JP5526524B2 - Noise suppression device and noise suppression method - Google Patents

Description

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

Conventionally, a voice reproduction device that drives a load such as a speaker according to an input signal, a voice communication device that transmits voice between remote locations, and a voice recognition device that understands the meaning and the like by distinguishing and recognizing the type of voice , Etc. have been proposed. In each of these apparatuses, since the sound is accurately reproduced, transmitted, or recognized, it is preferable that the influence of noise included therein is removed.
As such noise suppression techniques, for example, those disclosed in the following Patent Documents 1 and 2 and Non-Patent Documents 1 to 5 are known.
JP 2007-226264 A JP 2005-257748 A Boll, S., “Suppression of acoustic noise in speech using spectral subtraction”, IEEE Trans.Vol.ASSP-27, No2, pp.113-120, 1979. M. Berouti, el al, “Enhancement of Speech Corrupted by Acoustic Noise”, Proceedings of ICASSP, pp. 201-211, 1979. Lim & Oppenheim, “Enhancement and Bandwidth Compression of Noisy Speech”, Proc.IEEE, Vol67, No12, pp.1586-1604, 1979 Y. Ephraim and D. Malah, “Speech Enhancement Using a Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator”, IEEE Trans.Vol.ASSP-32, No.6, pp1109-1121, 1984. Junko Sasaki, Masafumi Tanaka, “Study of low distortion noise reduction method using masking effect”, IEICE Technical Report, EA98-106, pp.37-42, 1998

The techniques disclosed in these documents are basically related to a technique for suppressing noise by subtracting the level of the amplitude spectrum in the frequency domain by an appropriate technique, that is, a so-called spectral subtraction method. even by the one of the technologies, certain noise suppression effect can be enjoyed.

However, these documents still have undisclosed and unsolved problems.
For example, the spectral subtraction method is based on a method of estimating a noise spectrum included in an input signal and subtracting the noise spectrum estimation value from an amplitude spectrum. In this case, there is a high possibility that so-called musical noise is generated. There is a problem. This is because the estimated noise spectrum here does not necessarily reflect the actual noise spectrum. That is, in some frequency bands, noise may still remain after the noise spectrum estimation value is subtracted, and in other frequency bands, excessive noise may occur. For this reason, if the amplitude spectrum after subtracting the noise spectrum is reconverted to the time domain, a composite of sine waves with a plurality of random frequencies may appear, and this is reproduced, which is very annoying. Noise (that is, musical noise) may be generated.

In order to suppress this musical noise, it has been proposed to perform processing for adding the original sound to the amplitude spectrum after subtracting the estimated noise spectrum value (see Non-Patent Document 5).
However, in this technique, since the original sound addition ratio is determined based on the estimation of the SNR of the input signal, the noise suppression processing is somewhat lacking in flexibility. For example, in Non-Patent Document 5, on the basis that the SNR is 10 to 15 [dB], the original sound addition ratio is 0.05 when the SNR is higher than that, and 0.5 when the SNR is lower than that. Thus, since the original sound addition ratio is set regardless of the amount of noise that is originally desired to be suppressed or to be suppressed, there are some problems in the noise suppression effect that is actually enjoyed. Such a situation also applies to the above-mentioned Patent Document 2.

Furthermore, in the input signal, there are mainly a portion occupied by speech (speech portion) and a portion almost free of it (noise portion).
Under such circumstances, for example, as in Non-Patent Document 1 described above, the spectral subtraction method is applied to the speech portion, but the noise is suppressed by applying a fixed gain to the noise portion. When the fixed gain value is too small, the background noise amount increases when the noise part switches to the voice part. When the fixed gain value is too large, the background noise amount decreases. The phenomenon of becoming can occur. If this is reproduced, for example, the listener is likely to feel unnaturalness in the sense of hearing.

  It is an object of the present invention to provide a noise suppression device and a noise suppression method that can solve at least a part of the above-described problems.

In order to solve the above-described problem, the noise suppression device according to the present invention estimates a noise spectrum included in an input signal for each of K frequency bands (where K is a natural number of 2 or more) based on the input signal. Noise spectrum estimating means, first gain calculating means for calculating a noise suppression gain for each of the K frequency bands based on an estimation result by the noise spectrum estimating means, and applying the noise suppression gain to the input signal The input signal is added at a first ratio determined based on the difference between the noise suppression gain and the target noise suppression gain indicating the target level of noise suppression to the first noise-suppressed signal obtained as a result of And original sound adding means.

According to the present invention, since the original sound adding means adds the input signal to the first noise-suppressed signal, there is a case where the amplitude spectrum is excessively drawn by the noise spectrum estimation value as described above, for example. Even if it occurs, the portion is made up of the original sound, so that the occurrence of the musical noise is suppressed extremely effectively.
In addition, in the present invention, the “first ratio” having the significance as the original sound addition ratio is determined based on the difference between the noise suppression gain and the target noise suppression gain, so that it is actually enjoyed. The noise suppression effect becomes extremely effective. This is because, in the case of the present invention, the “first ratio” is determined in relation to the target noise suppression gain, not by a simple comparison between the SNR estimated value and the reference value as in the above example. This is because the target noise suppression gain can be set based on the amount of noise that is originally desired to be suppressed or to be suppressed.
The “target noise suppression gain” referred to in the present invention can be given artificially through an operation unit or the like installed outside the noise suppression apparatus according to the present invention, or can be automatically generated by any appropriate technique. May be calculated automatically.

The noise suppression apparatus according to the present invention further comprises second gain calculation means for calculating an average value gain for the K frequency bands for the noise suppression gain, wherein the original sound addition means is configured to add the average signal to the input signal. The input signal is added to the second noise-suppressed signal obtained as a result of applying the value gain at a second ratio determined based on the difference between the average gain and the target noise suppression gain. May be.
According to this aspect, the input signal is added to the second noise-suppressed signal at the “second ratio”.
Here, in this aspect, one of the features is that “average value gain” is used. Assuming that the noise suppression gain is G (1), G (2),..., G (K), for example, Gave as the average gain is Cave = (G (1) + G (2) +... + G (K)) / K or the like. Further, if the amplitude spectrum obtained by converting the input signal to the frequency domain is Y (1), Y (2),..., Y (K), the output signal in the frequency domain is Gave. Y (1), Gave · Y (2),..., Gave · Y (K), etc. (this is the “second noise obtained as a result of applying the average gain” in this aspect. This is a specific example of “post-suppression signal”.)
Therefore, first, the existence of such an average value gain does not cause, for example, an excessive case of the amplitude spectrum due to the estimated noise spectrum as described above, so that the generation of musical noise is extremely effective. Repressed.
In addition, in this aspect, since the original sound is added at the second ratio determined based on the difference between the average value gain and the target noise suppression gain, the musical noise can be reduced. The suppression effect can be enjoyed more effectively.

It should be noted that “add the input signal to the second noise-suppressed signal at a second ratio” in the present embodiment and the above-mentioned “add the first noise-suppressed signal to the first noise suppression signal at the first ratio. The relationship of “adding” is as follows.
First, the “noise suppression gain” is involved in both the generation of the “signal after the first noise suppression” and the standing of the “first ratio”, while the former is the target of the original sound addition processing based on the latter, while the “second noise suppression” The conceptual distinction that “average gain” is involved in both the generation of “rear signal” and the standing of “second ratio”, and the former is the target of the original sound addition processing based on the latter, is confirmed.
Under this assumption, first, as long as the “average value gain” in the present embodiment is calculated, the “original sound adding means” according to the present invention performs the original sound addition by the “second ratio” based on the “average value gain”. In the present invention, a relationship in which the original sound is not added based on the “first ratio” can be established. In other words, in this case, this mode is in a priority position with respect to the mode including the “first ratio” described above.
Alternatively, an aspect in which the original sound addition by the “first ratio” is performed in a part of the input signal and the original sound addition by the “second ratio” is performed in the other part can be realized in the present invention (in this case) The former part is related to the “first noise-suppressed signal”, and the latter part is related to the “second noise-suppressed signal”.)
Further, depending on the case, not only the original sound addition by the “second ratio” is performed on the “second noise-suppressed signal” but also the original sound addition by the “second ratio” is performed on the “first noise-suppressed signal”. In particular, the present invention falls within the scope thereof (this aspect will be referred to later as one of various aspects of the present invention). This means that the “second ratio” determined based on the “average value gain” is applied not to the “average value gain” but also to the “first noise-suppressed signal” established by the involvement of the “noise suppression gain”. Implies that it may be (and vice versa in some cases).

