JP2011013383A - Audio signal correction device and audio signal correction method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an audio signal correction device excellent in improving quality of input audio signal.SOLUTION: The audio signal correction device includes: a feature parameter calculation means for calculating a feature parameter from the input audio signal; a collating means for collating the feature parameter with a plurality of reference models, to calculate likelihood that the feature parameter belongs to a predetermined reference model in the plurality of reference models; a first determination means which determines that a kind of the audio signal indicated by the reference model is a kind of the audio signal indicated by the feature parameter, when the likelihood satisfies a threshold condition; a second determination means which determines a mixing degree of a voice component and a non-voice component from the feature parameter, when the likelihood does not satisfy the threshold condition; and a correction means for correcting the input audio signal according to the kind of the audio signal determined by the first determination means, and according to the mixing degree.

Description

  The present invention relates to an audio signal correction apparatus and an audio signal correction method for correcting an audio signal to improve sound quality.

  In recent years, attention has been focused on high-quality and high-quality AV equipment, and research and development of such AV equipment has become active. For example, a technique is disclosed in which a specific signal is extracted from an input audio signal and signal processing is performed on the extracted specific signal to change a sense of distance or a sense of spread (see Patent Document 1). Specifically, a human voice specific signal included in the input audio signal is extracted, and the sense of distance is changed by a process of changing the frequency characteristics and volume of the extracted specific signal. In addition, a specific signal such as cheers or applause included in the input audio signal is extracted, and, for example, surround processing is performed on the extracted specific signal to change the sense of spread.

JP 2007-67858 A

  In the input audio signal processing technique disclosed in Patent Document 1, a specific signal is extracted from an input audio signal, a correction method is determined according to the extracted specific signal, and high sound quality is achieved.

  However, since the input audio signal contains signal components of various sound sources, it is not easy to improve the sound quality of the input audio signal. Simply determining the correction method in accordance with the extracted specific signal may prevent the selection of an appropriate correction method, making it difficult to achieve high sound quality.

  An object of the present invention is to provide an audio signal correction apparatus and an audio signal correction method that are excellent in improving the sound quality of an input audio signal.

  An audio signal correction apparatus according to an embodiment of the present invention divides an input audio signal into a plurality of frames in a predetermined time unit, and calculates a difference signal between an L channel component signal and an R channel component signal included in each divided frame. Difference signal calculating means, feature parameter calculating means for calculating one or more feature parameters from each difference signal calculated from each divided frame, the feature parameter and a plurality of reference models are collated, and the feature parameter is The feature parameter indicates the type of audio signal indicated by the predetermined reference model when the likelihood satisfies the threshold condition, and the matching means for calculating the likelihood belonging to the predetermined reference model among the plurality of reference models A first determination unit configured to determine the type of the audio signal; and the feature parameter when the likelihood does not satisfy the threshold condition. A second determination unit that determines the degree of mixing of the audio component and the non-speech component, and the input audio signal is corrected in accordance with the type of the audio signal determined by the first determination unit, and in accordance with the degree of mixing. Correction means for correcting the input audio signal.

  An audio signal correction method according to an embodiment of the present invention divides an input audio signal into a plurality of frames in a predetermined time unit, and calculates a difference signal between an L channel component signal and an R channel component signal included in each divided frame. Then, one or more feature parameters are calculated from each difference signal calculated from each divided frame, the feature parameters are compared with a plurality of reference models, and the feature parameter is a predetermined reference in the plurality of reference models. The likelihood belonging to the model is calculated, and when the likelihood satisfies the threshold condition, the type of the audio signal indicated by the predetermined reference model is determined as the type of the audio signal indicated by the feature parameter, and the determined audio When the input audio signal is corrected according to the type of signal and the likelihood does not satisfy the threshold condition, the feature parameter Determining a mixed degree of Luo speech component and a non-speech component, correcting the input audio signal according to the degree of mixing.

  ADVANTAGE OF THE INVENTION According to this invention, the audio signal correction apparatus and audio signal correction method excellent in the quality improvement of the input audio signal can be provided.

