the AdaBoost theory. The “feature vector” and “label” will be described in detail when the final discriminator deriving unit 31 is described.

  In the first embodiment, the learning data holding unit 21 holds one common group, and for this one group, “voice only”, “spray sound”, “paper breaking sound”, Although the case where a label is assigned and held for each sound source such as “phone call sound” has been described, the present invention is not limited to this. For example, when another sound source is selected as the sound source, or when a plurality of groups are held in duplicate instead of giving a label for each sound source to one group (a label is assigned to each sound source). The present invention can be similarly applied. That is, the sound source may be appropriately selected according to the environment in which the noise detection apparatus 10 according to the present invention is used, and any form of holding may be used.

  The final discriminator holding unit 22 assigns a final discriminator (final discriminator for discriminating a binary value of whether or not the noise is generated by a predetermined sound source) for each predetermined sound source. Hold on. Specifically, the final discriminator holding unit 22 holds the final discriminator derived by the final discriminator deriving unit 31 described later for each predetermined sound source, and the held final discriminator is a section sound source detection unit described later. 32 is used for the process.

  Here, since the final discriminator held by the final discriminator holding unit 22 is a discriminator for the section sound source detection unit 32 to detect noise, the final discriminator before the detection processing by the section sound source detection unit 32 is performed. Derived in advance by the deriving unit 31 and held in advance by the final discriminator holding unit 22. In the first embodiment, the noise detection apparatus 10 includes the final classifier deriving unit 31 and the final classifier derived by the final classifier deriving unit 31 is held by the final classifier holding unit 22. However, the present invention is not limited to this, and the processing for holding the learning data and deriving the final discriminator from the learning data is also performed in a case where another device different from the noise detection device 10 performs. The invention can be applied as well. In this case, the noise detection apparatus 10 holds the final discriminator by being input in advance by the user of the noise detection apparatus 10.

  The final discriminator will be described with an example. For example, the final discriminator holding unit 22 holds the final discriminator as shown in FIG. That is, the final discriminator holding unit 22 “identifies whether or not there is only sound” as the identification content “final discriminator of“ sound only ””, and “identifies whether or not it is a spray sound” as the identification content "Final sound detector for spray sound", "Identify whether the sound is a paper breaker" as the identification content, "Final discriminator for the paper break sound", and "Phone call sound" as the identification content “Final identifier of“ phone call tone ”” etc. are stored. In the first embodiment, the final classifier is derived using the AdaBoost theory. The AdaBoost theory will be described in detail when the final classifier deriving unit 31 is described. Further, the item “identification content” shown in FIG. 5 is given for convenience of explanation, and is not necessarily an item that the final discriminator holding unit 22 should hold.

  The input data temporary storage unit 23 temporarily stores data to be detected by the noise detection device 10. Specifically, the input data temporary storage unit 23 temporarily stores the data to be detected input by the input unit 11, and the temporarily stored data to be detected is the section sound source detection unit 32 described later. It is used for processing by.

  Here, since the data held in the input data temporary storage unit 23 is data for the section sound source detection unit 32 to detect noise, it is input in advance by the user of the noise detection device 10 or each time noise detection is performed. The input data temporary storage unit 23 temporarily stores the input data. Further, the input data temporarily stored in the input data temporary storage unit 23 may be deleted immediately after the processing by the section sound source detection unit 32 is completed, or may be stored for a predetermined period as necessary. In addition, the period during which the input data temporary storage unit 23 stores the input data can be appropriately changed according to the operation.

  The detection result storage unit 24 stores a detection result of data to be detected. Specifically, the detection result storage unit 24 stores detection results detected (or smoothed after detection) by the section sound source detection unit 32 and the smoothing unit 33 described later, and the stored detection results are stored in the output unit 12. Is output by.

  Here, the detection result stored in the detection result storage unit 24 may be deleted immediately after the output processing by the output unit 12 is completed, or may be stored for a predetermined period as necessary. The period during which the result storage unit 24 stores the detection result can be changed as appropriate according to the operation.