Further, in the noise suppression device according to the present invention, the original sound adding means calculates a smoothing ratio obtained by smoothing the first or second ratio on the time axis, and outputs the smoothed ratio to the first or second noise-suppressed signal. The input signal may be added at the smoothing ratio.
According to this aspect, a smoothing ratio obtained by smoothing the first or second ratio on the time axis is calculated. Therefore, after the “first ratio” or “second ratio” is established based on the aforementioned “difference”, the smoothing ratio is still subjected to the smoothing process. Note that “smoothing on the time axis” means that the calculated smoothing ratio is OGsmt-T (1), OGsmt-T (2),... OGsmt-T (r), in time series. ... (R is an appropriate integer), for example, using an appropriate smoothing coefficient δ, OGsmt−T (r) = δ · OGsmt−T (r−1) + (1−δ) · It is calculated as og or the like (however, og is “first ratio” or “second ratio”).
According to this, an abrupt change with the passage of time of the first or second ratio (more precisely, the smoothed first or second ratio, that is, the “smoothing ratio” referred to in this aspect). Therefore, the continuity and consistency of the noise suppression process is maintained.
In the present invention, as will be described later, it is preferable that processing for each frame divided over time is performed. In that case, the “time axis” in this aspect is more specifically, It can be assumed as an axis that is considered when the frames are arranged one by one. A more specific example of this point will be described with respect to an embodiment to be described later, particularly with respect to equation (7).
Further, in relation to this aspect, “first ratio” or “second ratio” referred to in this aspect is “original sound addition ratio og” and “smoothing ratio” is “original sound” in the embodiment described later. The addition ratio OG _t ”will be specifically described.

Further, in the aspect of the present invention including the concept such as the “average gain”, the input signal is detected over time by detecting the presence or absence of the sound included in the input signal. Voice detection means for classifying a frame and a noise frame not including the voice is further provided, and the second noise-suppressed signal is obtained by applying the average gain to a portion corresponding to the noise frame in the input signal. You may comprise so that a result may be obtained.
According to this aspect, the above average gain is applied to the noise frame, more preferably only to the noise frame, to generate the second noise-suppressed signal (and the second noise-suppressed signal is , Receiving the original sound addition process by the second ratio). In the noise frame, since the musical noise is relatively easy to generate, the present embodiment applies the average gain as if aiming at it, so that the optimum mode for obtaining the musical noise suppression effect is achieved. It can be said that.
In this aspect, the term “included” or “not included” in the voice should not be understood in an absolute sense. For example, two modes of a frame that is filled with “all speech” and a frame that “no speech exists” are assumed conceptually, but “voice frame” and “noise frame” refer to both extremes. ”Is not limited to this case, and this mode is limited to the case where only the latter is“ noise frame ”and all other cases are“ voice frames ”. It's not done. That is, this aspect does not require that “speech” should not be included in the noise frame even when it is recognized as a “noise frame”. The distinction of “noise frame” may be made on the basis of a suitable intermediate point in the above two cases.
In the above meaning, the term “included” or “not included” in this aspect, or the distinction between “voice frame” and “noise frame” in this aspect is relative. it can.

In this aspect, the first noise-suppressed signal may be configured to be obtained as a result of applying the noise suppression gain to a portion corresponding to the voice frame in the input signal.
According to this aspect, the noise suppression gain described above is applied to the speech frame to generate a first noise-suppressed signal (and the first noise-suppressed signal is subjected to original sound addition processing by the first ratio). Receive). This aspect can coexist with the aspect just described, but in this case, preferably, only the normal noise suppression gain without using the averaging process is used in the voice frame, and the average in the noise frame is used. That is, only the average value gain that has undergone the conversion processing is used. In view of the fact that the presence of noise is not so noticeable in speech frames and the opposite is the case in noise frames, the processing content according to this aspect enjoys a very rational, efficient, and effective noise suppression effect. to enable.

In this aspect, the original sound adding means may be configured to add the input signal to the first noise-suppressed signal at the second ratio instead of the first ratio.
According to this aspect, not only the original sound addition by the “second ratio” is performed on the “second noise-suppressed signal” but also the original sound addition by the “second ratio” is performed on the “first noise-suppressed signal”. become. That is, in this case, not the “average value gain” but also the “first noise-suppressed signal” established by the involvement of the “noise suppression gain”, the original sound by the “second ratio” determined based on the “average value gain” An addition process is performed.
According to this, since what is used as the original sound addition ratio is unified into the “second ratio”, processing efficiency and the like are realized.

Further, in an aspect in which a “noise suppression gain” is applied to a “voice frame” to obtain a “first noise-suppressed signal”, the original sound adding means is configured to input the input to the first noise-suppressed signal related to the voice frame. When trying to add a signal, the second ratio already calculated for the noise frame nearest to the voice frame is assumed to be the first ratio in the voice frame, and the signal after the first noise suppression is You may comprise so that an input signal may be added.
According to this aspect, in the original sound addition process in the voice frame, the original sound addition ratio (that is, the “second ratio”) used in the noise frame processed most recently is changed to the original sound addition ratio (that is, “first”). Is used as a percentage. That is, in this case, although the actual value is the value of the “second ratio” of the most recent noise frame, it is imitated as if it is the value of the “first ratio” in the voice frame. That is.
For this reason, in this aspect, the noise suppression processing executed in the most recent noise frame is succeeded to the noise suppression processing in the subsequent speech frame, and switching from the noise frame to the speech frame is performed. In this case, the consistency of the noise suppression process is maintained. As a result, it is possible to prevent the occurrence of a phenomenon in which the noise amount changes abruptly in the switching scene.
In this aspect, there is an aspect in which the original sound is added to the signal after the first noise suppression at the first ratio, and the original sound is added to the signal after the second noise suppression at the second ratio, and the former is included in the input signal. It is assumed that the latter is related to the speech frame and the latter is related to the speech frame.

Alternatively, in an aspect in which a “noise suppression gain” is applied to a “voice frame” to obtain a “first noise-suppressed signal”, the original sound adding unit is configured to input the input to the second noise-suppressed signal related to the noise frame. When trying to add signals, after calculating the second ratio in the noise frame as a temporary second ratio, the temporary ratio is calculated using the first or second ratio in the frame immediately before the noise frame. A smoothing ratio obtained by smoothing the second ratio on the time axis is calculated, and the input signal is added to the second noise-suppressed signal, assuming that the smoothing ratio is the second ratio, and the audio frame When the input signal is to be added to the first noise-suppressed signal according to the above, the first or second ratio in the frame immediately before the voice frame is set to the first or second ratio in the voice frame. As being 1 ratio, adds the input signal to the first noise suppression after signal may be configured to.
According to this aspect, the first or second ratio for each of the voice frame and the noise frame is preferably established. Since the smoothing ratio is calculated for the noise frame, the consistency and continuity of the noise suppression process is maintained, and for the voice frame, the noise spectrum in the “most recent noise frame” as described above is maintained. (I.e., according to this aspect, if the second rate for a certain noise frame has already been set, the value of the second rate is set to the new first value as long as the audio frame continues thereafter). Will continue to be maintained as a percentage value.) A more specific example of this point will be described with respect to an embodiment to be described later, particularly with respect to equation (7).
In this aspect, when simply referred to as “frame”, it may be “voice frame” or “noise frame”.
Further, in this aspect, there is an aspect in which the original sound is added to the signal after the first noise suppression at the first ratio and the original sound is added to the signal after the second noise suppression at the second ratio, and the former is included in the input signal. It is assumed that the latter is related to the speech frame and the latter is related to the speech frame.
Further, in relation to this aspect, “first ratio”, “second ratio”, “smoothing ratio” and the like referred to in this aspect are “original sound addition rate og” or “original sound addition” in the embodiment described later. In this case, it should be noted that there is no one-to-one relationship between one word in the former and one word in the latter.

In the noise suppression device according to the present invention, the first or second ratio may be obtained by the following equation (A).
og = max (0, tg−G) (A)
Here, og is the first or second ratio to be obtained, tg is the target noise suppression gain, G is the noise suppression gain or the average gain, and max (a, b) is a or b Means a function that returns the larger value.
According to this aspect, the first or second ratio is suitably calculated. A more specific example of this point will be described with respect to an embodiment to be described later, particularly with respect to formula (5).