1 is a block diagram showing a schematic configuration of an audio signal correction apparatus according to an embodiment of the present invention. It is a block diagram which shows schematic structure of a real-time characteristic analysis module. It is a figure which shows an example of distribution of a characteristic parameter. It is a flowchart which shows an example of characteristic parameter generation. It is a flowchart which shows an example of determination of a signal sound classification. It is a flowchart which shows an example of audio | voice reference model correction | amendment.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram showing a schematic configuration of an audio signal correction apparatus according to an embodiment of the present invention. As shown in FIG. 1, the audio signal correction apparatus includes a real-time characteristic analysis module 1, a selection module 2, and a sound correction module 3. Furthermore, the sound correction module 3 includes a cheering scene input module 31, a live commentary scene input module 32, other scene input modules 33, a first sound quality correction module 34, a second sound quality correction module 35, and a third sound quality correction module 36. Yes.

  In the present embodiment, the audio correction module 3 is configured to appropriately correct the audio signal according to the characteristics of the audio signal of the content such as sports broadcast, but the present invention is limited to this configuration. It is not something. For example, when correcting an audio signal of another content, the audio correction module 3 is configured to appropriately correct the audio signal according to the characteristics of the audio signal of the other content.

  The real-time characteristic analysis module 1 extracts feature parameters from the input audio signal, and analyzes the type of audio signal (scene) in real time from the extracted feature parameters. For example, in the case of content such as sports broadcasts, the real-time characteristic analysis module 1 determines, in real time, the audio signal of the cheering scene, the audio signal of the live commentary scene, and the audio signal of other scenes (for example, normal play) Scene audio signal).

  The selection module 2 selects the output destination of the input audio signal according to the analysis result that changes in real time. For example, the selection module 2 selects an input to the cheering scene input module 31 based on detection of the audio signal corresponding to the cheering scene, and inputs the input audio signal to the cheering scene input module 31. The selection module 2 selects an input to the live commentary scene input module 32 based on detection of the audio signal corresponding to the live commentary scene, and inputs the input audio signal to the live commentary scene input module 32. The selection module 2 selects an input to the other scene input module 33 based on detection of an audio signal corresponding to the other scene (for example, an audio signal corresponding to the normal play scene), and the input audio signal is input to the other scene input module. Input to 33. The selection module 2 selects an input to the cheering scene input module 31, selects an input to the live commentary scene input module 32, or other to the scene input module 33 according to the analysis result that changes in real time. Or select the input.

  The first sound quality correction module 34 corrects the input audio signal from the cheering scene input module 31 by a correction method corresponding to the cheering scene. For example, in a cheering scene (including support, booing, etc.), if the environmental sound is much louder than the live sound so that the live commentary sound becomes difficult to hear, the first sound quality correction module 34 greatly increases the sound component contained in the input audio signal. To make it easier to hear the actual situation.

  The second sound quality correction module 35 corrects the input audio signal from the live commentary scene input module 32 using a correction method corresponding to the live commentary scene. For example, in the live commentary scene, the live sound tends to be louder than the environmental sound, but in order to make it easier to hear the live sound, the second sound quality correction module 35 slightly emphasizes the sound component included in the input audio signal. , Make the actual situation easier to hear.

  The third sound quality correction module 36 corrects the input audio signal from the other scene input module 33 by a correction method corresponding to the other scene (for example, a normal play scene). For example, in the normal play scene, the sound in the field is the main, and the third sound quality correction module 36 corrects the input audio signal for wide stereo in order to enhance the presence.

  Next, the above-described real-time characteristic analysis module 1 will be described with reference to FIG. FIG. 2 is a block diagram showing a schematic configuration of the real-time characteristic analysis module 1. As shown in FIG. 2, the real-time characteristic analysis module 1 includes an LR difference signal calculation module 11, a feature parameter calculation module 12, a reference model matching module 13, a threshold determination module 14, and a voice / non-voice mixing degree determination module 15. Yes. Further, the feature parameter calculation module 12 includes a power fluctuation calculation module 121, a zero-crossing frequency calculation module 122, a spectrum fluctuation calculation module 123, and the like.

  The LR difference signal calculation module 11 divides an input audio signal into a plurality of frames in a predetermined time unit, and calculates a difference signal between an L (Left) channel component signal and an R (Right) channel component signal included in each divided frame. To do. For example, the LR difference signal calculation module 11 divides the input audio signal into a plurality of frames in units of about several hundreds msec, further divides each frame into subframes in units of about several tens of msec, and LRs from each subframe. A difference signal is calculated.

