JP2005203981A - Device and method for processing acoustic signal - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reproduce clear sound without depending on any sound environments by improving audibility of speech, or the like, according to the distinction between voice and non-speech, surrounding noise, noise, or the degree of reverberation in an acoustic signal processor, such as a hearing aid and a television. <P>SOLUTION: In the acoustic signal processor 20 provided in the hearing aid, or the like, various acoustic signals are sorted out at a learning phase processor 27 (a learning step). In a signal processing execution phase processor 28, the acoustic signal sorted out in the learning phase is used for signal processing to allow a user to listen to the amplification, or the like of an input acoustic signal (a signal processing execution step). <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to an acoustic signal processing device that facilitates hearing and assists hearing, such as a hearing aid, for example, and in particular, an acoustic signal processing device and an acoustic device that are suitable for appropriate signal processing depending on the sound environment. The present invention relates to a signal processing method.

  In general, an acoustic signal processing apparatus is used for receiving and reproducing various sounds, and an appropriate signal processing method or signal processing depending on the sound environment around the user or the type of sound input to the acoustic signal processing system. Often have different parameters (amplification, etc.). For example, an acoustic signal processing device for a hearing aid has different amplification characteristics depending on the type of voice or non-voice, ambient noise or reverberation (hereinafter referred to as a sound environment). For this reason, in order to achieve the maximum speech intelligibility, mode switching and parameter adjustment are required for each environment.

Conventionally, each of these mode switching and parameter adjustment operations is manually performed by the user in accordance with the sound environment. In addition, recent hearing aids are provided with a plurality of mode switching functions with different amplification parameters. On the other hand, when it is difficult for the user to hear, the user himself can switch the mode using a button or the like.
Conventionally, various methods for switching or controlling the mode for each environment have been proposed.

For example, gain control in a voice automatic gain control apparatus has been proposed (see Patent Document 1). The automatic gain control device described in Patent Document 1 determines whether or not to perform gain control by comparing the calculated input acoustic signal level with a threshold value. As a result, unnecessary gain control with respect to the end portion of speech and background noise can be suppressed, and a larger maximum gain can be obtained.
In addition, a hearing aid device has been proposed in which the sound is clearly heard in consideration of the quality of environmental noise (see Patent Document 2). In this listening method, the level distribution of the input acoustic signal is detected to determine whether the environmental noise is stationary or non-stationary, and a hearing aid process is performed accordingly. As a result, it is possible to clearly hear the sound to be listened to regardless of the quality of the environmental noise.

  Furthermore, Patent Literature 3 discloses a correction method for a hearing aid. The hearing aid described in Patent Document 3 is provided with an acoustic sensor for identifying ambient noise and necessary speech separately from a microphone (microphone) that inputs speech to be heard, and is based on a noise sensor obtained from the acoustic sensor. Thus, spectral characteristics are changed, and spectral correction and gain correction are adjusted in real time. As a result, the hearing aid recognizes ambient noise and the target voice, and operates with high accuracy and adapts to the acoustic environment of the person using the hearing aid.

  In addition, a method for determining the characteristics of a hearing aid signal processing unit according to an acoustic parameter extracted from an input acoustic signal has been proposed (see Patent Document 4). The hearing aid described in Patent Document 4 converts an acoustic parameter of an input acoustic signal into a fitting parameter indicating a hearing aid characteristic, describes a mapping relationship, and automatically adjusts the fitting parameter from the acoustic parameter of the input sound. As a result, an adaptive characteristic hearing aid can be obtained in which the hearing aid processing characteristic is determined adaptively to the environmental sound.

  Furthermore, a dynamic range compression type auditory compensation processing method has been proposed (see Patent Document 5). The hearing compensation processing method described in Patent Document 5 uses a loudness curve of a normal hearing person and a user, and gradually reduces the gain for an input sound below a preset sound pressure in accordance with the sound pressure of the input sound. is there. By reducing the amplification factor of minute noise, masking in the time direction due to noise in the silent part before and after the input speech can be improved.

In addition, a circuit that collectively implements clustering and labeling, which are basic processes in speech recognition and the like, with an analog circuit has been proposed (see Patent Document 6). The speech recognition circuit described in Patent Document 6 includes a similarity circuit that outputs a feature based on a self-organization algorithm, and a matrix circuit that performs a matrix operation on an output signal of the similarity circuit. A corresponding voltage signal is received, and a matrix calculation output that is closest to a previously prepared pattern is output as a recognition result. Thereby, a small-scale circuit voice recognition circuit suitable for a semiconductor integrated circuit can be realized.
JP-A-11-220345 JP 2001-128296 A JP 2000-13895 A JP 2002-369292 A JP-A-10-94095 JP 2002-279393 A

However, when the user goes out of the house, the acoustic signal processing device such as a hearing aid needs to frequently perform a mode switching operation when the surrounding sound environment changes abruptly. Therefore, there is a problem that the user performs a complicated operation such as manual switching or is forced to be difficult to hear.
On the other hand, the conventional technique is not capable of switching modes appropriately according to the sound environment. For example, both the automatic gain control device described in Patent Document 1 and the hearing aid device described in Patent Document 2 monitor only the input sound signal level, and the spectral characteristics other than the input sound signal level or the input sound signal It does not monitor the type. For this reason, each device can control the gain etc. according to the level of ambient noise, but it does not describe the state of the surrounding sound environment, such as the type of voice or non-voice, the presence or absence of reverberation, Fine control is not possible. Further, in the hearing aid device described in Patent Document 2, the level values of the detected level distribution are scattered and the sound quality is deteriorated.

  Furthermore, since the hearing aid described in Patent Document 3 needs to be provided with an acoustic sensor in addition to the voice input microphone, the cost of the hearing aid increases. Moreover, although the correction method described in Patent Document 3 patterns noise using spectral information of an acoustic sensor input and controls the amplification characteristics of speech, details of the patterning method are not disclosed. Therefore, even if the hearing aid can pattern extreme sounds such as siren sounds, it is difficult to classify the surrounding environment.

  Patent Document 4 does not describe what procedure is used to map the acoustic parameters and the characteristics of the hearing aid signal processing unit, and the effect is not always obtained. Further, the hearing aid described in Patent Document 4 does not use parameters that characterize sound. Further, when the mapping process does not operate, there is a high possibility that the amplification characteristic is completely different from the appropriate amplification characteristic, and stable audio processing is not necessarily performed.

Also, Patent Documents 5 and 6 do not disclose a technique for monitoring spectral characteristics other than the input sound signal level or the type of the input sound signal.
The present invention has been devised in view of such problems. For example, in the acoustic signal processing of a hearing aid, the input acoustic signal is identified, the acoustic signal is appropriately analyzed, and the surroundings are obtained based on the characteristic amount obtained by the analysis. By identifying the sound environment and processing acoustic signals such as amplification characteristics based on the identified information, stable and appropriate acoustic signal processing can be performed in any sound environment, making it easy to hear and stable. Another object of the present invention is to provide an acoustic signal processing apparatus and acoustic signal processing method capable of reproducing highly clear voice.

  For this reason, the acoustic signal processing device of the present invention includes a feature amount output unit that outputs first feature amount data representing a feature amount of an input acoustic signal, two-dimensional plane coordinates, and a second assigned to the coordinates. Self-Organizing Map (SOM) information data in which feature quantity data and group identification information for identifying a plurality of groups in which a plurality of learning acoustic signals are grouped in a two-dimensional plane are associated with each other. (Hereinafter referred to as SOM information data)) based on the first feature quantity data from the feature quantity output section, and the SOM information data held in the self organization map holding section. The group identification information searched by the search unit for searching for the group identification information corresponding to the first feature quantity data and the search unit of the plurality of signal processing type information representing the processing type of the acoustic signal A signal processing type information output unit that outputs signal processing type information corresponding to the information, and a signal processing unit that processes the input acoustic signal based on the signal processing type information output from the signal processing type information output unit (Claim 1).

  The acoustic signal processing device according to the present invention includes a feature amount output unit that outputs first feature amount data representing a feature amount of an input acoustic signal, coordinates in a multidimensional space, and second features assigned to the coordinates. A self-organizing map holding unit for holding SOM information data in which quantity data is associated with group identification information for identifying a plurality of groups in which a plurality of learning acoustic signals are grouped in a multidimensional space; and a feature quantity output unit A search unit for searching for group identification information corresponding to the first feature amount data based on the first feature amount data output from the SOM information data stored in the self-organizing map holding unit; A signal processing type information output unit for outputting signal processing type information corresponding to the group identification information searched by the searching unit among the plurality of signal processing type information representing the signal processing type, and a signal processing Based on the signal processing type information output from the component type information output unit and a signal processing unit for processing an input sound signal is characterized in that it is configured (claim 2).

The signal processing type information output unit may be configured as a parameter adjustment unit that adjusts and outputs a parameter related to a set value necessary for processing the acoustic signal based on the group identification information searched by the search unit. (Claim 3).
In addition, the acoustic signal processing device of the present invention learns based on the feature amount output unit that outputs the first feature amount data representing the feature amount of the learning acoustic signal, and the first feature amount data from the feature amount output unit. A learning unit that groups acoustic signals in a two-dimensional plane, coordinates of the two-dimensional plane, second feature amount data assigned to the coordinates, and group identification that identifies a plurality of groups grouped in the learning unit It is characterized by comprising a self-organizing map holding unit for holding SOM information data in which each information is associated (claim 4).