On the other hand, in order to solve the above problem, the noise suppression method according to the present invention calculates the noise spectrum included in the input signal for each of K frequency bands (where K is a natural number of 2 or more) based on the input signal. A noise spectrum estimating step for estimating; a first gain calculating step for calculating a noise suppression gain for each of the K frequency bands based on an estimation result in the noise spectrum estimating step; and the noise suppression gain for the input signal. In the first noise-suppressed signal obtained as a result of application, the input signal is input at a first ratio determined based on the difference between the noise suppression gain and the target noise suppression gain indicating the target level of noise suppression. An original sound adding step of adding.

  According to the present invention, it is obvious that the above-described operational effects that are not essentially different from the operational effects described with respect to the noise suppression device according to the present invention are achieved.

The noise suppression method according to the present invention further includes a second gain calculation step of calculating an average value gain for the K frequency bands for the noise suppression gain, and in the original sound adding step, the average signal is added to the input signal. The input signal is added to the second noise-suppressed signal obtained as a result of applying the value gain at a second ratio determined based on the difference between the average gain and the target noise suppression gain. It may be configured.
According to this aspect, of the various aspects of the noise suppression device according to the present invention described above, this is essentially different from the operational effects described regarding the aspect in which the input signal is added to the second noise-suppressed signal at the second ratio. It is clear that there is an effect that does not become.
It should be noted that “add the input signal to the second noise-suppressed signal at a second ratio” in the present embodiment and the above-mentioned “add the first noise-suppressed signal to the first noise suppression signal at the first ratio. The relationship of “adding” is the same as described above.

Further, in the noise suppression method according to the present invention, by detecting the presence or absence of speech included in the input signal, the input signal is analyzed over time with respect to speech frames including the speech and noise not including the speech. And further comprising a voice detection step of segmenting into frames, wherein the second noise-suppressed signal is obtained as a result of applying the average gain to a portion corresponding to the noise frame of the input signal. Also good.
According to this aspect, among the various aspects of the noise suppression device according to the present invention described above, operational effects that are not essentially different from the operational effects described regarding the aspect in which the average value gain is applied to the noise frame are exhibited. Is obvious.
Note that the meaning of the term “included” or “not included” in this embodiment is the same as described above.

  In addition to the above, a more specific aspect of the present invention or an effect achieved by the present invention will be clarified in the description of the embodiment starting immediately after.

<First Embodiment>
In the following, a first embodiment according to the present invention will be described with reference to FIG. In addition to FIG. 1 mentioned here, in each drawing referred to below (for example, including the graph of FIG. 6 and the like), the ratio of dimensions of each part is appropriately different from the actual one. There is.

  As shown in FIG. 1, the noise suppression apparatus 1 includes a time / frequency conversion unit 10, a noise spectrum estimation unit 20, a noise suppression gain calculation unit 30, a noise period / noise suppression gain calculation unit 40, an original sound addition rate calculation unit 50, An original sound addition gain calculation unit 60, a frequency / time conversion unit 70, and a sound detection unit 80 are included.

The time / frequency conversion unit 10 performs Fourier transform on the time domain input signal to convert it to a frequency domain signal. The Fourier transform is preferably performed through dividing the input signal into a predetermined number of frames over time and applying an appropriate window function to the frames.
The signal in the frequency domain is divided into an amplitude spectrum and a phase spectrum, and the phase spectrum is sent as it is to a frequency / time conversion unit 70 described later. On the other hand, an amplitude spectrum is sent to each part after the noise spectrum estimation part 20 mentioned later, and receives various processes mentioned later.

The time domain input signal is also supplied to the voice detector 80. The voice detection unit 80 detects the presence or absence of a voice signal in the input signal. As described above, when the input signal is divided into frames, voice detection is performed for each frame (in the first embodiment, such processing is assumed). Here, “speech” means sound that is meaningful to a person, such as conversation, spoken language, music, various signals, and the like. That is, when the input signal is reproduced by an appropriate reproducing means, the relationship that the sound is obtained by reproducing the “voice signal” in the input signal is established.
This audio signal is detected based on, for example, whether or not the level of the input signal exceeds a predetermined threshold. However, the present invention can employ various methods other than this. For example, a method of estimating the probability of occurrence of an audio signal using a probability / statistical method may be employed, or the detection target may not be based on the input signal but after the Fourier transform. A method using a signal (that is, a signal in the above-described frequency domain) may be employed.
In the following description, a frame that is determined by the sound detection unit 80 to determine that an audio signal is present may be referred to as a “voice frame”, and a frame that is determined to be absent may be referred to as a “noise frame”. . The existence / non-existence here has no absolute meaning. As described above, since the presence / absence of the audio signal may be determined based on the predetermined threshold, the possibility that the “noise frame” includes what can be called an audio signal strictly speaking is not excluded.

ここで、Ｎ_ｔ（ｎ）は、現に処理中であるフレームにおける雑音スペクトル推定値、Ｎ_ｔ−１（ｎ）は、その直前のフレームにおける雑音スペクトル推定値（したがって、“ｔ”は、現に処理中であるフレームそれ自体を表現する添え字である。）、Ｙ（ｎ）は入力された振幅スペクトル、ｎは周波数帯域（に付けられた番号。なお、周波数帯域はＮ個に分割される。なお、このＮは、本発明にいう「Ｋ個の周波数帯域」のＫ以下（＝Ｎ≦Ｋ）である。）、βは平滑化係数である。また、式（１）中、ｃａｓｅ・Ａとあるのは、雑音スペクトル推定部２０が雑音フレームを処理する場合を表現し、ｃａｓｅ・Ｂとあるのは、音声フレームを処理する場合を表現している。
このように、雑音スペクトル推定部２０は、現に処理しているフレームが、雑音フレームであるか音声フレームであるかに応じて、雑音スペクトル推定値Ｎ_ｔ（ｎ）を求めるために利用する式を変更する。すなわち、音声フレーム処理時（ｃａｓｅ・Ｂ）には、その直前の雑音スペクトル推定値をそのまま用いて雑音スペクトル推定値Ｎ_ｔ（ｎ）を求め、雑音フレーム処理時（ｃａｓｅ・Ａ）には、入力した振幅スペクトルを時間軸上で平滑化することで、雑音スペクトル推定値Ｎ_ｔ（ｎ）を求める。 The noise spectrum estimation unit 20 calculates an estimated value of the noise spectrum based on the amplitude spectrum. Particularly in the first embodiment, a noise spectrum for each frequency band divided into a predetermined number is estimated based on the following equation (1).

Here, N _t (n) is a noise spectrum estimation value in the frame currently being processed, N _t−1 (n) is a noise spectrum estimation value in the immediately preceding frame (thus, “t” is actually processed). And Y (n) is an input amplitude spectrum, n is a frequency band (number assigned to it), and the frequency band is divided into N parts. Note that N is equal to or less than K (= N ≦ K) of “K frequency bands” in the present invention, and β is a smoothing coefficient. In Equation (1), “case · A” represents a case where the noise spectrum estimation unit 20 processes a noise frame, and “case · B” represents a case where an audio frame is processed. Yes.
As described above, the noise spectrum estimation unit 20 uses an expression used to obtain the noise spectrum estimation value N _t (n) depending on whether the frame currently processed is a noise frame or a speech frame. change. That is, at the time of voice frame processing (case · B), the noise spectrum estimation value N _t ( n ) is obtained using the noise spectrum estimation value immediately before that as it is, and at the time of noise frame processing (case · A), the input The noise spectrum estimation value N _t (n) is obtained by smoothing the measured amplitude spectrum on the time axis.