  For example, the detection result storage unit 24 stores the detection result as illustrated in FIG. 6. That is, the detection result storage unit 24 holds information related to the noise source of noise present in the data for each section divided by the data frame. Here, the upper half of FIG. 6 is a detection result before smoothing of the data subjected to the detection process by the section sound source detection unit 32, and the lower half of FIG. The detection result after performing the smoothing process with respect to data by the smoothing part 33 is shown. The difference between the two will be described in detail when the smoothing unit 33 is described.

  In the first embodiment, the method in which the detection result storage unit 24 stores the detection results as shown in FIG. 6 has been described. However, the present invention is not limited to this, and the detection results are stored in other forms. The detection result stored in the detection result storage unit 24, such as a method for performing the detection and a method for not storing the detection result before smoothing, can be appropriately changed according to the operation.

  The control unit 30 performs various controls in the noise detection apparatus 10, and particularly those closely related to the present invention include a final discriminator derivation unit 31, a section sound source detection unit 32, and a smoothing unit as shown in FIG. And a conversion unit 33. The final discriminator deriving unit 31 corresponds to the “final discriminator deriving unit” described in the claims, and the section sound source detection unit 32 corresponds to the “detecting unit” described in the claims. The smoothing unit 33 corresponds to “smoothing means” described in the claims.

  The final discriminator deriving unit 31 finishes learning by learning a discriminator (a discriminator for discriminating binary whether data is noise caused by a predetermined sound source, hereinafter, a weak discriminator) from learning data. The final classifier for each predetermined sound source is derived using boosting for deriving the final classifier. Specifically, the final discriminator deriving unit 31 derives the final discriminator from the learning data held by the learning data holding unit 21 using the AdaBoost theory, and the derived final discriminator is finally discriminated. It is held in the vessel holder 22.

  In the first embodiment, the noise detection apparatus 10 holds learning data in the learning data holding unit 21 in advance, and the final discriminator derivation unit 31 derives the final discriminator from the learning data held in advance. However, the present invention is not limited to this, and a case where a process different from the noise detection apparatus 10 performs the process of holding the learning data and deriving the final discriminator from the learning data is performed. In addition, the present invention can be similarly applied. In this case, the noise detection apparatus 10 may not include the final discriminator deriving unit 31.

  Hereinafter, the final classifier deriving process by the final classifier deriving unit 31 will be described in detail. The final discriminator derivation process is performed using boosting. In the first embodiment, an example will be described in which the booster is one of boosting processes (AdaBoost).

  The outline of the AdaBoost procedure will be described with reference to FIG. 7. In AdaBoost, first, a weak classifier is learned from learning data weighted with initial weights (steps S701 to S702) (step S703). The weight of the learning data is updated so that the weight of the learning data that has been erroneously identified by the weak classifier is increased (step S705). Next, a new weak classifier is learned from the learning data weighted with the updated new weight (step S703), and again the weight of the learning data that has caused erroneous identification with the new weak classifier is increased. Then, the weight of the learning data is updated (step S705). In this way, weak classifiers are automatically generated (learned), and finally, a final classifier (final classifier) is generated by weighted majority of a plurality of weak classifiers generated (step S704). Step S706).

  The AdaBoost procedure described above will be described in more detail with reference to FIG. 8. In AdaBoost, first, as shown in FIG. 8A, a learning feature in which “feature vectors” and “labels” are associated with each other. Data is provided (corresponding to step S701). Here, the “feature vector” and “label” are, for example, a group of learning data held by the learning data holding unit 21 and “spray sound or otherwise” in the first embodiment. Information.

  Next, in AdaBoost, as shown in FIG. 8B, the weight of learning data is initialized (corresponding to step S702). That is, for example, if the “label” of some learning data is “spray sound”, it is divided by twice the number of data in which “label” becomes “spray sound” in the entire learning data Is the initial weight of the learning data. Similarly, for example, if the “label” of a certain learning data is “other (other than spray sound)”, the data in which “label” is “other (other than spray sound)” in the entire learning data. The initial weight of the learning data is obtained by dividing the value by twice the number of. The reason for dividing by 2 times is for normalization.