  Furthermore, in the acoustic signal processing method of the present invention, the learning acoustic signal is based on the first feature amount data representing the feature amount of the learning acoustic signal, and the second feature amount assigned to the coordinates on the two-dimensional plane. A learning step for generating SOM information data in which data and group identification information for identifying a plurality of groups grouped by the learning unit are associated with each other, and a learning step for the first feature data in the learning step A search step for searching for neighboring coordinates located in the vicinity of the first feature value data from the feature value output unit among a plurality of coordinates in the two-dimensional plane held in the SOM information data generated in step S; The signal processing type information corresponding to the group identification information of the neighboring coordinates searched in the search step among the plurality of signal processing type information representing the processing type of And a sound signal processing step for performing sound signal processing based on the signal processing type information output in the signal processing type information output step (Claim 5). .

According to the acoustic signal processing device of the present invention, sound with high clarity is reproduced stably in any sound environment.
In addition, according to the acoustic signal processing device of the present invention, clear sound can be obtained, which greatly contributes to improvement of the user's audibility.
Furthermore, regardless of the type of speech or non-speech, any acoustic signal can retain its feature value through the learning phase, so that a customized function can be further exhibited.

  In addition, by using a memory having an SOM function (SOM memory), memory data that cannot be uniquely identified is represented on a two-dimensional plane. Therefore, a designer can visually grasp the memory data and relates to acoustic signal processing. The operability is remarkably improved, and the feature amount of the acoustic signal can be efficiently classified and arranged by improving the operability. With this classification and arrangement, the acoustic signal processing apparatus can accurately and reliably identify the surrounding sound environment, and can stably perform acoustic signal processing such as amplification characteristics on the input acoustic signal. Thereby, clear voice or the like can be obtained.

Embodiments of the present invention will be described below with reference to the drawings.
(A) Description of One Embodiment of the Present Invention FIG. 1 is a block diagram of an acoustic signal processing apparatus according to a first embodiment of the present invention. The acoustic signal processing device 20 shown in FIG. 1 is used, for example, in a hearing aid, a radio circuit or a television audio circuit, etc., depending on the type of voice or non-voice, ambient noise, noise or reverberation, etc. Appropriate processing (acoustic signal processing) is performed so that voices and the like can be easily heard. In this acoustic signal processing, the amplification characteristic is changed for the acoustic signal such as increase / decrease in the level of the input acoustic signal, enhancement or reduction of a part of the spectrum band of the acoustic signal, and voice intelligibility is ensured. Yes.

(1) Learning Phase and Signal Processing Execution Phase The acoustic signal processing method in the acoustic signal processing device 20 of the present invention is mainly provided with two types of processing phases, and a learning phase (learning step) in which various acoustic signals are classified and arranged. And a signal processing execution phase (signal processing execution phase) for performing signal processing such as amplification of the input acoustic signal in order for the user to listen to the input acoustic signal using the acoustic signals classified and arranged in this learning phase. Step).

  The acoustic signal processing device 20 performs a learning phase during production of hearing aids, radios, televisions, and the like, and groups a large number of learning acoustic signals and records the grouped data. That is, the learning phase is performed before the user uses the product. In the following description, the user does not execute the learning phase but always uses the product in the signal processing execution phase.

  That is, according to the acoustic signal processing method of the present invention, in the acoustic signal processing device 20 provided in a hearing aid or the like, various acoustic signals are classified and organized in the learning phase processing unit 27 (learning step), and the signal processing execution phase processing unit 28 is processed. The signal processing is performed for the user to listen to the amplification of the input sound signal using the sound signals classified and arranged in the learning phase (signal processing execution step).

(2) Configuration of Acoustic Signal Processing Device 20 This acoustic signal processing device 20 includes an audio input unit 50, a feature analysis unit (feature amount output unit) 21, an SOM learning unit (self-organizing learning unit) 26, and an SOM. A coordinate search unit 22, an SOM information storage memory (self-organizing map holding unit) 23, a signal processing type information output unit 24, an acoustic signal processing unit 25, an amplifier 51, and an earphone 52 are provided. Yes. These blocks (circuit blocks) cooperate to perform a learning phase and a signal processing execution phase. Hereinafter, each block will be described in detail.

(2-1) Voice input unit 50
The voice input unit 50 acquires voice and ambient sounds, converts them into acoustic signals, and inputs the acoustic signals to the feature analysis unit 21. For example, a microphone and an amplifier are provided.
(2-2) Feature analysis unit 21
The feature analysis unit 21 outputs a feature amount vector (feature amount vector data: first feature amount data) representing the feature amount of the input acoustic signal, and functions as the feature amount output unit 21.

(2-3) Acoustic signal feature quantity and feature quantity vector Here, the acoustic signal feature quantity is, for example, the waveform of the acoustic signal, the level of the acoustic signal, the repetition period of the acoustic signal having a repetitive waveform, or the acoustic signal. A characteristic or property of the acoustic signal itself, such as a spectral component (power spectrum).
The feature quantity vector is a set of {1.1, 1.3,..., 1.2}, etc., which represents a feature quantity such as “A” by a plurality of elements. For example, the feature vector of the acoustic signal having the time width TW is the time t ₀ , t ₁ ,..., T _(n in _-1), respectively, quantizing the waveform of the audio signal (sampling), and quantized components _{F INPUT (1,1,0), F INPUT} (1,1,1), ..., F INPUT (1, 1, n-1) is output as a feature vector.

In the following description, “F _INPUT ” may be referred to as a feature quantity or a feature quantity vector.
The feature analysis unit 21 operates in both the learning phase and the signal processing execution phase. In the learning phase, the feature analysis unit 21 receives a large number of learning acoustic signals and extracts feature amounts F _INPUT (x, y, k) of the respective acoustic signals (k represents a natural number from 0 to n−1). . The extracted feature value F _INPUT (x, y, k) is stored as a feature value F _SOM (x, y, k) (second feature value data) in the SOM information storage memory 23 described later. Then, in the signal processing execution phase, the feature amount F _INPUT of input acoustic signal characteristic analysis section 21 is extracted (x, y, k), the feature amount F _SOM (x stored in the SOM information storage memory 23, y, k). Therefore, the feature analysis unit 21 performs feature analysis of the input acoustic signal in the signal processing execution phase, and also performs feature analysis of the learning acoustic signal in the learning phase as preprocessing.

Note that the noise component of the audio signal is suppressed by digital signal processing, and a clear audio signal is obtained.
(2-4) Difference between Feature Quantity Vector and Coordinate In general, the concepts of “vector” and “coordinate” are often equivalent, but in the following description, the feature quantity vector and the coordinate are different. As described above, the feature vector means a set of elements F _INPUT (x, y, 0) to F _INPUT (x, y, n−1), and the coordinate means an address of the SOM information storage memory 23. To do. In other words, n elements are held in association with the addresses of the SOM information storage memory 23.

(2-5) SOM information storage memory 23
In addition, the SOM information storage memory 23 has a plurality of two-dimensional plane coordinates, a feature vector F _SOM (x, y, k) assigned to the coordinates, and a plurality of learning acoustic signals grouped in the two-dimensional plane. SOM information data in which the group identification information for identifying each group is associated with each other is held.

  The SOM information storage memory 23 comprises a two-layer network having an input layer to which acoustic signal data is input and a competing layer of an attribute map represented by a two-dimensional plane. Random Access Memory) etc. Thereby, acoustic signals having similar feature amounts are grouped to obtain an attribute map. As is well known, SOM is used for self-organizing neural network technology.

(2-6) SOM coordinate search unit 22
Then, the SOM coordinate search unit 22 and the feature amount vector F _INPUT (x, y, k) from the feature analysis unit 21 and the SOM information data (coordinate, feature amount vector F _SOM ( x, y, k) and group identification information), the group identification information corresponding to the feature vector F _INPUT (x, y, k) is retrieved.

The feature quantity vector F _INPUT (x, y, k) input to the SOM learning unit 26 and the feature quantity vector F _SOM (x, y, k) held in association with each coordinate on the two-dimensional plane. Both use the same format, and the number of elements (number of data) of both feature vectors match. Accordingly, the same processing block can be shared in the learning phase and the signal processing execution phase. As a result, the SOM coordinate search unit 22 performs speech or non-speech on the input acoustic signal based on the feature vector F _INPUT (x, y, k) and the feature vector F _SOM (x, y, k). Identify the presence or absence of reverberation.

(2-7) Signal processing type information output unit 24
Further, the signal processing type information output unit 24 displays the signal processing type information corresponding to the group identification information searched by the SOM coordinate search unit 22 among the three types of signal processing type information representing the processing type of the acoustic signal. Output. Here, the signal processing type information is, for example, as shown in Table 1, identification information of three types of signal processing such as compression amplification processing, formant emphasis processing, noise suppression processing, etc., amplification degree of the input acoustic signal, etc. Parameter information necessary for signal processing. Specifically, the function of the signal processing type information output unit is exhibited by the classification determination unit 24 (see FIG. 3 described later) or the parameter adjustment unit (see FIG. 9 described later). Formant emphasis means amplifying the amplitude of the maximum portion of the speech spectrum waveform.