  The feature parameter calculation module 12 calculates a plurality of feature parameters from the LR difference signal calculated from each subframe. For example, the power fluctuation calculation module 121 calculates power in units of subframes. The zero crossing frequency calculation module 122 calculates a zero crossing frequency (the number of times the time waveform crosses zero in the amplitude direction) in units of subframes. The spectrum fluctuation calculation module 123 calculates the spectrum fluctuation (FFT power spectrum component fluctuation between the predetermined subframe and the immediately preceding subframe) in units of subframes. In the present embodiment, a case where the feature parameter calculation module 12 calculates three parameters of power, zero-crossing frequency, and spectrum variation will be described, but the present invention is not limited to this. For example, the feature parameter calculation module 12 can calculate n parameters. That is, the feature parameter calculation module 12 can be configured by n parameter value calculation modules from the first parameter calculation module to the n-th parameter value calculation module. The feature parameter calculation module 12 obtains statistics such as power, zero-crossing frequency, average of spectrum variation, and variance in units of frames, and calculates feature parameters based on the statistics.

  The reference model collation module 13 collates predetermined feature parameters with a plurality of reference models. For example, the reference model matching module 13 holds a reference model (distribution data) for each sound type (by scene) that is constructed in advance by learning, compares a predetermined feature parameter with each reference model, and determines whether the predetermined feature parameter is The probability that belongs to a predetermined reference model, that is, the likelihood is calculated.

  The threshold determination module 14 determines that the type of the audio signal (scene) indicated by the predetermined reference model is the type of the audio signal (scene) indicated by the predetermined feature parameter when the calculated likelihood satisfies the threshold condition. That is, the threshold determination module 14 can determine the type of the audio signal (scene) when the calculation likelihood satisfies the threshold condition.

  The voice / non-voice mixing degree determination module 15 determines the degree of mixing of the voice component and the non-voice component when the calculated likelihood does not satisfy the threshold condition, and the voice correction module 3 determines the input audio signal based on the degree of mixing. to correct. For example, the third sound quality correction module 36 controls the correction of the input audio signal based on the degree of mixing.

  Here, the cooperation between the threshold determination module 14 and the voice / non-voice mixing degree determination module 15 will be described. For example, the reference model matching module 13 matches the feature parameter with each reference model, calculates each likelihood that the feature parameter belongs to each reference model, and selects the maximum likelihood from the calculated likelihoods. . In this case, when the maximum likelihood satisfies the threshold condition, the threshold determination module 14 uses the audio signal (scene) whose feature parameter indicates the type of the audio signal (scene) indicated by the reference model used for calculating the maximum likelihood. ) Type. That is, the type of audio signal (scene) indicated by the feature parameter is determined. However, when the maximum likelihood does not satisfy the threshold condition, the threshold determination module 14 cannot determine the type of the audio signal (scene) indicated by the feature parameter. Therefore, the voice / non-voice mixing degree determination module 15 functions to determine the type of audio signal (scene) that could not be determined by the threshold determination module 14. That is, the voice / non-voice mixing degree determination module 15 determines (estimates) the degree of mixing of the voice component and the non-voice component based on the maximum likelihood, and the voice correction module 3 corrects the input audio signal based on the degree of mixing. To control.

  As described above, the selection module 2 and the sound correction module 3 correct the input audio signal according to the type of the audio signal (scene) determined in real time, or the sound according to the mixing degree determined in real time. The correction module 3 controls the correction degree of the input audio signal. Thereby, the audio signal correction apparatus can output an audio signal corrected in real time according to the type of the audio signal (scene).

  Next, each process described above will be described in more detail with reference to the flowcharts shown in FIGS. FIG. 4 is a flowchart illustrating an example of feature parameter generation. FIG. 5 is a flowchart illustrating an example of determination of the signal sound type. FIG. 6 is a flowchart illustrating an example of speech reference model correction.

  As shown in FIG. 4, the LR difference signal calculation module 11 divides the input audio signal into a plurality of frames in units of about several hundreds of milliseconds, and further divides each frame into subframes in units of about several tens of milliseconds ( S101), an LR difference signal is generated from each subframe (S102). In general, since human voice has a property of being localized toward the center, generation of an LR difference signal suppresses voice components. This makes it easy to identify a specific sound type such as cheers.

  Subsequently, the power fluctuation calculation module 121 calculates power in units of subframes (S103), the zero crossing frequency calculation module 122 calculates zero crossing frequencies in units of subframes (S104), and a spectral fluctuation calculation module. 123 calculates the spectral variation in units of subframes (S105), and further determines other discrimination information as necessary (S106). The feature parameter calculation module 12 obtains statistics such as power, zero-crossing frequency, average of spectrum variation, and variance in units of frames, and calculates feature parameters based on the statistics (S107, S108).