ここで、ｍａｘ（ａ，ｂ）は、ａ及びｂのうちいずれか大きい値を返す関数を意味する（以下、同じ。）。
この式（２）により、入力された振幅スペクトルＹ（ｎ）に対する雑音スペクトル推定値Ｎ_ｔ（ｎ）との間において、Ｙ（ｎ）＜Ｎ_ｔ（ｎ）が成立する場合は、Ｇ（ｎ）＝０となり、Ｙ（ｎ）＞Ｎ_ｔ（ｎ）が成立する場合は、Ｇ（ｎ）＝（Ｙ（ｎ）−Ｎ_ｔ（ｎ））／Ｙ（ｎ）となる。
この雑音抑圧ゲイン演算部３０で算出された雑音抑圧ゲインは、前記音声検出部８０によって区分された音声フレーム及び雑音フレームの別に応じて、雑音期間・雑音抑圧ゲイン演算部４０を介して又は直接に、原音加算ゲイン演算部６０に供給される。図１に示す雑音抑圧装置１は、このような処理を実現するためのスイッチを備える（図中弧線矢印参照）。 The noise suppression gain calculation unit 30 calculates a noise suppression gain based on the amplitude spectrum and the noise spectrum estimation value N _t (n) obtained by Expression (1). Particularly in the first embodiment, the noise suppression gain is calculated by the following equation (2).

Here, max (a, b) means a function that returns a larger value of a and b (hereinafter the same).
If Y (n) <N _t (n) holds between the noise spectrum estimation value N _t (n) for the input amplitude spectrum Y (n) according to this equation (2), G (n ) = 0, and when Y (n)> N _t (n) holds, G (n) = (Y (n) −N _t (n)) / Y (n).
The noise suppression gain calculated by the noise suppression gain calculation unit 30 is determined via the noise period / noise suppression gain calculation unit 40 or directly according to the voice frame and noise frame classified by the voice detection unit 80. The original sound addition gain calculation unit 60 is supplied. The noise suppression device 1 shown in FIG. 1 includes a switch for realizing such a process (see the arrow in the figure).

このｇは、式（３）の右辺から明らかなように、式（２）の雑音抑圧ゲインについての、周波数帯域ｎに関する平均値を意味する。
次いで、この式（３）の雑音抑圧ゲイン平均値ｇが、以下の式（４）によって平滑化される。

ここで、μは平滑化係数、Ｇ_ｔは、現に処理中である雑音フレームについての雑音抑圧ゲイン、Ｇ_ｔ−１は、その直前に処理した雑音フレームについての雑音抑圧ゲインである。
前述の式（１）のｃａｓｅ・Ａとして示される式もそうであるが、この式（４）では、現に処理中のフレームにおける雑音抑圧ゲインを求めるにあたって、その直前に処理されたフレームにおけるそれが参照されていることから、時間軸上で平滑化が行われているということがいえる（後述する式（７）についても同様である。）。
この式（４）中のＧ_ｔが、本雑音期間用ゲイン演算部４０において求められるべき、雑音期間に適用するための雑音抑圧ゲイン（以下、簡単のため、「雑音期間用ゲイン」ということがある。）である。
雑音期間用ゲイン演算部４０は、このようにして求められた雑音期間用ゲインＧ_ｔを、すべての周波数帯域に対して一律に適用する。以下では、この事情を表現するため、この一律に適用されるＧ_ｔを、Ｇ１（ｎ）と表現する。この場合、Ｇ１（０），Ｇ１（１），…，Ｇ１（Ｎ−１）のすべてが、Ｇ_ｔに等しい。 The noise period / noise suppression gain calculation unit 40 (hereinafter sometimes referred to as “noise period gain calculation unit 40” for simplicity) calculates a noise suppression gain to be applied to the noise frame. In the first embodiment, the following method is used to calculate the noise suppression gain.
First, g expressed by the following equation (3) is calculated based on the noise suppression gain G (n) obtained by equation (2).

As is apparent from the right side of Equation (3), g means an average value related to the frequency band n for the noise suppression gain of Equation (2).
Next, the noise suppression gain average value g of the equation (3) is smoothed by the following equation (4).

Here, μ is a smoothing coefficient, G _t is a noise suppression gain for a noise frame currently being processed, and G _t−1 is a noise suppression gain for a noise frame processed immediately before.
The same applies to the expression shown as case · A in the above-described expression (1). In this expression (4), when the noise suppression gain in the frame currently being processed is obtained, that in the frame processed immediately before is calculated. Since it is referred, it can be said that smoothing is performed on the time axis (the same applies to equation (7) described later).
G _t of the equation (4) is to be determined in the noise period for the gain calculation unit 40, the noise suppression gain to be applied to the noise period (hereinafter, for simplicity, be referred to as "noise period gain" Yes.)
Noise period gain calculation unit 40, thus the noise period for the gain G _t obtained by, applied uniformly to all frequency bands. In the following, in order to express this situation, G _t that is applied uniformly is expressed as G1 (n). In this case, G1 (0), G1 ( 1), ..., all G1 (N-1) is equal to _{G t.}

ここで、ｔｇは、目標雑音抑圧ゲインであり、以下の式（６）に基づいている。

この式（６）中のＴＧは、目標雑音抑圧量であり、ｄＢ単位で与えられる。このＴＧ（あるいは、ｔｇ）は、装置外部から図示しない操作部等を介することによって人為的に与えられたり、あるいは、何らかの適当な手法により自動的に演算されてよい。
以上の式（５）によれば、目標雑音抑圧ゲインｔｇと雑音期間用ゲインＧ_ｔとの間において、ｔｇ＜Ｇ_ｔが成立する場合は、ｏｇ＝０となり、ｔｇ≧Ｇ_ｔが成立する場合は、ｏｇ＝ｔｇ−Ｇ_ｔとなる。 The original sound addition rate calculation unit 50 calculates the original sound addition rate of the original sound signal with respect to the noise-suppressed signal. Particularly in the first embodiment, the original sound addition rate og is obtained based on the following equation (5).

Here, tg is a target noise suppression gain and is based on the following equation (6).

TG in the equation (6) is a target noise suppression amount, and is given in dB units. This TG (or tg) may be given artificially from the outside of the apparatus through an operation unit (not shown) or may be automatically calculated by some appropriate method.
According to equation (5) above, between the target noise suppression gain tg and noise period for the gain _{G t,} if tg _{<G t} is satisfied, if og = 0 becomes, tg ≧ _{G t} is satisfied is a og = _{tg-G t.}

ここでＯＧ_ｔは、現に処理中であるフレームにおける原音加算割合、ＯＧ_ｔ−１は、その直前のフレームにおける原音加算割合、λは平滑化係数である。なお、式（７）中のｃａｓｅ・Ａ及びｃａｓｅ・Ｂの意義は、上述の式（１）の場合と同様である（以下の式（８）においても同じである。）。
このように、原音加算ゲイン演算部６０は、現に処理しているフレームが、雑音フレームであるか音声フレームであるかに応じて、原音加算割合ＯＧ_ｔを求めるために利用する式を変更する。すなわち、音声フレーム処理時（ｃａｓｅ・Ｂ）には、その直前の原音加算割合をそのまま用いて、原音加算割合ＯＧ_ｔを求め、雑音フレーム処理時（ｃａｓｅ・Ａ）には、前記の原音加算率ｏｇを時間軸上で平滑化することで、原音加算割合ＯＧ_ｔを求める。 The original sound addition gain calculator 60 calculates the noise suppression gain after the original sound addition based on the original sound addition rate og. In the first embodiment, the following method is used to calculate the noise suppression gain.
First, OG _t expressed by the following equation (7) is calculated based on the original sound addition rate og obtained by equation (5).

Here, OG _t is the original sound addition ratio in the frame currently being processed, OG _t-1 is the original sound addition ratio in the immediately preceding frame, and λ is the smoothing coefficient. In addition, the meanings of case · A and case · B in the formula (7) are the same as in the case of the above-described formula (1) (the same applies to the following formula (8)).
In this way, the original sound addition gain calculation unit 60 changes the expression used to obtain the original sound addition ratio OG _t depending on whether the frame currently being processed is a noise frame or an audio frame. That is, at the time of audio frame processing (case · B), the original sound addition ratio OG _t is obtained using the original sound addition ratio immediately before it, and at the time of noise frame processing (case · A), the above-mentioned original sound addition ratio The original sound addition ratio OG _t is obtained by smoothing og on the time axis.

ここで、Ｇ１（ｎ）は、上で説明したように、雑音フレームにおいて、すべての周波数帯域に対して一律に適用される雑音期間用ゲインを表している。
この式（８）によれば、前述の式（７）における場合分けに応じて、原音加算後の雑音抑圧ゲインＧ２（ｎ）（以下、簡単のため、「修正後ゲインＧ２（ｎ）」ということがある。）が求められる。 Next, the original sound addition gain calculation unit 60 obtains the noise suppression gain after the original sound addition based on the following equation (8).