  Subsequently, in AdaBoost, as shown in FIG. 8C, after learning with the weak classifier from the learning data weighted with the initial weight, the weight of the weak classifier itself is determined and its weak classification is performed. The weight of the learning data is updated so that the weight of the learning data that has been erroneously identified by the device increases (corresponding to steps S703 to 705). That is, for example, from the learning data of “spray sound” weighted with the initial weight in FIG. 8B, as shown in (2.1), the weak classifier is set so as to minimize the misclassification. After learning (referred to as weak classifier 1), the weight of the weak classifier 1 itself is determined and expressed by formula (2.4) as shown in formulas (2.2) and (2.3). As described above, the weight of the learning data is updated so that the weight of the learning data that has been erroneously identified by the weak classifier 1 is increased.

  In this way, AdaBoost automatically generates (learns) a plurality of weak classifiers from weak classifier 1 to weak classifier T, and finally, a plurality of weak classifiers are generated as shown in FIG. 8D. A final discriminator (final discriminator) is generated by the weighted majority of the weak discriminators (corresponding to step S706). For example, a final discriminator of “spray sound” is generated. In other words, the weak discriminator sets a threshold value so that the weighted error is minimized in each dimension (1 to T), and selects a dimension that further minimizes the weighted error.

  In this way, the final discriminator deriving unit 31 according to the first embodiment derives a final discriminator for each predetermined sound source from the learning data by using the AdaBoost theory as described above. Specifically, the final discriminator deriving unit 31, as shown in the equation (1.1) in FIG. 9, “label” of the learning data (a common group) held by the learning data holding unit 21. ”Is replaced for each predetermined sound source, and a final discriminator for each predetermined sound source is derived as shown in Equation (1.2).

  Here, as described above, AdaBoost is basically two-class discrimination, but if noise removal (suppression) is considered, the noise source is identified (what noise is mixed in the voice) It is desirable to remove (suppress) noise using noise data stored in advance for each sound source when removing (suppressing) noise. For this reason, the final discriminator deriving unit 31 realizes multi-class by extending AdaBoost to be adaptable to multi-class problems (creating a plurality of binary discriminators of one class versus other classes). In addition, the section sound source detection unit 32 can perform identification of noise types (sound sources).

  The section sound source detection unit 32 detects the noise section and the noise source of the input noise superimposed speech data. Specifically, the section sound source detection unit 32 uses the final discriminator for each predetermined sound source stored in the final discriminator holding unit 22 by using the noise superimposed voice data stored in the input data temporary storage unit 23. By identifying each frame and determining the final classifier with the highest score identified as one of the binary values as a result of identification, the data noise section is a section divided by frames. In addition, it is detected that the noise source included in the data is the predetermined sound source indicated by the determined final discriminator, and the detection result is stored in the detection result storage unit 24, or by the smoothing unit 33. It is used for processing.

  Hereinafter, the detection process of the section sound source detection unit 32 will be described with reference to FIG. 10. The section sound source detection unit 32 first converts the noise superimposed speech data (step S1001) stored in the input data temporary storage unit 23 to the final. Each final discriminator for each predetermined sound source held by the discriminator holding unit 22 is used to identify each frame (step S1002). In other words, the section sound source detection unit 32 inputs a logarithmic mel filter bank, which is a feature amount for each frame, to the Multi-class AdaBoost, and “whether it is only sound” or “whether it is a spray sound”. , “Whether it is a sound that breaks paper”, “Whether it is a phone call sound”, and so on. Then, as a result of the identification, by determining the final discriminator having the highest score identified as one of the two values (step S1003), the data noise section is a section divided by frames. In addition, it is detected that the noise source included in the data is a predetermined sound source indicated by the determined final discriminator.