  Subsequently, the signal sound type (scene type) is determined (S200). As shown in FIG. 5, the reference model collation module 13 collates the feature parameter with the reference model of each sound type (S201). An example of reference model construction is GMM (Gaussian Mixture Model). A reference model using GMM is expressed by a mixed sum of normal distributions having dimensions of the number of feature parameters.

That is, as shown in the following (Formula 1), x is a D-dimensional vector of feature parameters x = {x1, x2,. A matrix, αk is a weight of each distribution when a plurality of normal distributions are weighted and expressed as a linear sum.

  N calculated by these represents the likelihood and is a mixed sum of the D-dimensional feature parameter normal distribution. If a feature parameter in a certain frame is generated from a mixed Gaussian distribution which is a reference model, the likelihood of the signal is obtained by the above (Equation 1), and this likelihood value is usually logarithmized. The feature parameter is compared with each reference model, the likelihood that the feature parameter belongs to each reference model is calculated, the maximum likelihood is selected from each calculated likelihood, and this is used to calculate the maximum likelihood The determined reference model is set as a candidate for determination.

  Here, the reliability of the reference model may differ depending on the sound type of the reference model or the amount of learning data for obtaining the reference model. A reference model of a sound type that is likely to have a stationary property in a spectrum structure or the like, such as a big cheer or music, is ideal with a small deviation of normal distribution and high convergence. For example, as shown in FIG. 3, cheers are sound types that are likely to have stationary properties, and the reliability of the reference model constructed from the distribution of feature parameters corresponding to cheers is relatively high.

  On the other hand, in the reference model of the sound type including the environmental sound peculiar to each content as in the play scene and the reference model of the sound type such as human speech, the stationary property does not appear so much and the spread of the distribution tends to be large. Not only that, it is difficult to collect sufficient learning data for constructing a reference model of sound types including environmental sounds peculiar to each content. A reference model of a sound type such as speech is relatively difficult to construct ideally due to various noise components depending on the recording environment and the influence of a minute silence (silence) section included in the speech. For example, as shown in FIG. 3, the play scene and the voice do not show much stationary properties, and the spread of the distribution tends to increase, and the reference model constructed from the distribution of feature parameters corresponding to the play scene and the voice The reliability of the reference model constructed from the distribution of corresponding feature parameters is relatively low.

  Therefore, it is important to determine the type of audio signal (scene) by the threshold determination module 14 and the mixing degree determination by the sound / non-voice mixing degree determination module 15. The distribution deviation of a reference model of a sound type that is likely to show stationary properties such as cheers (including cheering noises and boos) is small. For this reason, when the feature parameter of cheers and the reference model of cheers are collated, the collation result is close to the expected value (average value), and the likelihood tends to increase. Therefore, it is possible to determine a sound type such as cheers relatively easily. After collating the feature parameter with the reference model, the calculated likelihood is compared with the threshold value α, and if frames having a likelihood exceeding the threshold value α continue continuously β times or more (S202, YES) (S203, YES) At this time, the sound type to which the input audio signal belongs is determined as cheering (S204).

  At this time, if the frame likelihood when confirmed as cheering, or the likelihood information smoothed by multiple frames that lasted more than β times in the past, the reliability of information that is cheering can be conveyed more accurately, Fine correction control is possible for sound quality correction in the subsequent stage.

  When the likelihood that exceeds the threshold α cannot be obtained (S202, NO), or the frames having the likelihood that exceeds the threshold α do not continue β times or more continuously (S203, NO), the audio signal ( It is difficult to determine the exact type of scene. That is, the mixing degree determination by the voice / non-voice mixing degree determination module 15 is applied to input audio signals other than those determined as cheers in the threshold determination module 14 (S300).

  As shown in FIG. 6, in the actual content (input audio signal), the audio signal component and the play scene signal component are often not clearly separated, and simply the input audio signal is audio. Alternatively, the determination that the input audio signal is a play scene is not appropriate. Therefore, the degree of likelihood (mixing degree) that is a non-voice component with respect to the probability that is a sound component is output as a parameter, and sound quality correction such as wide stereo is controlled according to the degree of mixing.

  The speech / non-speech mixing degree determination module 15 includes a first likelihood obtained by matching the feature parameter with the speech reference model and a second likelihood obtained by matching the feature parameter with the non-speech reference model. The above-described degree of mixing is calculated from the difference between the degrees or the ratio between the first likelihood and the second likelihood.