(2-8) Acoustic signal processing unit 25
The acoustic signal processing unit 25 processes the input acoustic signal based on the signal processing type information output from the signal processing type information output unit (classification determination unit or parameter adjustment unit) 24.
(2-9) Amplifier 51
The amplifier 51 amplifies the acoustic signal processed by the acoustic signal processing unit 25 and outputs an amplified signal. The amplification degree of the amplifier 51 can amplify a signal at three types of amplification degrees A, B, and C, for example. The earphone 52 is for listening to the amplified signal from the amplifier 51 from the outside. Thus, the input acoustic signal is processed by the acoustic signal processing unit 25 so as to obtain a clear sound, and the user can obtain the processed speech through the earphone 52.

(2-10) SOM learning unit 26
Further, the SOM learning unit 26 groups the learning acoustic signals in the two-dimensional plane based on the feature vector F _INPUT (x, y, k) from the feature analysis unit 21 (groups by mapping to the two-dimensional plane). To do. The SOM learning unit 26 inputs a number of environmental sound signals as learning signals in advance before the acoustic signal is input to the feature analysis unit 21. The SOM information storage memory 23 analyzes the characteristics of each learning signal. Used to map to Here, environmental sound means sound in various environments such as various places and various times.

(3) Learning phase processing unit 27
Next, the learning phase processing unit 27 will be described with reference to FIG.
FIG. 2 is a block diagram of the learning phase processing unit 27 according to the first embodiment of the present invention. 2 having the same reference numerals as those shown in FIG. 1 are the same as those shown in FIG.

The feature analysis unit 21 outputs a feature quantity vector (feature analysis data obtained by quantizing the learning acoustic signal and feature analysis of the learning acoustic signal) representing the feature quantity of the learning acoustic signal. It functions as a feature quantity output unit. The feature analysis unit 21 converts an input acoustic signal into a spectrum region using FFT (Fast Fourier Transform) as a signal processing type, analyzes a spectrum waveform obtained by the conversion, and analyzes a power spectrum (power spectrum density). ) Level or waveform information is output as a feature vector F _INPUT (x, y, k).

  Furthermore, as the type of feature analysis processing, various analysis algorithms such as filter bank processing, linear prediction analysis processing, and mel cepstrum processing can be used in addition to FFT processing. Here, the filter bank processing uses the filter bank output as a feature amount. The filter bank processing is to perform spectrum conversion on the input acoustic signal and divide the spectrum band occupied by the input acoustic signal into a plurality of sub-bands. The linear prediction analysis process uses a linear prediction coefficient as a feature quantity and obtains a linear prediction coefficient using autocorrelation calculation. The mel cepstrum processing uses MFCC (Mel Filtered Cepstrum Coefficient) as a feature quantity, and calculates the logarithm of the power spectrum of the input speech and generates MFCC using Mel transformation and cosine transformation. .

  The input sound signal is an analog sound signal input from, for example, a microphone, and the analog sound signal is converted from analog to digital. A digital acoustic signal output from a speech decoding unit such as a mobile phone can be used. The input sound signal data is once held in a buffer, and the held input sound signal data is read by the feature analysis unit 21 at a certain time interval and subjected to feature analysis or feature extraction. The time required for this processing is called a frame (unit frame). Specifically, the input sound signal data is successively held in the buffer, and when the number of holds becomes, for example, 100, the feature analysis unit 21 reads 100 pieces of input sound signal data, and the feature analysis is performed. Therefore, a frame is a set of waveform data obtained by dividing waveform data of an acoustic signal at regular time intervals, and feature extraction is performed in units of frames.

  Furthermore, the feature analysis unit 21 can adjust the time width of the waveform required for frame processing according to the type of the input acoustic signal. For example, an acoustic signal having a relatively short time width of about 1 msec (milliseconds) to about 0.1 msec of a waveform has a small number of quantizations and is called a short time frame (short frame). About one is enough. On the other hand, a relatively long acoustic signal having a time width of, for example, about 1 sec to 2 sec is read into the feature analysis unit 21 by combining a plurality of short-time frames.

The feature vector F _INPUT (x, y, k) from the feature analysis unit 21 is also sent to the SOM learning unit 26 as one short frame or a plurality of short frames according to the type of the input acoustic signal. Entered. That is, the feature vector input to the SOM learning unit 26 can be expressed by vector data or a scalar value of one frame, or can be expressed by vector data obtained by combining parameters of a plurality of frames. Input / output.

This parameter refers to spectral characteristics, sound pressure level, time waveform, and the like.
More specifically, assuming that the number of divisions n is 16, for example, in one short frame, the feature vector F _INPUT (x, y, k) is obtained for each of the 16 divided times t _{0 to} t _15. It is generated by acquiring each element F _INPUT (x, y, 0), F _INPUT (x, y, 2),..., F _INPUT (x, y, 15). Further, the value of the time width TW can be variously changed depending on the case, and the time width TW can be several msec to several sec.

Thus, in the learning phase, the feature signal F _INPUT (x, y, k) is generated and generated from the acoustic signal input to the speech input unit 50 by the feature analysis for each short time frame in the feature analysis unit 21. The feature vector F _INPUT (x, y, k) is taken into an input buffer (not shown). The feature quantity vector F _INPUT (x, y, k) is expressed in a size corresponding to the frame length and stored in the SOM information storage memory 23.

(4) Coordinate, Grouping and SOM Information Storage Memory 23 Next, with reference to FIG. 4A and FIG. 4B, coordinates and grouping will be described, and FIG. 5A and FIG. An implementation example of the SOM information storage memory 23 will be described with reference to b).
FIG. 4A is a diagram illustrating an example of the SOM network according to the first embodiment of the present invention. The SOM learning unit 26 provides the SOM network shown in FIG. 4A at the address of the SOM information storage memory 23. In addition, the SOM network is provided with 10 × 10 = 100 coordinates (represented by circles and circles) each having 10 vertical and horizontal dimensions. The SOM information storage memory 23 holds a group to which each coordinate of each group belongs and a unique feature vector F _SOM (x, y, k) that is different between each coordinate in association with each other.

FIG. 4B is a diagram showing an example of two-dimensional plane grouping in the SOM information storage memory according to the first embodiment of the present invention. In the SOM network shown in FIG. 4B, for example, three types of groups A to C are generated. The SOM learning unit 26 also holds identification information for identifying the acoustic signal group to which each coordinate (represented by a circle) on the two-dimensional plane belongs together with the feature vector F _SOM (x, y, k). . For this reason, the SOM learning unit 26 holds group attribute information for each coordinate. Accordingly, the specific coordinate value and the coordinates near the two-dimensional plane hold group attribute information close to the group attribute information of the specific coordinate. Accordingly, when the two-dimensional plane is observed as a whole, the coordinates where the feature amount vectors F _INPUT (x, y, k) are close to each other are held close to each other on the two-dimensional plane. Based on the feature quantity, a feature quantity vector F _INPUT (x, y, k) is held.

FIG. 5A is a diagram showing an example of a memory area of the SOM information storage memory 23 according to the first embodiment of the present invention. The address 0x0000 of the SOM information storage memory 23 shown in FIG. 5A includes F _INPUT (x, y, 0) and the like held in association with the coordinates (1, 1) shown in FIG. Stores a pointer in which a feature vector is stored. Note that 0x represents a hexadecimal number, and the address value is an example. Similarly, at the addresses 0x0001 to 0x0073 in the SOM information storage memory 23, feature quantity vectors F _INPUT (x, y, k) held in association with the coordinates (1, 2) to the coordinates (10, 10), respectively. A pointer (vector pointer) is stored.

Accordingly, the SOM information storage memory 23 is a group of memory space addresses such as 0x0001 to 0x0073, the feature quantity vector F _SOM (x, y, k) assigned to the addresses, and a plurality of learning acoustic signals in the memory space. SOM information data in which the group identification information for identifying the plurality of groups is associated with each other is held.
(5) Feature Quantity Vector Holding Area FIG. 6 is a view for explaining a feature quantity vector F _INPUT (x, y, k) holding area according to the first embodiment of the present invention. In the memory area indicated by the pointer stored in the coordinates (1, 1) of the SOM information storage memory 23 shown in FIG. 6, group identification information to which 16 data (for example, spectrum values) and coordinates (1, 1) belong Is stored. Specifically, 16 elements F _INPUT (x, y + 1, 0), F _INPUT (x, y + 1, 1),..., F _INPUT (x, y + 1, 15) are held as spectral values. Similarly, 16 data and group identification information are respectively stored in the memory areas indicated by the pointers stored at coordinates (1, 2) to (10, 10). For example, the coordinates (10, 10) are stored at coordinates (10, 10). , 16 elements F _INPUT (x + 10, y + 10, 0), F _INPUT (x + 10, y + 10, 1),..., F _INPUT (x, y + 1, 15) are held as spectral values.

As described above, the acoustic signal processing device 20 according to the present invention has the coordinates (1, 1) to the coordinates (10, 10) on the two-dimensional plane and the feature vector F _INPUT (x, y, k) in the learning phase. And the group type are held in association with each other. When the learning phase is completed, as shown in FIG. 4B, the feature quantity vector F _SOM (x) associated with the coordinates (1, 1) to the coordinates (10, 10) at each address of the SOM information storage memory. , Y, k) and group attribute information is held in each address. That is, each coordinate is held in association with different unique feature vector data by a pointer (see FIG. 6). In the initial process in the learning phase, random values are set for the feature vectors.