  The detection processing of the section sound source detection unit 32 will be described with a specific example. The section sound source detection unit 32 uses the final discriminator for each sound source to store the input noise superimposed speech data in units of data frames. Identify with Here, the data frame is a group of data divided by a certain time (for example, “20 ms”) and indicates a data section.

  For example, the section sound source detection unit 32 converts the input data into “sound-only final identifier”, “spray sound final identifier”, “paper-breaking sound final identifier”, “phone call sound” Using a final discriminator such as “final discriminator”, the section of “frame No. 100” of data is identified. Then, for example, “frame No. 100” of the data is identified with a score of “−0.3” as one of the two values as a result of identification using the “final classifier only for voice”. The Similarly, “frame No. 100” of the data is identified with a score of “−0.1” in any one of the binary values as a result of identification using the “final identifier of spray sound”. As a result of identification using “the final discriminator of the sound that breaks the paper”, one of the two values is identified with a score of “−0.2”, and the “final discriminator of the phone call tone” As a result of the identification used, any one of the two values (here, “phone call tone”) is identified with a score of “0.5”.

  Subsequently, the section sound source detection unit 32 determines the final discriminator having the highest score identified by any one of the binary values. For example, the section sound source detection unit 32 determines that the final discriminator identified with the highest score of “0.5” as one of the two values is the “final discriminator of the telephone call sound”. judge.

  Then, the section sound source detection unit 32 detects that the noise section of the data is a section divided by this frame, and that the noise source existing in the data is the sound source indicated by the determined final discriminator. . For example, the section sound source detection unit 32 detects that the noise section of the data is a section divided by “Frame No. 100” and that the sound source of the noise present in the data is “phone call sound”. To do.

  Thus, the section sound source detection unit 32 extends AdaBoost so that it can be applied to the multi-class problem (by making the highest result among the plurality of binary discriminators the identification result). Multi-class is realized, and the type of noise (sound source) can be identified.

  In the first embodiment, after detection by the section sound source detection unit 32, smoothing processing by the smoothing unit 33 (step S1004), output processing by the output unit 12 (step S1005), and the like are performed. In the first embodiment, the section sound source detection unit 32 identifies the input noise-superimposed speech data in units of frames, and also detects that the data noise section is a section divided by frames. However, the present invention is not limited to this, and the present invention is similarly applied to the case where the section sound source detection unit 32 does not rely on a method of identifying input noise superimposed speech data in units of frames. Can be applied.

  The smoothing unit 33 performs smoothing on frames with different results when consecutive frames of input data include frames with different results of identification from other frames. Specifically, the smoothing unit 33 is a detection result detected by the section sound source detection unit 32, and a result of identification different from that of other frames with respect to the detection result stored in the detection result storage unit 24. If the frame is included, smoothing is performed on frames with different results, and the result of smoothing is stored in the detection result storage unit 24.

  For example, when the frames detected as “noise” are continuous and there is a frame that is detected as “speech” very slightly, the smoothing unit 33 is not “speech”, and the frame is erroneous. Think of it as the “noise” that caused the identification. Specifically, for example, the smoothing unit 33 sets, for example, the largest number of detection results among the three frames before and after and a total of seven frames, and repeats until there is no change. The smoothing unit 33 regards a section that is continuously determined as “noise” as one “noise” even if the detection results differ from frame to frame.

  As illustrated in FIG. 6, for example, when focusing on “frame No. 101”, the smoothing unit 33 has the largest number of detection results among the three frames before and after “frame No. 101” and the total of seven frames. Is a “phone call tone”, and therefore the detection result of “frame No. 101” is also a “phone call tone”.

  In addition, since the smoothing unit 33 according to the first embodiment targets, for example, noise with a duration of 200 ms or longer, what is determined as noise with a duration shorter than that is considered to be “welling up”. For example, noise is cut out for noise of 100 ms or less, which is half of the duration of 200 ms.