  However, since the feature parameter is extracted from the LR difference signal, there is a possibility that the degree of mixing cannot be calculated accurately. That is, in the LR difference signal, a voice component that is easily localized at the center is suppressed. For this reason, the background sound is more easily reflected in the feature parameter distribution than the sound, and it is difficult to accurately calculate the degree of mixing of the sound and the play scene. On the other hand, if the D-dimensional feature parameter is extracted from the PCM signal before generating the LR difference signal, the processing load is doubled.

  Therefore, the voice / non-voice mixing degree determination module 15 detects a voice / non-voice mixing degree by paying attention to a simple feature parameter. For example, the voice / non-voice mixing degree determination module 15 detects the voice / non-voice mixing degree by pitch estimation. Pitch is the fundamental frequency that appears due to vocal cord vibrations in human (voiced) utterances. Generally, the value is around 100Hz for men and around 200Hz for women. There are various pitch estimations, but simple extraction includes autocorrelation in the time domain.

As shown in the following (formula 2), the difference in amplitude between the signal s (n) of a certain frame and the signal s (n + τ) shifted by time τ is accumulated for N frames. Τ that minimizes ε (τ) is set as a basic period, and its reciprocal is regarded as an estimated pitch value (S301).

  By checking whether or not the estimated pitch value in each subframe is present in the general human voice pitch range {lowPitch, uppPitch}, the existence ratio c1 in the frame is calculated (S302).

  Further, the reliability of pitch estimation itself is not necessarily high. For this reason, the voice / non-voice mixing degree determination module 15 can strengthen the reliability of this parameter by paying attention to other voice characteristics. For example, the voice / non-voice mixing degree determination module 15 detects the voice / non-voice mixing degree based on the number of zero crossings of the signal. If the switching between consonants and vowels is frequent like speech, the number of zero crossings varies greatly. Therefore, if the zero-crossing number variability is large and the estimated pitch value is within the standard human voice pitch range, the probability that the signal is voice is high. Here, regarding the zero-crossing number variability, when the value of the average zero-crossing number in the frame is defined as meanZC, the ratio c2 of subframes in which the calculated value exceeds meanZC is calculated (S303).

  The likelihood of speech is corrected using the above ratios c1 and c2. The likelihood obtained by collation with the GMM model is expressed by (Equation 3), and this likelihood is usually logarithmized.

N '= ln (N) (Formula 3)
At this time, for the logarithmic likelihood Nsp ′ output from the speech model, a correction term shown in the following (Equation 4) is introduced with respect to the likelihood before logarithmization (S304).

Nsp '= ln (Nsp + c1 * c2) (Formula 4)
From this Nsp 'and logarithmic likelihood Nnsp' output from the non-speech model (as in equations 3 and 4, the likelihood Nnsp obtained by matching with the GMM non-speech model is corrected by c1 and c2) Then, the degree of mixture of the non-speech probability with respect to the speech likelihood is determined as shown in (Equation 5) below, and is output (S305).

Ratio = Nnsp'−Nsp ′ (Formula 5)
As described above, for sound types such as speech and non-speech that cannot be easily matched with the reference model, information that quantifies the degree of mixture of the probability of speech and non-speech components is output, and the sound quality correction processing in the subsequent stage Then, the degree of correction is controlled based on this information.

  Here, the effects of the above-described audio signal correction apparatus will be summarized.

  (1) The audio signal correction apparatus can finely control sound quality correction according to the type of audio signal (scene) in real time.

  (2) The audio signal correction device extracts a feature parameter from the LR difference signal, determines a sound type such as a cheer with high reliability of the reference model based on a result of matching the feature parameter and the reference model, and determines a sound type Sound quality correction is controlled based on the result. By using the characteristic parameter extracted from the LR difference signal, it is possible to improve the accuracy of determining the sound type such as cheers.

  (3) The audio signal correction device obtains the mixing degree of the voice / non-voice component with low reliability of the reference model, and controls the sound quality correction according to the mixing degree.

  (4) The audio signal correction apparatus detects the degree of voice / non-voice mixing from simple feature quantities such as pitch estimation and the number of zero crossings of the signal. As a result, the audio signal correction apparatus can estimate the degree of voice / non-voice mixing with a small amount of processing, and even if the feature parameter is extracted from the LR difference signal (even if the voice component is suppressed), The degree of voice / non-voice mixing can be detected with high accuracy.