(6) Coordinate Search Using SOM Next, coordinate search in the SOM network will be described with reference to FIG.
The SOM learning unit 26 includes the feature vector F _INPUT (x, y, k) input from the feature analysis unit 21 and the feature vector F _SOM (x, y, k) held in association with each coordinate. The Euclidean distance between each element is calculated, and the nearest coordinates closest to the input feature vector F _INPUT (x, y, k) (small Euclidean distance) are searched.

FIG. 7 is a diagram for explaining a search for neighboring coordinates according to the first embodiment of the present invention. The two-dimensional plane shown in FIG. 7 is a coordinate plane realized in the SOM information storage memory 23. Here, if the feature vector input at a desired time t is w (t) and the current feature vector at each coordinate is m _i (t) (i represents an index indicating the coordinate), w. The Euclidean distance between (t) and m _i (t) is represented by | w (t) −m _i (t) | (where “||” represents an absolute value). Therefore, the SOM learning unit 26 searches for i that minimizes the Euclidean distance | w (t) −m _i (t) | in order to obtain the nearest coordinates closest to the input feature vector.

Specifically, the coordinates (x, y) located at the center (represented by a white circle) of the two-dimensional plane shown in FIG. 7 are the time width TW for the waveform of the input acoustic signal having the time width TW. 16 divided time t _0, t _1, ..., each quantized component F _INPUT in _{t 15 (x, y, 0} ), ..., F INPUT (x, y, 15) an element F _INPUT of ( x, y, 0). First, the SOM coordinate search unit 22 uses, for example, (x-1, y, 0) as a search start coordinate, and is between F _INPUT (x, y, 0) and a coordinate (x-1, y, 0). The Euclidean distance is calculated and the calculated Euclidean distance is held in a buffer (not shown). Thereafter, the SOM coordinate search unit 22 repeats calculation and holding in the order of (x, y-1, 0), (x + 1, y, 0), (x, y + 1, 0), and further (x-1, y y + 1,0), (x-2, y), (x-1, y-1,0), (x, y-2), (x + 1, y-1,0), (x + 2, y), ( Repeat calculation and holding in the order of x + 1, y + 1). Then, the coordinate at which the minimum Euclidean distance among the stored Euclidean distances is obtained is acquired as the nearest coordinate closest to the coordinates (x, y, 0).

Next, the SOM learning unit 26 obtains the coordinates i _c (i = i _c ) at which the Euclidean distance is minimum, and then uses the following equations (1) to (3) to obtain coordinates near the coordinates i _c. The feature vector m _i (t) at is updated. Here, t represents time, and α (t) represents a learning rate coefficient described later.
m _i (t + 1) = m _i (t) + h _ci (t) [w (t) −m _i (t)] (1)
h _ci = α (t) (when i is in the vicinity of _ic ) (2)
h _ci = 0 (when i is not in the vicinity of i _c ) (3)
Here, whether or not i is in the vicinity of _ic is determined using Expression (4).

| a − a _c | <4 and | b − b _c | <4 (4)
Then, the SOM learning unit 26 uses the coordinates i that satisfy Equation (4) as the neighborhood coordinates. Note that the coordinates i are (a, b), and the coordinates of i _c are ( _ac , b _c ). Further, i in the formula (4) is a coordinate on a two-dimensional plane, a represents an x-axis value, and b represents a y-axis value.
Further, the learning rate coefficient α (t) shown in Expression (2) satisfies the condition of 0 <α (t) <1. Further, the learning rate coefficient α (t) is set as shown in, for example, Expression (5) using a function that decreases with time.

α (t) = α ₀ (1−t / TL) (5)
Α ₀ is an appropriately selected constant coefficient, and is set to about 0.3, for example. t represents the time from the start of learning or times t _{0 to} t ₁₅ , and TL represents the time required for learning.
Thereby, the SOM learning part 26 performs learning calculation shown in Formula (1) using many acoustic signals. For example, when a total of k acoustic signals are input to the SOM learning unit 26, the SOM learning unit 26 uses the feature parameters of one of the k acoustic signals at each time t _{0 to} t ₁₅ , The value held in the SOM information storage memory is updated. Since learning is performed at each time t _{0 to} t ₁₅ , learning requires time TL.

  Further, after the learning is completed, the SOM learning unit 26 holds the SOM network information (feature value vector values at each coordinate) in the SOM information storage memory. Therefore, the input acoustic signal is mapped to the coordinates on the virtual two-dimensional plane in the SOM information storage memory, and the coordinates on the two-dimensional plane are held in a state where each unique feature vector is assigned. This retained feature vector is specifically related to a specific sound, the feature vector of the sound “a” is stored in the desired coordinates, and “a” is stored in the other coordinates. The feature vector of sound is stored.

The acoustic signal waveform data w (t) held in the input buffer is subjected to feature analysis by the feature analysis unit 21 every subframe long time obtained by dividing the short time frame length into, for example, n (n represents a natural number). , N acoustic signal waveform data w (t ₀ ), w (t ₁ ),..., W (t _n−1 ) are generated and output. Here, t _{0 to} t _n-1 represent time, respectively, and in the following description, w (t ₀ ), w (t ₁ ),..., W (t _n-1 ) are expressed as w (t _j (J represents an integer of 0 to n−1). That is, the feature amount of the acoustic signal having the subframe length is output as a plurality of elements (scalar value w (t _j )). Accordingly, in the following description, these elements w (t _j ) are referred to as feature quantity vectors w (t). In other words, the feature analysis unit 21 outputs a feature quantity vector w (t) composed of n elements w (t _j ).

  Then, the feature vector generated by the feature analysis is input to the SOM learning unit 26 shown in FIG. 2 and learned using a self-organizing map (SOM). This learning is performed by inputting and processing various environmental sounds, and mapping the processed environmental sounds on a virtual two-dimensional plane. In addition, the SOM learning unit 26 assigns a unique feature vector to each coordinate on the two-dimensional plane. In addition, the SOM learning unit 26 desirably groups a large number of various environmental sounds before shifting to the signal processing execution phase.

As a result, after learning, an attribute map in which each input environmental sound is mapped to desired coordinates on the virtual two-dimensional plane is obtained, and a unique feature vector is displayed on the two-dimensional plane. Assigned to each coordinate.
Furthermore, the SOM learning unit 26 preferably divides the two-dimensional plane into regions so that each acoustic signal is appropriately grouped in the two-dimensional plane. This area division is performed based on the grouping of each environmental sound before moving to the signal processing execution phase. Specifically, the SOM learning unit 26 divides a two-dimensional plane into regions based on groups, and sorts and arranges a plurality of acoustic signals on the divided planes. Assign a signal processing mode.

An appropriate acoustic signal processing mode is determined for each grouped group by conducting a listening experiment or the like in advance. Thus, the second feature vector and grouping information are assigned to each coordinate on the two-dimensional plane.
Subsequently, the classification determination unit (determination unit) 24 determines a signal processing mode corresponding to the group identification information searched by the search unit among a plurality of signal processing modes representing the processing type of the acoustic signal. Thus, it functions as a signal processing type information output unit. The classification determination unit 24 determines a group corresponding to the searched neighboring coordinates. That is, it is determined to which group the neighboring coordinates are included. Further, the classification determination unit 24 selects an appropriate signal processing mode from among the plurality of signal processing modes based on the determined group. Then, the acoustic signal processing unit 25 processes the acoustic signal based on the signal processing mode selected by the classification determination unit 24.

(7) Acoustic signal processing method of the present invention With this, the acoustic signal processing method of the present invention analyzes the characteristics of the input acoustic signal waveform and identifies the current acoustic signal using the SOM from the feature amount (voice The sound signal processing mode such as signal amplification is switched based on the identified information.
Therefore, in the acoustic signal processing method of the present invention, the SOM learning unit 26 uses the feature amount vector representing the feature amount of the learned acoustic signal for the learning acoustic signal, and the feature amount assigned to the coordinates. SOM information data in which each vector is associated with group identification information for identifying, for example, three types of groups A to C grouped by the SOM learning unit 26 is generated (learning step). In this learning step, a plurality of parameters corresponding to each of the plurality of groups are acquired.

  Next, the SOM coordinate search unit 22 uses the feature amount from the feature amount output unit among the plurality of coordinates in the two-dimensional plane held in the SOM information data generated in the learning step for the feature amount vector in the learning step. Search for nearby coordinates located in the vicinity of the vector (search step). Specifically, the input acoustic signal (acoustic signal waveform) is subjected to feature analysis for each short time frame by the feature analysis unit 21, and the feature quantity vector subjected to the feature analysis is stored in each SOM information storage memory 23. Compared with the feature quantity vector assigned to the coordinates, the nearest coordinates closest to the inputted feature quantity vector among the many coordinates held on the two-dimensional plane are searched. Here, the same type of feature vector as that in the learning phase is used.

Further, the classification determination unit 24 outputs signal processing type information corresponding to the group identification information of the neighboring coordinates searched in the search step among the plurality of signal processing type information indicating the processing type of the acoustic signal (signal processing). Type information output step).
The acoustic signal processing unit 25 performs acoustic signal processing based on the signal processing type information output in the signal processing type information output step (acoustic signal processing step). In the acoustic signal processing step, one or more parameters acquired in the learning step are changed based on the group identification information assigned to the neighboring coordinates searched in the search step (change step). Furthermore, acoustic signal processing is performed using the one or more parameters changed in the changing step.