  In the first embodiment, the technique has been described in which the smoothing unit 33 performs smoothing on different result frames and outputs the smoothed results as detection results. However, the present invention is not limited to this. Instead, the present invention can be similarly applied to a method in which a smoothing process is not performed by the smoothing unit 33 and an unsmoothed result is output as a detection result.

[Procedure of processing by the noise detection apparatus according to the first embodiment]
Next, a procedure (example) of processing performed by the noise detection apparatus according to the first embodiment will be described with reference to FIG. FIG. 11 is a flowchart of a process procedure performed by the noise detection apparatus according to the first embodiment.

  First, the noise detection apparatus 10 determines whether or not the section sound source detection unit 32 has received input of data to be detected (step S1101). When the input of data is not received (No at Step S1101), the noise detection apparatus 10 returns to the process of determining whether or not the input of the data to be detected is received in the section sound source detection unit 32.

  On the other hand, when the input of data is accepted (Yes in step S1101), the noise detection apparatus 10 identifies one frame of the input data with the final discriminator of a predetermined sound source in the section sound source detection unit 32 ( Step S1102).

  Subsequently, the noise detection apparatus 10 determines whether or not the section sound source detection unit 32 has identified all sound sources with the final discriminators (step S1103). When not identifying with all the final discriminators of the sound source (No at Step S1103), the noise detection apparatus 10 changes the final discriminator in the section sound source detection unit 32 (Step S1104), and changes the predetermined discriminator after the change. The process returns to the process of identifying one frame of the input data by the final discriminator of the sound source.

  On the other hand, when the final sound discriminators of all sound sources have been identified (Yes at step S1103), the noise detection device 10 has the highest score identified by any one of the two values in the section sound source detection unit 32. A high final discriminator is determined (step S1105).

  In the noise detection device 10, the section sound source detection unit 32 indicates that the data noise section is a section divided by frames, and that the noise discriminator present in the data is indicated by the determined final discriminator. It is detected that the sound source is a predetermined sound source (step S1106).

  Next, the noise detection apparatus 10 determines whether or not all the frames are detected in the section sound source detection unit 32 (step S1107). If all the frames have not been detected (No at step S1107), the noise detection apparatus 10 changes the frame at the section sound source detection unit 32 (step S1108) and is input by the final discriminator of the predetermined sound source. Return to the process of identifying one frame of the data.

  On the other hand, when all the frames have been detected (Yes at Step S1107), the noise detection apparatus 10 performs a smoothing process at the smoothing unit 33 (Step S1109), and outputs the detection result to the output unit 12. (Step S1110), the process is terminated.

  In this way, the noise detection apparatus 10 according to the first embodiment can identify the type of noise (sound source).

[Effect of Example 1]
As described above, according to the first embodiment, the final discriminator that identifies the binary value of whether or not the noise-superimposed speech data that is present with the noise superimposed on the speech section is noise from a predetermined sound source is predetermined. Is stored for each sound source, and the input noise superimposed speech data is identified using each final discriminator for each predetermined sound source, and as a result of the identification, one of the binary values is identified. By determining the final discriminator having the highest score, it is detected that the noise source existing in the data is a predetermined sound source indicated by the determined final discriminator, so that the noise type (sound source) is identified. It becomes possible.

  In addition, according to the first embodiment, a discriminator that holds a plurality of data including noise superimposed voice data as learning data and discriminates whether or not the data is noise caused by a predetermined sound source is used for learning. The final classifier for each given sound source is derived from the stored learning data using boosting that derives the final classifier that has been learned by learning from the data, so the type of noise (sound source) is appropriate. Can be identified.

  Also, according to the first embodiment, the noise detection device derives the final discriminator using Adaboost as boosting, so that it is possible to appropriately identify the type of noise (sound source).

  In addition, according to the first embodiment, the noise detection apparatus further identifies that the noise-superimposed speech data is identified in units of frames and further detects that the data noise section is a section divided by frames. In addition, it is possible to detect a noise interval.