Here, G1 (n) represents the noise period gain that is uniformly applied to all frequency bands in the noise frame, as described above.
According to this equation (8), the noise suppression gain G2 (n) after adding the original sound (hereinafter referred to as “corrected gain G2 (n)” for simplicity) according to the case classification in the aforementioned equation (7). May be required).

  The multiplier 11 shown in FIG. 1 applies the corrected gain G2 (n) obtained as described above to the amplitude spectrum Y (n). That is, an operation of S (n) = G2 (n) · Y (n) is performed, and as a result, an amplitude spectrum S (n) after noise suppression to be finally obtained is obtained.

  Finally, the frequency / time conversion unit 70 is based on the amplitude spectrum S (n) after noise suppression obtained as described above and the phase spectrum directly supplied from the time / frequency conversion unit 10. Generate a time domain output signal. In the first embodiment, since the Fourier transform is applied in the time / frequency conversion unit 10, the frequency / time conversion unit 70 performs an inverse Fourier transform.

Next, the operation, operation, and effect of the noise suppression device 1 according to the first embodiment described above will be described with reference to FIGS. 2 to 4 in addition to FIG. 1 already referred to.
First, the time / frequency converter 10 performs a Fourier transform on the input signal, and further decomposes it into an amplitude spectrum Y (n) and a phase spectrum as shown in FIG. 1 (step S101 in FIG. 2). ). At this time, the time / frequency conversion unit 10 performs processing for each frame as described above.
In parallel with this, the voice detection unit 80 detects the presence or absence of a voice signal included in the input signal (step S102 in FIG. 2). This detection process enables a process of separating the input signal into a voice frame and a noise frame. The voice detection unit 80 also performs this process.

Next, the noise spectrum estimation unit 20 obtains a noise spectrum estimation value N _t (n) for each frequency band n having a predetermined width by using the amplitude spectrum Y (n) and the equation (1) described above. In this case, as described above, different processing is performed depending on whether the currently processed frame is a noise frame or a voice frame (see step S103 in FIG. 2). As shown in FIG. 2, the noise spectrum estimation value N _t (n) is calculated and thereafter the output signal generation process (step S104 in FIG. 2) by the multiplier 11 shown in FIG. Depending on the distinction between a frame and an audio frame, a process with substantially different contents is developed. Therefore, in the following, the processing for the noise frame will be described first, and the processing for the audio frame will be classified into [I] and [II].
In addition, such a classification process is based on switching of the switch according to the detection result of the audio | voice detection part 80, as shown in FIG.

[I] First, in the noise frame processing, the noise spectrum estimation value N _t (n) is obtained from the equation shown as case · A in the equation (1) (step S201 in FIG. 2). As described above, this is due to the smoothing process of the input amplitude spectrum Y (n).

Next, the noise suppression gain G (n) is calculated based on the noise spectrum estimation value N _t (n) and the equation (2) (step S202 in FIG. 2). This is due to the action of the noise suppression gain calculator 30 of FIG. As described above, when Y (n)> N _t (n) is satisfied, G (n) = (Y (n) −N _t (n)) / Y (n). , G (n) = 0. According to this, for example, a noise suppression gain as shown in FIG. 3C is obtained (in FIG. 3B, the above-described noise spectrum estimation value N _t (n), FIG. 3A). In each, the amplitude spectrum of the input signal is illustrated.)

Next, an average value g of the noise suppression gain G (n) with respect to the frequency band is obtained by Equation (3) and Equation (4), and a smoothing process is performed for the noise period. A gain _Gt is obtained (step S203 in FIG. 2). The averaging and smoothing the noise period for the gain G _t passed through becomes the common G1 (n) the entire frequency band. This is due to the action of the noise period gain calculation unit 40.
Thus, in the first embodiment, the noise suppression gain G (n) obtained by the equation (2) is not used as it is, but the frequency according to the equation (3) is used for the G (n). Using the noise period gain G _t after performing averaging on the band and smoothing on the time axis according to Equation (4) as the noise period gain G1 (n) for the entire frequency band, One of its major features.
Note that FIG. 3D illustrates an example when the averaging process is performed on the noise suppression gain G (n) (see also the broken line shown in FIG. 3C).

Next, the original sound addition rate og is obtained by the above-described noise period gain G _t and the above equation (5) (step S204 in FIG. 2). This is due to the action of the original sound addition rate calculation unit 50 in FIG. Here, the setting of the target noise suppression gain tg or the target noise suppression amount TG works as one dominant factor. That is, if the noise period gain G _t is larger than the target noise suppression gain tg, the original sound addition rate og is set to 0, and if not, the original sound addition rate og (in accordance with the noise period gain G _t ( That is, og = tg−G _t ) is set. The proper use of both has the significance of determining how to enjoy the sound quality improvement effect brought about by adding the original sounds in relation to the target noise suppression amount TG. That is, in the latter case, and within the framework defined by the target noise suppression quantity (i.e., the portion corresponding to the difference between tg and G _t) by adding the original sound that promote quality improvement and main purpose, in the former case, Since G _t > tg is satisfied and there is no room for improvement in sound quality, the original sound addition rate og is set to 0 (in this case, it is rather suppressed that the amount of noise increases). After all, the above formulas (5) and (6) are based on the observance of the target noise suppression amount, and the process of improving the sound quality within the frame when there is still room for adding the original sound. There is significance to realize.
Thus, in the first embodiment, that the original addition rate og is obtained by utilizing a noise period for the gain G _t, 1 Tsugaaru of its major characteristics.

Next, the original sound addition ratio OG _t is obtained from the above-described original sound addition rate og and the expression shown as case · A in the expression (7) (step S205 in FIG. 2). The original sound addition ratio OG _t is obtained by smoothing the original sound addition ratio og on the time axis as described above. Then, the noise suppression gain after the original sound addition, that is, the corrected gain G2 (n) is obtained from the original sound addition ratio OG _t thus obtained and the above equation (8). The above is due to the action of the original sound addition gain calculation unit 60.
In this case, the corrected gain G2 (n) is a gain determined in consideration of the noise period gain G1 (n) that has been subjected to the above-described averaging / smoothing and the degree of original sound addition. It has implications.

In consideration of the case immediately after the start-up of the apparatus, it is preferable that an initial value as a value corresponding to N _t-1 (n) in the formula (1) is appropriately determined (as such an initial value). N _t-1 (n) can be used in the calculation process of the noise spectrum estimation value N _t (n) in the speech frame processing described later. The same can be said for G _t-1 (n) in the equations (4) and (7).

[II] On the other hand, in the audio frame processing, basically, almost the same processing as the above-described noise frame processing is executed. That is, the noise spectrum estimation value N _t (n) and the noise suppression gain G (n) based thereon are obtained (see steps S301 and S202 in FIG. 2), and the corrected gain G2 (n) based on the original sound addition ratio OG _t (Steps S303 and S304 in FIG. 2) is the same as the noise frame processing.
However, this audio frame processing has the following differences or points of caution compared to noise frame processing.

(I) The noise spectrum estimation value N _t (n) is obtained not by the equation shown as case · A in the equation (1) but by the equation shown as case · B (step S301 in FIG. 2). Since this equation is N _t (n) = N _t−1 (n), it can be said that the voice frame processing is processing that maintains the current state. More specifically, if the previous audio frame is a noise frame, the noise spectrum estimation value N _t−1 (n) calculated in the noise frame is used as it is in the audio frame processing. On the other hand, if the previous audio frame is an audio frame and further is a noise frame, the noise spectrum estimate N _t− calculated in the noise frame is calculated. ₂ (n) is used as it is in the audio frame processing as it is.
In short, in a speech frame, the estimated noise spectrum value N _tp (n) (p is the number of frames up to the nearest noise frame counted from the immediately preceding frame of the speech frame. (Including both ends)) will be used.

(Ii) The same thing can be said in the calculation process of the original sound addition ratio OG _t performed using the equation (7). That is, since the expression shown as case · B in Expression (7) is OG _t = OG _t−1 , in this case as well, the current state is maintained in the audio frame processing (step of FIG. 2). (See S303).
If the expression is consistent with the above case, in the voice frame, the original sound addition ratio OG _t-p (n) (p is calculated from the immediately preceding frame of the voice frame, and is calculated from the latest noise frame. This means that the number of frames up to the noise frame (including both ends) is used.