(8) Learning Phase Processing Operation With this configuration, the learning phase processing flow will be described with reference to FIG.
FIG. 8 is a flowchart for explaining the learning phase processing according to the first embodiment of the present invention. When the sound input unit 50 inputs the acoustic signal of the environmental sound to the feature analysis unit 21 (step A1), the feature analysis unit 21 performs feature analysis processing of the various methods described above (step A2). Then, the SOM coordinate search unit 22 searches for a coordinate on the SOM closest to the feature amount parameter (step A3). Thereafter, it is determined whether or not the update by the feature amount parameter of each coordinate has been performed, for example, by the method shown in Expression (1) (step A4).

Here, the SOM coordinate search unit 22 searches whether or not all environmental sounds have been processed (step A5). If the processing is completed, the SOM information is stored in the SOM information storage memory 23 through the YES route. Store (step A6). If the process remains in step A5, the process after step A1 is performed through the NO route.
In this way, when the learning phase is completed, the feature quantity vector used for comparison in the signal processing execution phase is stored in the SOM information storage memory 23.

(9) Signal processing execution phase processing unit 28
Next, the signal processing execution phase processing unit 28 will be described with reference to FIG.
FIG. 3 is a diagram for explaining the signal processing execution phase processing unit 28 according to the first embodiment of the present invention. The classification determination unit 24 shown in FIG. 3 determines a signal processing mode corresponding to the group identification information searched by the SOM coordinate search unit 22 among a plurality of signal processing modes representing the processing type of the acoustic signal. is there. An example of this acoustic signal processing mode is compression amplification processing, formant enhancement processing, noise suppression processing, etc., as shown in Table 1 described later. Further, the determination of the signal processing mode is performed based on the group identification information and the signal processing mode held in the signal processing mode holding memory provided in the classification determination unit 24.

In addition, what has the same code | symbol as what was shown in FIG. 3 and mentioned above represents the same thing as them. Further, the signal processing mode signal can be held in the conversion table shown in Table 1.
Thereby, the feature analysis unit 21 performs feature analysis of the acoustic signal waveform input from the microphone. The type of signal processing in the feature analysis unit 21 is the same as the type of signal processing in the learning phase. The main reason is that, for example, the subband width of the acoustic signal, the division width of the power spectrum of the acoustic signal, and the like are set in common. Therefore, the feature analysis processing in the signal processing execution phase is performed using the FFT processing, filter bank processing, linear prediction analysis processing, mel cepstrum processing, and the like used in the learning phase.

Next, the SOM coordinate search unit 22 searches for the SOM coordinates. Specifically, the SOM coordinate search unit 22 searches for a coordinate closest to the feature vector w (t) of the input acoustic signal in the two-dimensional plane (see FIG. 5) of the SOM information storage memory 23. That is, the SOM coordinate search unit 22 searches for i that minimizes the Euclidean distance | w (t) −m _i (t) | in order to obtain the nearest coordinates closest to the input feature vector. Here, w (t), m _i (t) and | w (t) −m _i (t) | are respectively feature quantity vectors F _INPUT (x, y, k), held in the learning phase feature vector F _SOM represents the Euclidean distance (x, y, k) and w and (t) m _i and (t).

And the signal processing classification information corresponding to the group identification information searched in the SOM coordinate search part 22 among the several signal processing classification information showing the processing classification of an acoustic signal is output. That is, the signal processing is selected based on the feature vector that is classified and held.
Next, the classification determination unit 24 determines the type of the acoustic signal used for the acoustic signal processing based on the group (group attribute) of the coordinate ic that minimizes the Euclidean distance | w (t) −m _i (t) |. Then, the acoustic signal processing mode is output to the acoustic signal processing unit 25. As the acoustic signal processing mode, for example, the modes shown in Table 1 can be used.

The acoustic signal processing unit 25 performs signal processing of the input acoustic signal using the acoustic signal processing mode instructed (or notified) from the classification determination unit 24, and outputs the signal to an earphone or the like via the amplifier 51.
Thereby, in the signal processing execution phase, signal processing is executed based on an attribute map in which acoustic signals having similar feature amounts are grouped.

In the signal processing execution phase, acoustic signal processing is always performed, but feature analysis, SOM coordinate search, and classification determination processing for determining the acoustic signal processing mode are performed at appropriate intervals (for example, 1 to 2 seconds). ).
(10) Processing Operation of Signal Processing Execution Phase Processing Unit 28 With such a configuration, the signal processing execution phase according to the first embodiment of the present invention will be described with reference to FIG.

  FIG. 9 is a flowchart for explaining the acoustic signal processing method according to the first embodiment of the present invention. In step B1, the classification determination unit 24 determines whether or not the acoustic signal processing mode (signal processing mode) needs to be confirmed when the acoustic signal processing mode is changed when the acoustic signal processing device 20 is activated or after the activation. If it is determined that the confirmation is necessary, the route is YES and the feature analysis unit 21 performs feature analysis processing (step B2). If the classification determining unit 24 determines in step B1 that the acoustic signal processing mode need not be confirmed, the classification determining unit 24 passes the NO route, holds the current signal processing mode in the table shown in Table 1 (step B6), and the acoustic signal processing unit. 25 executes acoustic signal processing (step B5).

When the feature analysis is performed (step B2), the SOM coordinate search unit 22 searches for the coordinates on the SOM plane closest to the feature parameter (step B3), and the classification determination unit 24 based on the searched coordinates. Inputs an acoustic signal processing mode to the acoustic signal processing unit 25 (step B4), and acoustic signal processing is performed (step B5).
After step B5, the classification determining unit 24 determines in step B7 whether or not the process has ended. When the process ends, the process passes through the YES route, and if the process has not ended. The process after step B1 is repeated again through the NO route.

As described above, according to the acoustic signal processing method of the present invention, signal processing is appropriately performed in accordance with the input acoustic signal, so that a stable and highly clear voice or acoustic signal is acquired, and the user's trouble is Pleasure is removed and high quality sound can be heard.
Further, in this manner, the level of the spectral region with a low noise level can be emphasized according to the spectral characteristics, and the level of the spectral region with a high noise level is emphasized, so the audibility is improved. For example, when the user goes out of the house, the noise level is suppressed to increase the voice level, and in a noisy environment, the user can hear a sound with high clarity.

(11) Search Method Using Region Division Method Next, the region division method will be described in detail. In the learning phase, after the learning calculation, in the last stage, the number of groups and the sound are selected and divided into regions. For the selection, the system designer (or SOM learning unit 26) determines a plurality of representative sounds for each group in advance. This sound is selected from the sound used for the learning calculation or the sound prepared separately from the learning calculation.

  Next, the acoustic signal processing device 20 inputs the plurality of environmental sounds in order, performs the same processing as in the signal processing execution phase, and calculates to which coordinates the input sound is mapped on the two-dimensional plane. The Then, a numerical value representing the group type is written as the group identification information of the coordinates, and this is repeatedly executed. Here, in the case where mapping is made to a coordinate where some numerical value has already been written, the numerical value indicating the type of sound group input this time is overwritten as the group identification information of that coordinate.

  As a specific example, in the learning phase, the acoustic signal processing device 20 sets the group identification information of three types of groups of sound, noise (such as a cleaner), and siren sound (such as an ambulance) to “1”. ”,“ 2 ”, and“ 3 ”. Then, when the acoustic signal input for learning is analyzed as speech by feature analysis, the feature quantity vector is mapped to, for example, coordinates (2, 3). In this case, the feature analysis unit 21 inputs “1” as the group identification information of the coordinates (2, 3) to the SOM coordinate search unit 22.

  In addition, after all sounds prepared for learning are input, unset coordinates for which group identification information has not been set are coordinates closest to the unset coordinates and group identification information is set. Group identification information of assigned coordinates is assigned. The search method for the unset coordinates (x, y, k) is such that the SOM coordinate search unit 22 uses (x-1, y, k), (x, y-1, k), (x + 1, y, k). , (X, y + 1, k), (x-1, y-1, k), (x-1, y + 1, k), (x + 1, y-1, k), (x + 1, y + 1, k) in this order. Then, it is searched whether group identification information is set for each coordinate. When the group identification information is not set for all coordinates in this search range, the SOM coordinate search unit 22 further performs (x−2, y, k), (x, y−2, k), (x + 2, The group identification information of each coordinate is searched in the order of y, k), (x, y + 2, k).

Therefore, the SOM coordinate search unit 22 corresponds to neighboring coordinates located in the vicinity of the feature quantity vector from the feature quantity output unit among the plurality of coordinates in the two-dimensional plane held in the SOM information data of the SOM information storage memory 23. Output group identification information.
Thus, since the group identification information for each coordinate is retrieved and processed, a clear acoustic signal can be obtained.

  As described above, according to the acoustic signal processing device 20 and the acoustic signal processing method of the present invention, various acoustic feature vectors are databased in the SOM information storage memory 23 in the learning phase, and in the signal processing execution phase. The inputted current acoustic signal is subjected to feature analysis. Then, the surrounding sound environment is identified based on the current feature vector of the acoustic signal and the feature vector stored in the SOM information storage memory 23, and the amplification characteristics and the like are adjusted according to the identified environment. Therefore, in any sound environment, stable and appropriate sound signal processing can be performed, and it is easy to hear and stable and highly clear sound can be reproduced.