  In addition, according to the first embodiment, the noise detection apparatus has a result in which the identification result identified by the final classifier determined by the detection unit is different from the other frames in consecutive frames of the input data. When the frame is included, smoothing is performed on frames with different results, so that the noise type (sound source) can be accurately identified.

  So far, the outline, features, configuration, processing procedure, and the like of the noise detection apparatus 10 have been described as the first embodiment. Next, as the second embodiment, an evaluation experiment by the noise detection apparatus 10 according to the present invention is described. Will be described. In addition, the evaluation experiment in Example 2 is mainly intended to evaluate the recall rate, the matching rate, and the noise identification rate of the noise detection apparatus 10 according to the present invention.

[Experimental conditions]
First, the experimental conditions of the evaluation experiment in Example 2 will be described. For “learning data”, utterance data of 21 male speakers × 10 utterances from the continuous speech database for research provided by ASJ is used, and “evaluation data” to be verified is also provided by ASJ. From the research continuous speech database, the speech data of 5 male speakers × average 240 utterances was used. For the “noise”, three types of data “spray sound”, “sound to break paper”, and “phone call sound” were used from non-voice dry sources provided by RWCP.

  As the “learning data”, utterance data and data obtained by superimposing each “noise” adjusted for SNR on the utterance data were used. Further, the SNR of the “learning data” was randomly changed between “−5 dB” and “5 dB”. On the other hand, the “evaluation data” used was obtained by superimposing 1 to 3 “noises” with the SNR adjusted for a duration of 200 ms or more on one utterance. However, there is no data in which another noise is superimposed in the section where the noise is superimposed. Further, the SNR of the “evaluation data” is “−5 dB”, “0 dB”, and “5 dB”.

  In the evaluation experiment in Example 2, the SNR was obtained by the formulas shown in (A) to (C) of FIG. In addition, “logarithmic mel filter bank” was used as the feature amount. For both “learning data” and “evaluation data”, the frame width is “20 ms”, the frame shift is “10 ms”, and the pre-emphasis of “1- (0.97z to the (−1) power)” and the Hamming window are used. ing.

[Noise detection]
The determination in the evaluation experiment in Example 2 will be described. First, from the viewpoint of detection only, a section that is correctly detected is determined as “correct” even if the noise type (sound source) is different. And In addition, an error margin is determined, and an error that is within the margin of the correct answer data is also determined as “correct answer”. The margin was set to “30 ms” in the evaluation experiment in Example 2. In addition, the detection interval is too large, “false detection”, and the detection interval is too small, “not detected”.

  For the evaluation, the detection rate (Detection rate) shown in FIG. 13A, the recall rate shown in FIG. 13B (Recall rate), and the precision rate (Precision rate) shown in FIG. 13C. The following three are used. Here, the detection rate, the recall rate, and the matching rate are calculated using the number of correct answers “Tp”, the number of false detections “Fp”, the number of undetections “Tn”, and the total number of noises “Ta” in the detected section. , (A) to (C) in FIG. Here, although the detection rate and the recall rate are essentially the same, in the evaluation experiment in Example 2, since the noise having a large section was evaluated as “false detection”, the detection rate and the recall rate are Since different values may come out, both are shown.

  The result of the evaluation experiment is as shown in FIG. For all SNRs, the detection rate, reproducibility, and matching rate are “95% or higher”, and good results are obtained. It is confirmed that “noise” with a strength of “5 dB” or higher can be detected. It was done.

[Noise discrimination]
By the way, in the above, even if the section is correct, even if the noise type (sound source) is different, it is determined as “correct”. Next, the noise discrimination rate among the noises in which the section is correctly determined is evaluated. Further, in addition to the detection rate, the correct answer rate of noise was determined for the correct section and the correct noise identification result. The result is shown in FIG.