(Iii) The calculation of the noise suppression gain G (n) itself is performed in the same manner using the above equation (2) regardless of whether it is a speech frame or a noise frame. In step S202 of FIG. 2, the boxes corresponding to [noise frame processing] and [voice frame processing] are connected and drawn symbolically (note that the expression ( The value of N _t (n) in 2) is naturally different for both frames depending on the case · A and case · B in equation (1).

(Iv) In the audio frame processing, the processing related to the equations (3) and (4), that is, the averaging / smoothing processing for the noise suppression gain G (n) is not performed (step S203 and FIG. 2). (See the right side of the figure). As a result, a state in which a valid noise period gain G _t does not exist is present, so that the processing related to the equation (5), that is, the processing for calculating the original sound addition rate og is also not performed. (See step S204 in FIG. 2 and the right side in the figure).

(V) The corrected gain G2 (n) that is finally calculated is determined not by the equation shown as case · A in the equation (8) but by the equation shown as case · B (see FIG. 2). Step S304). In this case, the noise period gain G1 (n) that has undergone averaging and smoothing is used during the noise frame processing, whereas the noise suppression gain G ( The difference is that n) is used as it is.

After the above processes [I] and [II], the corrected gain G2 (n) is obtained in any case. The original amplitude spectrum Y (n) is added to the corrected gain G2 (n). If so, the amplitude spectrum S (n) after noise suppression is calculated (step S104 in FIG. 2).
In FIG. 3E, for the sake of simplicity, the result of simply multiplying the amplitude spectrum Y (n) of FIG. 3A by the averaged noise suppression gain (ie, g) of FIG. 3C. It is shown. In the first embodiment, as described above, in addition to this, the gain is further adjusted in consideration of the degree of original sound addition (see the expression (8), particularly the role of OG _t (n), see ) And FIG. 3 (E) well represent the essence of processing assuming that such consideration for the original sound addition processing is omitted (in equation (8), OG _t (n) = 0). Then, the corrected gain G2 (n) is simply equal to G1 (n) or G (n).)

According to the noise suppression device 1 having the configuration and operation described above, the following effects can be obtained.
First, according to the noise suppression device 1 of the first embodiment, the noise included in the input signal is suppressed extremely suitably. Here, the term “preferably” includes the following points in particular in the first embodiment.

(1) First, according to the first embodiment, generation of so-called musical noise can be extremely effectively prevented. Here, the musical noise means noise generated after the noise spectrum estimation value is subtracted from the amplitude spectrum of the input signal.
For example, the noise suppression gain based on the noise spectrum estimation value can be easily obtained by using (Y (n) −N (n)) / Y (n) in the equation (2). Assuming a mode in which this is applied as it is by the multiplier 11 shown in FIG. 1, the amplitude spectrum S (n) after noise suppression is S (n) = {(Y (n) −N (n)) / Y. (N)} · Y (n) = Y (n) −N (n). That is, in this case, the noise spectrum-suppressed amplitude spectrum S (n) is obtained by simply subtracting the noise spectrum estimation value from the amplitude spectrum of the input signal.
However, since the estimated noise spectrum value in this case is merely an “estimated value”, it does not necessarily reflect the actual noise spectrum. Therefore, in some frequency bands, noise may still remain after subtracting the noise spectrum estimate, and in other frequency bands, too much may occur (in this case, the negative amplitude spectrum is As long as it cannot be considered, it is set to 0.) In FIG. 4, such a situation is conceptually expressed. For example, solid lines in FIG. 4C remain undrawn (see reference “KN”), and broken lines are drawn too much (see reference “HS”). 4 (A) and (B) are exactly the same as FIGS. 3 (A) and 3 (B), and the part designated by the symbol HSt in FIG. Occasionally, this is an example when Y (n) -N (n) = 0 holds true.)
When such an amplitude spectrum S (n) is subjected to inverse Fourier transform in the time domain, the signal looks like a combination of sine waves having a plurality of random frequencies, and if this is reproduced, it is very annoying. It will be heard as sound. This is musical noise.
As described above, the musical noise is mainly caused by the fact that the actual noise spectrum that is unknown to the strict sense does not match the estimated noise spectrum value.

In the first embodiment, the occurrence of such musical noise is extremely effectively suppressed. This is because at the time of noise frame processing, the averaged / smoothed noise period gain G _t is used to obtain a corrected gain G2 (n), which is applied to the amplitude spectrum Y (n). This is because (see FIG. 3E). As a result, noise suppression is performed while maintaining the frequency structure originally possessed by the amplitude spectrum, so that musical noise is extremely difficult to occur.

(1-i) It should be noted that each of the averaging performed when obtaining the noise period for the gain G _t (Formula (3)) and smoothed (the formula (4)) of, there is an inherent meaning. As is clear from FIG. 3, the purpose of the former is mainly to lead to the effect of suppressing the musical noise, and the purpose of the latter is mainly to say that the continuity of noise suppression processing as seen through time. Is to maintain. According to the latter, since the abrupt change with the passage of time of the noise period gain G _t (n) does not occur, for example, when the signal included in the noise frame is reproduced, the listener is In other words, the other smoothing processes performed in the first embodiment (that is, the case A of the expression (1) and the case A of the expression (7)) are not fundamental. In fact, it has a meaning that is not essentially different from this.)

(2) Secondly, the above (1) relates to the prevention of the occurrence of musical noise related to noise frame processing. In this regard, according to the first embodiment, the generation of musical noise related to voice frame processing is also prevented. Better realized. As described above, in the audio frame processing, the noise suppression gain G (n) (see Equation (2)) that does not undergo averaging / smoothing is used as it is, and the corrected gain G2 ( n) is obtained (case · B in formula (8) or [II] (v) described above).

(3) Thirdly, according to the first embodiment, the consistency of the noise suppression process is maintained in the scene of switching from the noise frame to the voice frame. This is because, as described above, at the time of the speech frame processing, as the noise spectrum estimation value N _{t (n),} calculated N _{t-p (n)} is adapted to be utilized in the most recent noise frames (Refer to the description of [II] (i) above).
In short, the above (2) and (3) are required. In the first embodiment, while effective noise suppression is performed on a speech frame, the noise suppression processing (particularly, the effect) at the time of noise frame processing is respected. Thus, a contrivance is made so that the flow between both frames becomes more natural. According to this, when the noise suppression apparatus 1 of the first embodiment is connected to some kind of sound reproduction means, the listener can change the volume feeling related to noise in the scene of switching from the noise frame to the sound frame, etc. It does not give a sense of incongruity in hearing.

In order to suppress the musical noise in the voice frame, S (n) = Y (n) −αN (n) is used instead of S (n) = Y (n) −N (n) described above, Although a method of increasing the value of α (> 0) is also conceivable, this suffers from a drawback that the possibility of severe deterioration of sound quality becomes extremely high. However, if α is made small, the suppression of musical noise becomes insufficient.
In addition, a constant value (noise floor) is added to a portion where the amplitude spectrum after noise suppression becomes 0 (that is, a portion indicated by symbols HS and HSt) indicated by a broken line in FIG. It is also conceivable to suppress musical noise by adopting a technique to do this. This is based on the idea of masking the remaining part KN (or making it inconspicuous) by putting on the part HS and HSt, so to speak, clogs (so this technique and the above-mentioned). When using together with the method using α, α may be set smaller, and in this case, the effect of preventing deterioration of sound quality is also obtained.
However, such addition of the noise floor means to increase the absolute amount of noise, so there is not only a problem from the viewpoint of achieving the original purpose of noise suppression. Depending on the setting of the amount of floor, there is a high possibility that the noise suppression effect is very insufficient.

  From this point of view, it is clear that the noise suppression device 1 of the first embodiment is extremely advantageous. That is, in the first embodiment, since the amount of subtraction is not mechanically increased as in the case of the use of α, there is almost no possibility that the sound quality will deteriorate, and the noise floor is simply added. Such a process is not performed, and the noise suppression effect once performed is not sacrificed. And as already mentioned, the musical noise is effectively suppressed despite this being the case.