(A1) Description of First Modification The acoustic signal processing unit 25 in the first embodiment has changed the processing type based on the acoustic signal processing mode input from the classification determination unit 24. In the acoustic signal processing method of this modification, instead of switching the acoustic signal processing mode, the processing type for the acoustic signal is changed and adjusted using a parameter related to signal processing such as the degree of amplification or amplification characteristics.

  FIG. 10 is a block diagram of a signal processing execution phase processing unit 28a according to a first modification of the first embodiment of the present invention. On the output side of the SOM coordinate search unit 22 shown in FIG. Part 30 is provided. The memory 31 stores group identification information and parameters in association with each other. This parameter is a value for identifying, for example, the amplification degree of the amplifier 51 necessary for the amplification process of the acoustic signal.

  The parameter adjustment unit 30 adjusts and outputs a parameter related to the setting value necessary for processing the acoustic signal based on the group identification information searched by the search unit, and includes a signal processing type information output unit. Is functioning as Then, the parameter adjustment unit 30 adjusts and outputs a parameter related to a setting value necessary for the amplification process of the acoustic signal based on the group identification information searched by the SOM coordinate search unit 22.

In other words, the parameter adjustment unit 30 serving as the signal processing type information output unit includes the SOM coordinate search unit 22 in the signal processing type information such as compression amplification processing, formant emphasis processing, or noise suppression processing that represents the processing type of the acoustic signal. The signal processing type information corresponding to the group identification information retrieved in is output.
Thereby, in the parameter adjustment part 30, the parameter for applying to an acoustic signal process is changed appropriately based on group identification information.

The memory 31 can also be provided in the parameter adjustment unit 30, in a block other than the parameter adjustment unit 30, or in a memory (buffer) in another block.
With such a configuration, in the acoustic signal processing method according to the first modification, in the learning phase, the parameter adjustment unit 30 learns appropriate parameters for each classified group, inputs from an external device, or is manually set. Or the like is previously stored in a memory or the like. In this state, the learning phase of the first modification basically performs the same processing as the acoustic signal processing in the first embodiment.

In the signal processing execution phase, the SOM coordinate search unit 22 determines to which group the coordinates on the SOM information storage memory 23 obtained from the input acoustic signal belong.
Next, unlike the processing in the first embodiment, the parameter adjustment unit 30 changes the parameters of the acoustic signal processing based on the obtained group information. Then, the acoustic signal processing is executed using the changed parameter.

In this way, in the first modified example, in addition to obtaining the effects of the first embodiment, the process can be simplified by using the parameter adjustment unit 30.
(A2) Description of Second Modification The acoustic signal processing device 20 of the second modification can use manual processing by a user's manual operation in addition to automatic processing of the acoustic signal processing device 20 itself.

  FIG. 11 is a block diagram of a signal processing execution phase processing unit according to a second modification of the first embodiment of the present invention. The signal processing execution phase processing unit 28b shown in FIG. A classification determination / parameter adjustment unit (classification determination and parameter adjustment unit) 30a and a mode / parameter forced change unit 32 are provided. Here, the classification information correction unit 29 can rewrite the SOM information data held in the SOM information storage memory 23. Further, the mode / parameter forced change unit 32 inputs an appropriate correction signal to the classification information correction unit 29 based on input data from the user.

  Further, the classification determination / parameter adjustment unit (classification determination and parameter adjustment unit) 30a determines an appropriate grouping (classification determination) and an appropriate amplification degree of the amplifier based on the coordinate data from the SOM coordinate search unit 22, for example. Are input to the acoustic signal processing unit 25. This function is, for example, a method of directly inputting information data used by a user for a mode or parameter, a method of inputting information data input via a wireless line or a wired line, or information data relating to various modes and parameters in advance. This is realized by using a ROM or the like to be stored.

  Here, the flow of correction of SOM information data will be described in further detail. Information data manually input by the user is read by the mode / parameter forced change unit 32, and data including the mode or parameter correction content (correction instruction data) is classified from the mode / parameter forced change unit 32. The information is stored in the SOM information storage memory 23 via the information correction unit 29. The mode or parameter changed by the mode / parameter forced change unit 32 is input to the audio signal processing unit 25. Therefore, in the signal processing execution phase, the processing mode / parameter can be forcibly changed based on an instruction using a user input operation or the like.

  As a result, when the user uses or listens to the acoustic signal processing device 20 and determines that the mode or parameter is not appropriate and performs a button press or the like, the classification determination / parameter adjustment unit 30a performs the button operation. Based on the input, it detects that the sound quality is not appropriate, and corrects the group identification information of the acoustic signal. Further, the classification attribute information correction unit starts correction upon detection of an input stop. Further, the classification attribute information correction unit holds the feature quantity vector, the signal processing type information at the time of the input operation, or the parameter associated with each group in the SOM information storage memory 23.

With such a configuration, in the acoustic signal processing method in the second modification of the first embodiment of the present invention, the processing in the learning phase is the same as the processing in each learning phase in the first embodiment and the first modification. .
In the signal processing execution phase in the second modified example, processing steps described below are added.

  Next, in the additional signal processing execution phase, when the sound being heard by the user at the time of executing the signal processing cannot be clearly heard, the button operation or the like is switched so that the user himself / herself becomes an appropriate mode / parameter. Here, when the classification determination / parameter adjustment unit 30a detects that the sound quality is not appropriate, the classification determination / parameter adjustment unit 30a immediately switches to the mode / parameter designated by the user. Subsequently, the classification determination / parameter adjustment unit 30a corrects both the group of the SOM coordinates at that time and the group of the peripheral coordinates of the SOM coordinates to a group corresponding to the mode / parameter selected by the user. . Further, the classification determination / parameter adjustment unit 30a temporarily holds the feature amount information of the acoustic signal input at the time of the correction, and temporarily holds the changed mode / parameter changed by the user.

Further, when the user stops the acoustic signal processing device 20 in the correction phase, the acoustic signal processing device 20 shifts to the correction phase, and when the shift to the correction phase is started, the classification determination / The parameter adjustment unit 30a corrects the group information.
Then, the classification determination / parameter adjustment unit 30a and the SOM coordinate search unit 22 cooperate to search for the closest coordinates in the two-dimensional plane that the temporarily stored feature amount is. Then, the classification determination / parameter adjustment unit 30a corrects the retrieved coordinates and surrounding group information to a group corresponding to the mode / parameter selected by the user.

Thus, in the signal processing execution phase, the processing mode / parameter is forcibly changed based on an instruction using a user input operation or the like.
Further, while the signal processing is stopped, the acoustic signal processing device 20 proceeds to the correction phase. In the correction phase, the SOM information is corrected according to the user response.
As described above, in the second modified example, the group identification information of the acoustic signal in the SOM information storage memory 23 is corrected by feedback from the user, and the mode / parameter switching information of the acoustic signal processing is used on the self organizing map. The group identification information in is corrected.

(A3) Third Modification In the third modification, an existing speech encoding processing unit is provided instead of the feature analysis unit 21 of the first embodiment.
FIG. 12 is a block diagram of a learning phase processing unit according to a third modification of the first embodiment of the present invention. In the learning phase processing unit 27a shown in FIG. 12, the feature quantity output unit 21a receives the speech coding parameters obtained by the speech coding process and inputs the speech coding parameters to the SOM information storage memory 23 as feature quantity vectors. Is configured as a speech encoding processing unit (existing speech encoding processing unit) 21a. The SOM learning unit 26 is input with encoding parameters relating to encoding of an acoustic signal and generated by the acoustic signal processing device 20 itself or the transmission side device.

In addition, it is the code | symbol respectively displayed in FIG. 12 and FIG. 13 demonstrated below, Comprising: The thing which has the same code | symbol as mentioned above represents the same thing as them.
Next, FIG. 13 is a block diagram of a signal processing execution phase processing unit according to a third modification of the first embodiment of the present invention. The signal processing execution phase processing unit 28c shown in FIG. 13 performs signal processing using the speech coding parameters obtained by the existing speech coding processing, and includes the speech coding processing unit 21a and the decoding A processing unit 35, an acoustic signal processing unit 25a, and a speaker 52a are provided. Here, the speech encoding processing unit 21a outputs demodulated information data and outputs encoding parameters obtained by the existing speech encoding processing. This encoding parameter is input to the SOM coordinate search unit 22 as a feature amount. The decoding processing unit 35 performs decoding processing on the information data output from the voice encoding processing unit 21a and outputs a received voice. The received voice is input to the acoustic signal processing unit 25a. . Furthermore, the acoustic signal processing unit 25a processes the input acoustic signal based on the signal processing type information output from the signal processing type information output unit (speech encoding processing unit) 21a, and performs speech enhancement processing and noise suppression. Processing is performed. The speaker 52a is used to sound the amplified signal from the amplifier 51.

  Therefore, the SOM coordinate search unit 22 is input with the encoding parameters related to the encoding of the acoustic signal and generated in the acoustic signal processing device 20 itself or the transmission side device. Furthermore, the speech input unit 50 and the speech encoding processing unit 21a serve as the encoded information input unit (50, 21a), and the encoding parameters generated in advance are input to the SOM coordinate search unit 22.