  As described above, the audio signal correction apparatus can finely perform sound quality correction according to multi-classification scenes in real time for various contents including sports. The audio signal correction apparatus can also apply the sound quality correction described above when reproducing recorded content.

  Furthermore, the audio signal correction apparatus can be applied to, for example, a TV, a DVD player, a DVD recorder, and an HDD recorder. Thereby, the user can enjoy the program broadcast or the DVD content with high sound quality according to the scene.

  The module described above may be realized by hardware, or may be realized by software using a CPU or the like.

  In addition, this invention is not limited to the said embodiment, In the implementation stage, it can change variously in the range which does not deviate from the summary. In addition, the embodiments may be appropriately combined as much as possible, and in that case, the combined effect can be obtained. Further, the above embodiments include inventions at various stages, and various inventions can be extracted by appropriately combining a plurality of disclosed constituent elements. For example, even if some constituent requirements are deleted from all the constituent requirements shown in the embodiment, the problem described in the column of the problem to be solved by the invention can be solved, and the effect described in the column of the effect of the invention Can be obtained as an invention.

DESCRIPTION OF SYMBOLS 1 ... Real time characteristic analysis module, 11 ... LR difference signal calculation module, 12 ... Feature parameter calculation module, 121 ... Power fluctuation calculation module, 122 ... Zero crossing frequency calculation module, 123 ... Spectral fluctuation calculation module, 13 ... Reference model collation module , 14 Threshold determination module, 15 Voice / non-voice mixing degree determination module, 2 Selection module, 3 Voice correction module, 31 Cheer scene input module, 32 Live commentary scene input module, 33 Other scene input module 34 ... First sound quality correction module, 35 ... Second sound quality correction module, 36 ... Third sound quality correction module

Claims (5)

Differential signal calculating means for dividing an input audio signal into a plurality of frames in a predetermined time unit and calculating a differential signal between the L channel component signal and the R channel component signal included in each divided frame;
Feature parameter calculating means for calculating one or more feature parameters from each difference signal calculated from each divided frame;
Collating means for collating the feature parameter with a plurality of reference models, and calculating a likelihood that the feature parameter belongs to a predetermined reference model among the plurality of reference models;
First determination means for determining that the type of the audio signal indicated by the predetermined reference model is the type of the audio signal indicated by the feature parameter when the likelihood satisfies a threshold condition;
Second determination means for determining a mixing degree of a speech component and a non-speech component from the feature parameter when the likelihood does not satisfy the threshold condition;
Correcting means for correcting the input audio signal according to the type of the audio signal determined by the first determining means, and correcting the input audio signal according to the degree of mixing;
An audio signal correction apparatus comprising: The collation means collates the feature parameter with each reference model, calculates each likelihood that the feature parameter belongs to each reference model, and selects the maximum likelihood from among the calculated likelihoods,
The first determination unit determines that the type of the audio signal indicated by the reference model used for calculating the maximum likelihood is the type of the audio signal indicated by the feature parameter when the maximum likelihood satisfies a threshold condition. And
The second determination means determines the mixing degree of the speech component and the non-speech component of the feature parameter based on the maximum likelihood when the maximum likelihood does not satisfy a threshold condition.
The audio signal correcting apparatus according to claim 1, wherein   2. The audio according to claim 1, wherein the second determination unit determines a degree of mixing of the voice component and the non-voice component of the characteristic parameter based on a fundamental frequency that appears due to vibration of a human vocal cord. Signal correction device.   The second determination means determines the degree of mixing of the speech component and the non-speech component of the feature parameter based on the variability of the number of zero crossings generated in response to switching between a consonant and a vowel. The audio signal correction apparatus described in 1. Dividing the input audio signal into a plurality of frames in a predetermined time unit, calculating a difference signal between the L channel component signal and the R channel component signal included in each divided frame;
Calculating one or more feature parameters from each difference signal calculated from each divided frame;
Collating the feature parameter with a plurality of reference models, calculating a likelihood that the feature parameter belongs to a predetermined reference model among the plurality of reference models;
When the likelihood satisfies a threshold condition, the type of the audio signal indicated by the predetermined reference model is determined as the type of the audio signal indicated by the feature parameter, and the input audio is determined according to the determined type of the audio signal. Correct the signal,
When the likelihood does not satisfy the threshold condition, the degree of mixing of the speech component and the non-speech component is determined from the feature parameter,
Correcting the input audio signal according to the degree of mixing;
An audio signal correction method characterized by the above.

Images