(4) According to the noise suppression apparatus 1 of the first embodiment, as described with reference to the above-described formulas (5) to (7) or steps S205 and S303 in FIG. As a result, the noise suppression effect is more effectively achieved. According to the original sound addition process, it is possible to expect the same effect as the noise floor addition process described above, that is, the masking effect of the remaining portion KN in FIG. 4C. (However, the noise floor is “constant” to the last. This is a decisive difference from the case of using “original sound”).
In the above description, in order to more clearly grasp the effect exerted by the noise suppression device 1 of the first embodiment, in comparison with the method using the α or the method using the noise floor, Although there is a part which has been described, the present invention does not have the intention to positively exclude the technique of suppressing musical noise using these α or noise floor. In other words, these methods and the present invention and various aspects thereof can be used in combination, and according to such a combined form, the effects of the present invention and various aspects thereof can be further emphasized while enjoying the advantages of the method. It becomes possible.

In addition, the first embodiment is characterized not by simply adding the original sound but by the following points.
(4-i) First, the ratio of original sound addition (that is, OG _t ) is determined based on the original sound addition ratio og determined according to the magnitude of the noise period gain G _t and the target noise suppression gain tg. It is like that. Specifically, as already described, in the original sound addition process, the target degree of noise suppression (that is, tg) is one of the dominant factors, and the original sound addition rate og is determined in relation to this. since it way, a process based on the noise period for the gain G _t, between the original addition process, by selectively used with balanced is performed more effectively noise suppression effects and musical noise suppression effect, Furthermore, the sound quality improvement effect is enjoyed.

(4-ii) Also in such an original sound addition process, OG _t-p calculated in the latest noise frame is used as the original sound addition ratio OG _t during the audio frame process. (Refer to the description of [II] (ii) above). This is the same as the above-described idea that the previous noise spectrum estimation value N _t-1 (n) is used as it is as the noise spectrum estimation value N _t (n) in a certain voice frame. In other words, even in the original sound addition process, the consistency of the noise suppression process is maintained in the scene of switching between the noise frame and the voice frame.

Second Embodiment
In the following, a second embodiment according to the present invention will be described with reference to FIGS. Note that the second embodiment is different from the first embodiment in that there are differences related to the sound detection processing, and the other points are exactly the same as the first embodiment unless otherwise specified. is there. Therefore, in the following description, the difference will be mainly described, and description of other points will be simplified or omitted. Further, the reference numerals on the drawings are also used except for the differences.

  As shown in FIG. 5, the noise suppression device 1 ′ according to the second embodiment has a configuration in which a voice detection unit 801 is connected to a subsequent stage of the noise suppression gain calculation unit 30. That is, the voice detection unit 801 detects the presence / absence of a voice signal in the input signal by using the noise suppression gain G (n) calculated by the equation (2), or the voice frame and the noise frame. And make a distinction.

ここで、ｇは、上記第１実施形態において利用されていた式（３）によって表現されるｇであって、要するに、Ｇ（ｎ）についての周波数帯域ｎに関する平均値である（第２実施形態は、このｇの演算を、雑音期間用ゲイン演算部４０だけでなく、音声検出部８０１も行う。むろん、両者の一方で行った演算の結果を、両者間で共用してもよい。）。
この式（９）のＶａｒは、表式から明らかな通り、Ｇ（ｎ）の分散を表す。 In the second embodiment, the following method is used to detect the presence or absence of an audio signal.
First, based on the noise suppression gain G (n) obtained by Expression (2), Var expressed by Expression (9) below is calculated.

Here, g is g expressed by Expression (3) used in the first embodiment, and in short, is an average value regarding the frequency band n for G (n) (second embodiment). The calculation of g is performed not only by the noise period gain calculation unit 40 but also by the voice detection unit 801. Of course, the result of the calculation performed by one of the two may be shared by both.
Var in the formula (9) represents the dispersion of G (n) as is apparent from the table.

Next, it is determined whether or not this Var exceeds a predetermined value. The significance of this judgment is as follows.
In general, the noise suppression gain G (n) calculated by the equation (2) shows a very different aspect between the case where the audio signal is included and the case where the audio signal is not included. FIG. 6 and FIG. 7 show an example thereof. The former is a calculation example of the noise suppression gain G (n) when a speech signal is included, and the latter is the noise suppression gain G (n) when it is not included. It is a calculation example. As is clear from the comparison of these figures, if the variance of G (n) in each case is calculated, it can be easily estimated that a large gap occurs between the two. In other words, if the variance value of G (n) for a certain frame is large to a certain extent, it can be judged with a certain degree of certainty that it includes an audio signal, otherwise it does not include an audio signal. It is.
This is the significance of the above-described judgment on the magnitude of Var. In other words, if there is a certain predetermined value VB, if Var> VB, the frame has an audio signal. Therefore, it is distinguished as “audio frame”, and if Var ≦ VB, There is no audio signal, so it is distinguished as a “noise frame”.

ここで、ＰＡ_ｔ（ｎ）は、現に処理中であるフレームにおける入力信号中の振幅スペクトルであって平滑化されたもの、ＰＡ_ｔ−１（ｎ）は、その直前のフレームにおける当該振幅スペクトルであって平滑化されたもの、αは平滑化係数、γ・βは制御パラメータである。また、式（１１）中、ｃａｓｅ・Ｃとあるのは、ＰＡ_ｔ（ｎ）＞Ｎ_ｔ−１（ｎ）が成立する場合を表現し、ｃａｓｅ・Ｄとあるのは、それ以外の場合を表現している。 In the configuration of FIG. 5, unlike the configuration of FIG. 1, the noise spectrum estimation unit 20 cannot use the detection result of the audio signal. That is, the noise spectrum estimation unit 20 calculates the noise spectrum estimation value N _t (n) without assuming the distinction between the voice frame and the noise frame.
The noise spectrum estimation value N _t (n) in such a case may be obtained by, for example, the following expressions (10) and (11).

Here, PA _t (n) is an amplitude spectrum in the input signal in the frame currently being processed and smoothed, and PA _t−1 (n) is the amplitude spectrum in the immediately preceding frame. The smoothing coefficient, α is a smoothing coefficient, and γ · β is a control parameter. In the formula (11), case · C represents a case where PA _t (n)> N _t−1 (n) is satisfied, and case · D represents a case other than that. expressing.

In this case, the combination of the formula shown as case · D in the formula (11) and the formula (10) is substantially the same as the formula shown as case · A in the formula (1).
On the other hand, the formula shown as case · C in the formula (11) does not correspond to the formula (1). However, as described above, that is, when PA _t (n)> N _t−1 (n) holds, that is, the amplitude spectrum in the frame currently being processed becomes the noise spectrum in the immediately preceding frame. Since it is activated when the estimated value is exceeded, it is not impossible for this case · C to be considered as a suggestion that the frame currently being processed may be a speech frame (a lot of n (= (0, 1, 2, 3,...), If such a condition is satisfied, the possibility is further increased, but it is merely a “suggestion”).
It can be said that these formulas (10) and (11) have commonality with the above formula (1) as long as they have the above meanings.
In any case, the noise spectrum estimation value is still preferably calculated.

According to such 2nd Embodiment, the following effects are show | played.
First, it is obvious that this second embodiment also provides the operational effects that are not essentially different from the operational effects achieved by the first embodiment. That is, also in the second embodiment, the effects (1) to (4) described in relation to the first embodiment are substantially the same.

In addition, according to the second embodiment, as is apparent from the comparison between FIG. 1 and FIG. 5, effects such as improved processing efficiency and simplified circuit configuration can be enjoyed. This is performed by using the noise suppression gain G (n) for the voice detection in the second embodiment instead of performing the voice detection in the first embodiment independently. This is due to the dependency.
In the present invention, since the calculation of the noise suppression gain G (n) is a process that must be performed, the voice detection process is also performed using the calculation result, which improves the efficiency and rationalization of the process. There is no doubt about guiding. Moreover, the detection performance is considerably high (see FIG. 6 and FIG. 7 comparison).

While the embodiments according to the present invention have been described above, the noise suppression device according to the present invention is not limited to the above-described embodiments, and various modifications can be made.
(1) In the first and second embodiments, the noise period gain G _t is averaged on the frequency axis and smoothed on the time axis, but the present invention is not limited to such a form. As described above, since the main purpose of the averaging process and the smoothing process are different, the smoothing process may be omitted depending on circumstances. As shown in FIG. 3E, even if only the averaging process is performed, the effect of suppressing the musical noise can be enjoyed to a certain extent.