  FIG. 15 shows that a high noise discrimination rate (noise discrimination rate) exceeding “99.5%” can be obtained in all SNRs. In other words, the noise type (sound source) can be correctly identified for most of the detected ones. An example of the output of the evaluation experiment that has been correctly detected and identified is shown in FIG. As shown in FIG. 3, the “phone call sound” cannot be discriminated only by the waveform, but all can be correctly detected by the noise detection device 10 according to the present invention, and the noise type (sound source) is correctly identified. is made of.

[Change in detection accuracy due to mismatch model]
By the way, in the above, the SNRs of “noise” of “evaluation data” and “noise” of “learning data” were equal. Next, the SNRs of “evaluation data” and “learning data” were changed. Let us examine how much the detection accuracy changes. In the same manner as described above, a model learned with “Study data” of “SNR-5 dB to 5 dB”, a model learned with “Study data” of “SNR-5 dB only”, and “Learning data” of “SNR 0 dB only” For each of the learned model and the model learned using the “learning data” of “SNR 5 dB only”, the detection rate, the recall rate, and the adaptation rate for the “evaluation data” of “SNR-10 dB to 10 dB” are calculated, Compare the differences. Note that the number of learnings of AdaBoost is similarly 1,000 times.

  The results are as shown in FIG. The relevance rate is almost determined by the SNR of the “learning data” used during learning because the number of false detections was not affected even if the “noise” SNR of the “evaluation data” was changed. The lower the SNR of the “learning data”, the higher the matching rate.

  In addition, as the SNR of “evaluation data” increases, the number of undetected items tends to increase. As shown in FIG. 16, when the “learning data” “SNR-5 dB” discriminator is used, the detection rate of “evaluation data” “SNR5 dB” is “76.7%”, “learning data” “SNR0 dB”. In the discriminator, the evaluation data “SNR10 dB” is reduced to the detection rate “69.9%”. In the discriminator learned for all, the matching rate decreased compared to “−5 dB” and “0 dB”, but the decrease in the detection rate and the recall rate was small.

[Change in noise discrimination accuracy by mismatch model]
Under the same conditions as above, the “noise” discrimination rate and the correct answer rate are evaluated. The results are shown in FIG. FIG. 17 shows that the discrimination rate decreases as the difference between the “noise” SNR of “learning data” and the SNR of “evaluation data” increases. When using the discriminator learned with “SNR-5 dB”, the discrimination rate of “evaluation data” and “SNR10 dB” decreases to “94.4%”, but shows a high value, with “SNR5 dB” When the learned classifier was used, the identification rate of “evaluation data” “SNR-10 dB” was a relatively low value of “80.1%”. The data learned with “SNR 0 dB” is “95.9%” for “evaluation data” “SNR 10 dB” and “93.7%” for “evaluation data” “SNR-10 dB”. Further, the correct answer rate of “noise” shows a high value when the model match is high, but on average, what was learned with “SNR−5 dB to 5 dB” showed the highest value.

  Although the embodiments of the present invention have been described so far, the present invention may be implemented in various different forms other than the above-described embodiments.

[System configuration, etc.]
Of all the processes described in the present embodiment, all or a part of the processes described as being automatically performed (for example, the process of deriving the final discriminator from the stored learning data) is manually performed. (For example, processing for deriving a final discriminator from learning data to be held by inputting a command as necessary), or processing described as being performed manually (for example, learning data) All or a part of the data may be automatically performed by a known method (for example, automatically downloaded via a network). In addition, the processing procedure, control procedure, specific name, and information including various data and parameters shown in the above-mentioned document and drawings can be arbitrarily changed unless otherwise specified.

  Each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated (for example, FIG. 2). In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. The detection result storage unit can be configured to be integrated before and after smoothing, for example. Further, all or any part of each processing function performed in each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

  The noise detection method described in the present embodiment can be realized by executing a program prepared in advance on a computer such as a personal computer or a workstation. This program can be distributed via a network such as the Internet. The program can also be executed by being recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, an MO, and a DVD and being read from the recording medium by the computer.