  Therefore, the feature quantity output unit 21a uses the encoding parameters related to the encoding of the acoustic signal and generated in the acoustic signal processing apparatus 20 itself or the transmission side apparatus as the feature quantity vector in the SOM information storage memory 23. The signal processing type information output unit outputs at least one of the signal processing type information and the encoding parameter based on the group identification information, and the acoustic signal processing unit (signal processing unit) 25a The signal processing type information output from the processing type information output unit is configured to execute different signal processing in accordance with each set value of the encoding parameter.

With such a configuration, in the learning phase, speech encoding parameters are output from the received data shown in FIG. 12, and in the SOM learning unit 26, the acoustic signals are classified and organized using the self-organizing map. Retained.
In the signal processing execution phase, the encoding parameters from the speech encoding processing unit 21a shown in FIG. 13 indicate which group the acoustic signal input using the SOM information storage memory 23 in the SOM coordinate search unit 22 belongs to. Is identified. Further, the identification information obtained in the classification determining unit 24 is switched in the signal processing mode or parameter in the acoustic signal processing unit 25a. Then, different acoustic signal processing is performed depending on the designated signal processing mode or set parameters.

As described above, since the acoustic signal processing device 20 is mounted and operated in cooperation with the existing speech encoding device, the generalization of the acoustic signal processing device 20 is promoted.
In this way, appropriate signal processing can be performed according to the input acoustic signal, and a sound with high clarity can be heard stably.
(B) Description of Second Embodiment of the Invention In the second embodiment, the acoustic signal processing device 20 is linked to an existing speech encoding device provided in a mobile phone or the like.

  FIG. 14 is a block diagram of a learning phase processing unit according to the second embodiment of the present invention. The learning phase processing unit 42 shown in FIG. 14 performs radio demodulation processing on the encoding parameter generated on the transmission side and inputs it to the SOM learning unit 26. For example, the learning phase processing unit 42 is provided in a receiving unit such as a mobile phone. This is an acoustic signal processing device 20. The learning phase processing unit 42 includes an antenna 40a that receives an RF (Radio Frequency) signal, an RF reception unit 40b that down-converts and demodulates the radio signal from the antenna 40a, and outputs the demodulated signal. A baseband signal processing unit 40c for extracting information data including speech coding parameters by performing baseband processing on the demodulated signal from the receiving unit 40b, and an SOM learning unit 26 and an SOM information storage memory 23 are provided. It is configured.

  Here, the baseband signal processing unit 40c that outputs speech coding parameters, the antenna 40a, and the RF receiving unit 40b function as a feature amount output unit and an encoded information input unit. That is, the learning phase processing unit 42 demodulates acoustic signal data wirelessly transmitted from a remote place instead of the feature analysis unit 21 of the first embodiment, and learns using the demodulated data. . Therefore, in the second embodiment, the SOM learning unit 26 is input with the encoding parameters related to the encoding of the acoustic signal and generated by the acoustic signal processing device 20 itself or the transmission side device.

14 and FIG. 15 described below, and those having the same reference numerals as described above represent the same elements.
FIG. 15 is a block diagram of a signal processing execution phase processing unit according to the second embodiment of the present invention. The signal processing execution phase processing unit 43 shown in FIG. 15 is provided in a receiving unit such as a mobile phone.

  With such a configuration, a radio signal is demodulated, and an encoding parameter obtained by an existing speech encoding process is input to the SOM coordinate search unit 22, and the SOM coordinate search unit 22 is based on the encoding parameter. Sound signals are classified and stored using a self-organizing map (SOM information storage memory) 23. Further, the classification determining unit 24 identifies which group the acoustic signal input using the SOM information storage memory 23 belongs to based on the encoding parameter, and the signal processing mode based on the obtained group identification information. Alternatively, the parameters are switched, and different acoustic signal processing is performed depending on the mode designation / parameter setting.

As described above, the acoustic signal processing device 20 can be incorporated in a speech encoding device provided in a mobile phone or the like, and can be implemented in various acoustic signal processing.
(C) Others The present invention is not limited to the above-described embodiments and modifications thereof, and various modifications can be made without departing from the spirit of the present invention.

The learning phase is not only performed at the time of product production, but also allows the user to perform the learning phase. In this case, the learning phase and the signal processing execution phase are switched at a predetermined timing.
The function of the encoding input unit in the second embodiment can be provided in a light receiving processing device or the like in optical fiber communication in addition to a mobile phone or the like.

The SOM information storage memory 23 can also generate coordinates in a multidimensional space.
(D) Supplementary Note (Supplementary Note 1) A feature value output unit that outputs first feature value data representing a feature value of an input acoustic signal;
Corresponding to the coordinates of the two-dimensional plane, the second feature amount data assigned to the coordinates, and group identification information for identifying a plurality of groups in which a plurality of learning acoustic signals are grouped in the two-dimensional plane A self-organizing map holding unit for holding the attached SOM information data;
The group identification information corresponding to the first feature quantity data based on the first feature quantity data from the feature quantity output unit and the SOM information data held in the self-organizing map holding unit A search section for searching for,
A signal processing type information output unit that outputs signal processing type information corresponding to the group identification information searched by the search unit among a plurality of signal processing type information representing a processing type of an acoustic signal;
An acoustic signal processing apparatus comprising: a signal processing unit that processes the input acoustic signal based on the signal processing type information output from the signal processing type information output unit.

(Appendix 2) The search unit
Corresponds to neighboring coordinates located in the vicinity of the first feature value data from the feature value output unit among a plurality of coordinates in the two-dimensional plane held in the SOM information data of the self-organizing map holding unit The acoustic signal processing device according to appendix 1, wherein the group identification information is output.

(Appendix 3) The search unit
Note that, based on the first feature value data and the second feature value data, the input acoustic signal is configured to discriminate between speech or non-speech and the presence or absence of reverberation. The acoustic signal processing apparatus according to 1 or 2
(Appendix 4) The self-organizing map holding unit
Supplementary notes 1 to 3, wherein the group to which each coordinate of the plurality of groups belongs and the second characteristic amount data unique to each coordinate are stored in association with each other. The acoustic signal processing device according to any one of appendix 3.

(Appendix 5) The self-organizing map holding unit
The memory space address, the second feature amount data assigned to the address, and group identification information for identifying a plurality of groups in which a plurality of learning acoustic signals are grouped in the memory space are associated with each other. The acoustic signal processing device according to any one of Supplementary Note 1 to Supplementary Note 4, wherein the acoustic signal processing device is configured to hold SOM information data.

(Additional remark 6) The feature-value output part which outputs the 1st feature-value data showing the feature-value of an input acoustic signal,
Corresponding coordinates in the multidimensional space, second feature amount data assigned to the coordinates, and group identification information for identifying a plurality of groups in which a plurality of learning acoustic signals are grouped in the multidimensional space A self-organizing map holding unit for holding SOM information data;
The group corresponding to the first feature quantity data based on the first feature quantity data output from the feature quantity output unit and the SOM information data held in the self-organizing map holding unit A search unit for searching for identification information;
A signal processing type information output unit that outputs signal processing type information corresponding to the group identification information searched by the search unit among a plurality of signal processing type information representing a processing type of an acoustic signal;
An acoustic signal processing apparatus comprising: a signal processing unit that processes the input acoustic signal based on the signal processing type information output from the signal processing type information output unit.

(Supplementary note 7) The signal processing type information output unit
Appendix 1 characterized by being configured as a determination unit that determines a signal processing mode corresponding to the group identification information searched by the search unit among a plurality of signal processing modes representing a processing type of an acoustic signal. The acoustic signal processing device according to any one of?
(Appendix 8) The signal processing type information output unit
Appendix 1 to Appendix, wherein the parameter adjustment unit is configured to adjust and output a parameter related to a set value necessary for processing an acoustic signal based on the group identification information searched by the search unit. The acoustic signal processing device according to any one of 7.

(Supplementary note 9) The signal processing type information output unit
It is configured to output the signal processing type information acquired using a table in which group identification information and signal processing type information are associated with each other. Acoustic signal processing device.
(Additional remark 10) The feature-value output part which outputs the 1st feature-value data showing the feature-value of a learning acoustic signal,
A learning unit that groups the learning acoustic signals in a two-dimensional plane based on the first feature value data from the feature value output unit;
SOM information data in which the coordinates of the two-dimensional plane, the second feature value data assigned to the coordinates, and group identification information for identifying a plurality of groups grouped by the learning unit are associated with each other. An acoustic signal processing apparatus comprising a self-organizing map holding unit for holding

(Additional remark 11) The feature-value output part which analyzes the learning acoustic signal waveform obtained by sampling a learning acoustic signal, and outputs the feature-value data of this learning acoustic signal,
A learning unit that maps and groups the learning acoustic signals on a two-dimensional plane based on the feature amount data from the feature amount output unit;
SOM in which the coordinates of the two-dimensional plane, the second feature value data uniquely assigned to each coordinate, and group identification information for identifying a plurality of groups grouped by the learning unit are associated with each other. An acoustic signal processing apparatus comprising a self-organizing map holding unit for holding information data.