(2) In the first and second embodiments, the noise period gain G _t is obtained through the averaging process according to the equation (3) and the smoothing process according to the equation (4). However, the present invention is not concerned with the form of these formulas (3) and (4).
First, in the present invention, the noise suppression gain average value g is not limited to the form obtained by Expression (3).
That is, in Expression (3), g is calculated using all N frequency bands (N 0, 1, 2,..., N−1th frequency bands in total). May be calculated using, for example, only some of the frequency bands. In this case, it is conceivable to use a frequency band excluding both or one of the extremely low band (band close to the DC component) and the extremely high band (band close to the Nyquist frequency).
Further, in obtaining the noise suppression gain average value g, different weights may be applied to individual frequency bands. For example, a specific weighting factor is multiplied only for a specific frequency band, a weighting factor that is continuously or stepwise increased or decreased is multiplied for all frequency bands, and so on.
Next, in the present invention, the noise period gain G _t is not limited to the form obtained by the equation (4).
That is, in Expression (4), G _t is obtained by smoothing the noise suppression gain average value g on the time axis. This G _t is, for example, the average value of g of adjacent frames. May be calculated as

(3) In addition, in the first and second embodiments, the averaged / smoothed noise period gains _Gt to G1 (n) are applied to all frequency bands (Equation (8)). Case · A or see FIG. 3E), the present invention is not limited to such a form.
For example, the noise period gains _Gt to G1 (n) may be applied only to the frequency band excluding both or one of the above-described extremely low frequency range and extremely high frequency range. In this case, a gain that is a fixed value may be applied to the frequency band that is excluded from the application.

(4) In the first and second embodiments, the noise suppression gain G (n) is calculated by the equation (2), but the present invention is not limited to such a form. For example, besides this, the Wiener filter method, the MMSE (Minimum Mean-Square Error) method, etc. may be used (see the above-mentioned Non-Patent Documents 3 and 4). The SNR (voice (signal) / noise ratio) may be estimated, and the noise suppression gain G (n) may be obtained based on the SNR.

(5) In the second embodiment, in order to distinguish between a voice frame and a noise frame, the variance on the frequency axis for the noise suppression gain G (n) is taken according to the equation (9). The present invention is not limited to such a form.
For example, instead of the variance, the standard deviation may be used. Naturally, the variance on the time axis or the standard deviation may be used. In addition, based on the number of noise suppression gains G (n) for each frequency band that fall within the space defined by two predetermined reference values, a distinction is made between audio frames and noise frames. (For example, if the number is relatively large, it can be determined that the noise suppression gain G (n) is concentrated in a certain place, so that the degree of dispersion is small. The frame is identified as a noise frame, etc.). Furthermore, the various determination methods described above may be used in some cases. According to this, for example, the degree of dispersion is determined with reference to both the dispersion and the number of noise suppression gains G (n) that fit in the space.

It is a block diagram which shows the structure of the noise suppression apparatus which concerns on 1st Embodiment of this invention. It is a flowchart which shows the flow of the noise suppression process which concerns on 1st Embodiment. It is explanatory drawing for demonstrating the content of the noise suppression process which concerns on 1st Embodiment. It is explanatory drawing for demonstrating the content of the conventional noise suppression process. It is a block diagram which shows the structure of the noise suppression apparatus which concerns on 2nd Embodiment of this invention. It is a graph which shows the example of a calculation of the noise suppression gain G (n) in case an audio | voice signal is included. It is a graph which shows the example of calculation of noise suppression gain G (n) in case an audio signal is not included.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1,1 '... Noise suppression apparatus, 10 ... Time / frequency conversion part, 20 ... Noise spectrum estimation part, 30 ... Noise suppression gain calculation part, 40 ... Noise period / noise suppression gain calculation part (noise period) Gain calculation unit), 50 ... original sound addition rate calculation unit, 60 ... original sound addition gain calculation unit, 70 ... frequency / time conversion unit, 11 ... multiplier Y (n) ... amplitude spectrum of the input signal, N (n): Noise spectrum estimation value, G (n): Noise suppression gain, g: Noise suppression gain average value, G _t , G1 (n): Noise suppression gain for application to the noise period ( (Noise period gain), og ... original sound addition rate, tg ... target noise suppression gain, TG ... target noise suppression amount, OG ... original sound addition ratio, G2 (n) ... noise suppression gain after addition of original sound ( Corrected gain)

Claims (12)

Noise spectrum estimation means for estimating a noise spectrum included in an input signal for each of K frequency bands (where K is a natural number of 2 or more), based on the input signal;
First gain calculation means for calculating a noise suppression gain for each of the K frequency bands based on an estimation result by the noise spectrum estimation means;
The first noise-suppressed signal obtained as a result of applying the noise suppression gain to the input signal is determined based on a difference between the noise suppression gain and a target noise suppression gain indicating a target level of noise suppression. Original sound adding means for adding the input signal at a first ratio;
A noise suppression device comprising: A second gain calculating means for calculating an average value gain for the K frequency bands for the noise suppression gain;
The original sound adding means is
The second signal after noise suppression obtained as a result of applying the average gain to the input signal is input to the input signal at a second ratio determined based on the difference between the average gain and the target noise suppression gain. to add,
The noise suppression apparatus according to claim 1. The original sound adding means is
Calculating a smoothing ratio obtained by smoothing the first or second ratio on the time axis;
Adding the input signal at the smoothing rate to the first or second noise-suppressed signal;
The noise suppression apparatus according to claim 1 or 2, wherein Further comprising voice detection means for classifying the input signal into a voice frame including the voice and a noise frame not including the voice over time by detecting the presence or absence of the voice included in the input signal;
The second noise-suppressed signal is
Obtained as a result of applying the average gain to the portion corresponding to the noise frame of the input signal,
The noise suppression apparatus according to claim 2. The first noise-suppressed signal is
Obtained as a result of applying the noise suppression gain to the portion of the input signal corresponding to the speech frame,
The noise suppressor according to claim 4. The original sound adding means is
Adding the input signal to the first noise-suppressed signal at the second rate instead of the first rate;
The noise suppression apparatus according to claim 5. The original sound adding means is
When trying to add the input signal to the first noise-suppressed signal related to the speech frame,
Adding the input signal to the first noise-suppressed signal, assuming that the second ratio already calculated for the noise frame nearest to the voice frame is the first ratio in the voice frame;
The noise suppression apparatus according to claim 5. The original sound adding means is
When trying to add the input signal to the second noise-suppressed signal related to the noise frame,
After the second ratio in the noise frame is calculated as a temporary second ratio, the temporary second ratio is smoothed on the time axis using the first or second ratio in the frame immediately before the noise frame. And calculating the smoothed ratio, and assuming that the smoothed ratio is the second ratio, adding the input signal to the second noise-suppressed signal,
When trying to add the input signal to the first noise-suppressed signal related to the speech frame,
Adding the input signal to the first noise-suppressed signal, assuming that the first or second ratio in the frame immediately before the voice frame is the first ratio in the voice frame;
The noise suppression apparatus according to claim 5 or 7, wherein The first or second ratio is:
The noise suppression apparatus according to claim 1, wherein the noise suppression apparatus is obtained by the following expression (A).
og = max (0, tg−G) (A)
Where og is the first or second ratio to be determined,
tg is the target noise suppression gain,
G is the noise suppression gain or the average gain,
max (a, b) means a function that returns a larger value of a and b. A noise spectrum estimation step for estimating a noise spectrum included in the input signal for each of K frequency bands (where K is a natural number of 2 or more);
A first gain calculation step of calculating a noise suppression gain for each of the K frequency bands based on an estimation result in the noise spectrum estimation step;
The first noise-suppressed signal obtained as a result of applying the noise suppression gain to the input signal is determined based on a difference between the noise suppression gain and a target noise suppression gain indicating a target level of noise suppression. An original sound adding step of adding the input signal at a first ratio;
Including a noise suppression method. A second gain calculating step of calculating an average value gain for the K frequency bands for the noise suppression gain;
In the original sound adding step,
The second signal after noise suppression obtained as a result of applying the average value gain to the input signal has the input signal at a second ratio determined based on the difference between the average value gain and the target noise suppression gain. To be added,
The noise suppression method according to claim 10. A voice detecting step of detecting the presence or absence of a voice included in the input signal, and classifying the input signal into a voice frame including the voice and a noise frame not including the voice over time;
The second noise-suppressed signal is
Obtained as a result of applying the average gain to the portion corresponding to the noise frame of the input signal,
The noise suppression method according to claim 11.

Images