  As described above, the noise detection device and the noise detection method according to the present invention are useful for detecting “noise”, and are particularly suitable for identifying the type of noise (sound source).

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a diagram for explaining an overview and characteristics of a noise detection device according to a first embodiment. 1 is a block diagram illustrating a configuration of a noise detection apparatus according to a first embodiment. It is a figure for demonstrating an output part. It is a figure for demonstrating the data holding part for learning. It is a figure for demonstrating the last discriminator holding | maintenance part. It is a figure for demonstrating a detection result memory | storage part. It is a figure for demonstrating the last classifier derivation | leading-out process (AdaBoost). It is a figure for demonstrating the last classifier derivation | leading-out process (AdaBoost). It is a figure for demonstrating a section sound source detection process (Multi-class AdaBoost). It is a figure for demonstrating a section sound source detection process (Multi-class AdaBoost). 3 is a flowchart illustrating a processing procedure performed by the noise detection apparatus according to the first embodiment. It is a figure for demonstrating the calculation formula which calculates | requires SNR. It is a figure for demonstrating a detection rate, a reproduction rate, and a precision. It is a figure which shows the evaluation result of the noise detection apparatus which concerns on Example 2. FIG. It is a figure which shows the evaluation result of the noise detection apparatus which concerns on Example 2. FIG. It is a figure which shows the evaluation result of the noise detection apparatus which concerns on Example 2. FIG. It is a figure which shows the evaluation result of the noise detection apparatus which concerns on Example 2. FIG. It is a figure which shows the waveform which the telephone sound superimposed on the audio | voice. It is a figure which shows the waveform which each noise superimposed on the audio | voice.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Noise detection apparatus 11 Input part 12 Output part 13 Input / output control I / F part 20 Storage part 21 Learning data holding part 22 Final discriminator holding part 23 Input data temporary storage part 24 Detection result storage part 30 Control part 31 Final identification Deriving unit 32 Section sound source detection unit 33 Smoothing unit

Claims (6)

Final discriminator holding means for holding, for each predetermined sound source, a final discriminator for discriminating whether or not the data of the noise-superimposed speech existing by superimposing noise on the speech section is noise from a predetermined sound source;
The inputted noise superimposed speech data is identified using each final discriminator for each of the predetermined sound sources held by the final discriminator holding means, and as a result of the discrimination, any one of the binary values is identified. Detecting means for detecting that the sound source of the noise present in the data is a predetermined sound source indicated by the determined final identifier,
A noise detection apparatus comprising: Learning data holding means for holding a plurality of data including noise superimposed speech data as learning data;
The learning data holding is performed by using boosting for deriving a final discriminator whose learning is completed by learning a discriminator for discriminating whether or not the data is noise caused by a predetermined sound source from the learning data. A final discriminator deriving unit for deriving a final discriminator for each predetermined sound source from the learning data held by the unit;
The noise detection apparatus according to claim 1, further comprising:   The noise detection apparatus according to claim 2, wherein the final classifier deriving unit derives the final classifier using Adaboost as the boosting.   The said detection means identifies the data of the said noise superimposed audio | voice per frame, and further detects that the area of the noise of the said data is the area divided | segmented by the said frame. The noise detection device according to any one of the above.   In the case where consecutive frames of the input data include a frame whose result of identification identified by the final classifier determined by the detection means is different from that of other frames, the frame is different. The noise detection apparatus according to claim 4, further comprising a smoothing unit that performs smoothing on the resulting frame. A final discriminator holding step for holding, for each predetermined sound source, a final discriminator for discriminating whether the data of the noise-superimposed speech existing by superimposing noise on the speech section is noise from a predetermined sound source;
The input noise-superimposed speech data is identified using each final discriminator for each of the predetermined sound sources held in the final discriminator holding step, and as a result of the discrimination, one of the two values is identified. Detecting the final discriminator having the highest score identified in step (b) to detect that the noise source existing in the data is a predetermined sound source indicated by the determined final discriminator;
The noise detection method characterized by including.

Images