(Supplementary Note 12) The feature output unit
A speech coding parameter obtained by speech coding processing is input, and the speech coding parameter is input to the self-organizing map holding unit as the first feature value data. The acoustic signal processing device according to any one of Appendix 1 to Appendix 11.

(Supplementary note 13) The feature quantity output unit uses an encoding parameter related to encoding of an acoustic signal, which is generated by the acoustic signal processing apparatus itself or the transmission side apparatus, as the first feature quantity data. While inputting into the self-organizing map holding unit,
The signal processing type information output unit outputs at least one of the signal processing type information and the encoding parameter based on the group identification information,
The signal processing unit is configured to execute different signal processing according to each set value of the signal processing type information output from the signal processing type information output unit and the encoding parameter. The acoustic signal processing device according to any one of appendix 1 to appendix 11.

(Supplementary note 14) Any one of Supplementary notes 1 to 13, further comprising a classification attribute information correction unit that corrects the SOM information data held in the self-organizing map holding unit according to an input operation. The acoustic signal processing device according to one.
(Supplementary Note 15) The classification attribute information correction unit
Appendix 14 characterized in that the first feature amount data, the signal processing type information at the time of the input operation, or a parameter associated with each group is held in the self-organizing map holding unit. The acoustic signal processing device described.

(Supplementary Note 16) The classification attribute information correction unit
The acoustic signal processing device according to appendix 15, wherein the acoustic signal processing device is configured to start correction upon detection of an input stop.
(Additional remark 17) About the learning acoustic signal based on the 1st feature-value data showing the feature-value of a learning acoustic signal, the 2nd feature-value data allocated to this coordinate, and this learning A learning step for generating SOM information data in association with each of group identification information for identifying a plurality of groups grouped in a section;
For the first feature value data in the learning step, the first feature value output unit from among the plurality of coordinates in the two-dimensional plane held in the SOM information data generated in the learning step. A search step for searching for neighboring coordinates located in the vicinity of the feature amount data of
A signal processing type information output step for outputting signal processing type information corresponding to the group identification information of the neighboring coordinates searched in the searching step among a plurality of signal processing type information representing a processing type of an acoustic signal;
An acoustic signal processing method comprising: an acoustic signal processing step for performing the acoustic signal processing based on the signal processing type information output in the signal processing type information output step.

(Supplementary Note 18) The learning step acquires a plurality of parameters corresponding to each of the plurality of groups,
The acoustic signal processing step includes a changing step of changing one or a plurality of parameters acquired in the learning step based on group identification information assigned to the neighboring coordinates searched in the searching step;
18. The acoustic signal processing method according to appendix 17, wherein the acoustic signal processing is performed using the one or more parameters changed in the changing step.

  According to the acoustic signal processing device and the acoustic signal processing method of the present invention, appropriate sound processing can be performed under various sound environments. Therefore, clear voice or the like regardless of the type of voice or non-voice can be obtained, and voice or the like can be easily heard according to the level of ambient noise, noise or reverberation. For example, the sound reproduced by the hearing aid becomes clear, and the user can more easily listen to various acoustic signals, thereby effectively assisting hearing. It can also be used for radio or television audio processing, allowing the user to hear clear and non-audio.

Furthermore, regardless of the quality and level of the environmental sound, the user can hear a clear sound corresponding to the change in the sound in each environment. For example, the user can hear the sound from which unpleasant abnormal sounds are removed.
In addition, according to the acoustic signal processing device and the acoustic signal processing method of the present invention, the acoustic signal processing device and the acoustic signal processing method can be provided in, for example, a mobile phone and the like, and can be used for various telephones, terminal devices, or apparatuses in general.

1 is a block diagram of an acoustic signal processing device according to a first embodiment of the present invention. It is a block diagram of the learning phase process part which concerns on 1st Embodiment of this invention. It is a figure for demonstrating the process part of the signal processing execution phase which concerns on 1st Embodiment of this invention. (A) is a figure which shows an example of the SOM network which concerns on 1st Embodiment of this invention, (b) is an example of the grouping of the two-dimensional plane in the SOM information storage memory which concerns on 1st Embodiment of this invention. FIG. (A) is a figure which shows an example of the memory area of the SOM information storage memory which concerns on 1st Embodiment of this invention, (b) is a figure for demonstrating the SOM coordinate which concerns on 1st Embodiment of this invention. is there. It is a figure shown for demonstrating the holding | maintenance area | region of the feature-value vector which concerns on 1st Embodiment of this invention. It is a figure for demonstrating the search of the near coordinate which concerns on 1st Embodiment of this invention. It is a flowchart for demonstrating the process of the learning phase which concerns on 1st Embodiment of this invention. It is a flowchart for demonstrating the acoustic signal processing method which concerns on 1st Embodiment of this invention. It is a block diagram of a signal processing execution phase processing part concerning the 1st modification of a 1st embodiment of the present invention. It is a block diagram of the signal processing execution phase process part which concerns on the 2nd modification of 1st Embodiment of this invention. It is a block diagram of the learning phase process part which concerns on the 3rd modification of 1st Embodiment of this invention. It is a block diagram of the signal processing execution phase process part which concerns on the 3rd modification of 1st Embodiment of this invention. It is a block diagram of the learning phase process part which concerns on 2nd Embodiment of this invention. It is a block diagram of the signal processing execution phase processing unit according to the second embodiment of the present invention.

Explanation of symbols

20 acoustic signal processing device 21 feature analysis unit (feature amount output unit)
21a Speech encoding processing unit (speech encoding processing unit)
22 SOM coordinate search unit 23 SOM information storage memory (self-organizing map holding unit)
24 Signal processing type information output unit (determination unit, classification determination unit)
25, 25a Acoustic signal processing unit 26 SOM learning unit (self-organizing learning unit)
27, 27a, 40 Learning phase processing unit 28, 28a, 28b, 28c, 43 Signal processing execution phase processing unit 29 Classification information correction unit 30 Parameter adjustment unit 30a Classification determination / parameter adjustment unit 31 Memory 32 Mode / parameter forced change unit 35 Decoding processing unit 40a Antenna 40b RF receiving unit 40c Baseband signal processing unit 50 Audio input unit 51 Amplifier 52 Earphone 52a Speaker

Claims (5)

A feature quantity output unit that outputs first feature quantity data representing a feature quantity of the input acoustic signal;
Corresponding to the coordinates of the two-dimensional plane, the second feature amount data assigned to the coordinates, and group identification information for identifying a plurality of groups in which a plurality of learning acoustic signals are grouped in the two-dimensional plane A self-organizing map holding unit for holding attached self-organizing map (SOM) information data (hereinafter referred to as SOM information data);
The group identification information corresponding to the first feature quantity data based on the first feature quantity data from the feature quantity output unit and the SOM information data held in the self-organizing map holding unit A search section for searching for,
A signal processing type information output unit that outputs signal processing type information corresponding to the group identification information searched by the search unit among a plurality of signal processing type information representing a processing type of an acoustic signal;
An acoustic signal processing apparatus comprising: a signal processing unit that processes the input acoustic signal based on the signal processing type information output from the signal processing type information output unit. A feature quantity output unit that outputs first feature quantity data representing a feature quantity of the input acoustic signal;
Corresponding coordinates in the multidimensional space, second feature amount data assigned to the coordinates, and group identification information for identifying a plurality of groups in which a plurality of learning acoustic signals are grouped in the multidimensional space A self-organizing map holding unit for holding SOM information data;
The group corresponding to the first feature quantity data based on the first feature quantity data output from the feature quantity output unit and the SOM information data held in the self-organizing map holding unit A search unit for searching for identification information;
A signal processing type information output unit that outputs signal processing type information corresponding to the group identification information searched by the search unit among a plurality of signal processing type information representing a processing type of an acoustic signal;
An acoustic signal processing apparatus comprising: a signal processing unit that processes the input acoustic signal based on the signal processing type information output from the signal processing type information output unit. The signal processing type information output unit
The parameter adjustment unit configured to adjust and output a parameter related to a set value necessary for processing an acoustic signal based on the group identification information searched by the search unit, The acoustic signal processing device according to claim 2. A feature amount output unit that outputs first feature amount data representing the feature amount of the learning acoustic signal;
A learning unit that groups the learning acoustic signals in a two-dimensional plane based on the first feature value data from the feature value output unit;
SOM information data in which the coordinates of the two-dimensional plane, the second feature value data assigned to the coordinates, and group identification information for identifying a plurality of groups grouped by the learning unit are associated with each other. An acoustic signal processing apparatus comprising a self-organizing map holding unit for holding Based on the first feature value data representing the feature value of the learning acoustic signal, the learning acoustic signal is grouped in the two-dimensional plane coordinates, the second feature value data assigned to the coordinates, and the learning unit. A learning step for generating SOM information data in association with each of group identification information for identifying a plurality of groups,
For the first feature value data in the learning step, the first feature value output unit from among the plurality of coordinates in the two-dimensional plane held in the SOM information data generated in the learning step. A search step for searching for neighboring coordinates located in the vicinity of the feature amount data of
A signal processing type information output step for outputting signal processing type information corresponding to the group identification information of the neighboring coordinates searched in the searching step among a plurality of signal processing type information representing a processing type of an acoustic signal;
An acoustic signal processing method comprising: an acoustic signal processing step for performing the acoustic signal processing based on the signal processing type information output in the signal processing type information output step.

Images