JP5475697B2 - Noise suppressor, method and program thereof - Google Patents

Description

  The present invention relates to a noise suppression apparatus, a method thereof, and a program that are hardly affected by a local estimation error of a noise spectrum.

  The noise suppression device is a device that removes noise superimposed on a desired audio signal. FIG. 13 shows a functional configuration of a typical noise suppression apparatus 900, and its operation will be briefly described. The noise suppression apparatus 900 includes a spectrum calculation unit 90, a noise estimation unit 91, a gain calculation unit 92, a filter calculation unit 93, and a filter application unit 94.

  The spectrum calculation unit 90 calculates an observation spectrum that is a spectrum of the observation signal from the observation signal for each short-time frame having a predetermined time width. The noise estimation unit 91 calculates a noise spectrum that is a spectrum of noise included in the observation spectrum. The gain calculation unit 92 calculates a noise suppression gain. The noise suppression gain is a real vector of 0 or more and 1 or less defined for each frequency bin, and by multiplying this by the observed spectrum, an enhanced spectrum that is a spectrum in which noise is suppressed is obtained.

  The filter calculation unit 93 converts the noise suppression gain defined in the frequency domain into a time domain noise suppression filter. The filter application unit 94 obtains an enhanced signal in which noise is suppressed by applying a noise suppression filter to the observation signal. In addition, a configuration for obtaining an enhancement signal from an enhancement spectrum obtained by multiplying an observation spectrum by a noise suppression gain is also widely used.

  As for the gain calculation unit 92, a method of calculating a noise suppression gain called OMLSA (Optimally Modified Log-Spectral Amplitude Estimator) is known (Non-patent Document 1). The method will be described with reference to FIG. The OMLSA includes a speech existence probability calculation unit 920, a first desired speech estimation unit 921, a second desired speech estimation unit 922, and a gain determination unit 923.

  The voice presence probability calculation unit 920 calculates a voice presence probability vector. The speech presence probability vector is a vector over all frequency bins of the speech presence probability defined for each frequency bin. The voice existence probability in a certain frequency bin represents the probability that the voice exists in the observed spectrum in the frequency bin. The first desired speech estimation unit 921 calculates an estimated value of the spectrum of the desired speech as a first enhancement spectrum when it is assumed that speech exists in the observed spectrum in all frequency bins.

  The second desired speech estimation means 922 calculates the estimated value of the spectrum of the desired speech as the second enhancement spectrum when it is assumed that there is no speech in the observed spectrum in all frequency bins. Gain determining means 923 calculates a noise suppression gain from the speech presence vector, the first enhancement spectrum, and the second enhancement spectrum.

  The noise suppression apparatus 900 using OMLSA can achieve both a small voice distortion and a high noise suppression amount. Since voice has sparsity, it is known that voice energy tends to concentrate in some frequency bins. In such a frequency bin, it can be considered that sound exists. On the other hand, it can be considered that there is no sound in a frequency bin where there is almost no sound energy. In OMLSA, noise suppression gain is calculated while estimating whether or not speech exists for each frequency bin. Therefore, the energy of the desired speech is kept small while keeping the speech distortion in the frequency bin where the energy of the desired speech is concentrated. Noise can be significantly suppressed in frequency bins that are almost nonexistent.

I. Cohen, “Optimal speech enhancement under signal presence uncertainty using log-spectral amplitude estimator,” IEEE SP Letters, vol.9, no.4, pp.113-116, 2002.

  However, the OMLSA speech existence probability calculation has a problem that the speech existence probability vector cannot be obtained correctly when the noise is non-stationary. In the OMLSA speech existence probability calculation, the SN ratio in a frequency bin that is close in time or frequency to the frequency bin is obtained for each frequency bin, and the speech existence probability is calculated based on the magnitude.

  If the noise is non-stationary, the estimate of the noise spectrum tends to include local errors. For this reason, since an S / N ratio cannot be calculated | required correctly, a suitable audio | voice existence probability cannot be obtained. That is, OMLSA is not robust to local errors in noise spectrum estimates.

  The present invention has been made in view of the above problems, and by focusing on the general structural features of speech, the speech existence probability is highly accurate even when the noise spectrum estimation value includes an error. It is an object of the present invention to provide a noise suppression device, a method thereof, and a program that can be obtained. The general structural features of speech focused on in the present invention are that the logarithmic spectrum envelope can be accurately modeled with a mixed normal distribution and has a harmonic structure depending on the fundamental frequency. The logarithmic spectral envelope and the fundamental frequency physically correspond to phoneme and voice pitch, respectively.

  The noise suppression device of the present invention includes a spectrum calculation unit, a logarithm calculation unit, a noise statistic estimation unit, a gain calculation unit, a filter calculation unit, and a filter application unit, and the gain calculation unit An existence probability calculating means, a first desired speech estimating means, a second desired speech estimating means, and a gain determining means are provided.

The speech presence probability calculation means receives the observed log spectrum, the noise average log spectrum, and the noise variance log spectrum as inputs, and calculates a speech presence probability vector using a degraded speech log spectrum envelope model and a harmonic structure model. The first desired speech estimation means receives the observed logarithm spectrum, the noise average logarithm spectrum, and the noise variance logarithm spectrum as inputs, and the first emphasized logarithm that is an estimate of the logarithmic spectrum of the desired speech when speech is present in the observed logarithm spectrum. Calculate the spectrum. Second desired speech estimation unit is input with the observation logarithmic spectrum, calculating a second emphasis logarithmic spectrum is an estimate of the log spectrum of a desired sound when the audio during observation log spectrum does not exist. The gain determining means calculates the noise suppression gain with the observed log spectrum, the speech presence probability vector, the first emphasized log spectrum, and the second emphasized log spectrum as inputs.

  The speech existence probability calculating means of the noise suppression apparatus according to the present invention calculates the speech existence probability in consideration of the logarithmic spectrum envelope and the harmonic structure of the desired speech. That is, since the speech existence probability is calculated in consideration of the global structure of the logarithmic spectrum of the desired speech, it is less likely to be affected by local errors included in the estimated values of the noise average logarithmic spectrum and noise variance logarithmic spectrum. be able to.

The figure which illustrates the spectrum of the desired audio | voice and the deterioration audio | voice in the voiced area of an audio | voice. The figure which shows the prototype of a harmonic structure. The figure which shows the logarithmic spectrum envelope prototype of an audio | voice. The figure which shows the function structural example of the noise suppression apparatus 100 of this invention. The figure which shows the operation | movement flow of the noise suppression apparatus. The figure which shows the function structural example of the noise statistics estimation part 11. FIG. The figure which shows the function structural example of the gain calculation part 12. FIG. The figure which shows the operation | movement flow of the gain calculation part 12. FIG. The figure which shows the function structural example of the voice presence probability calculation means 120. FIG. The figure which shows the operation | movement flow of the voice presence probability calculation means 120. The figure which shows the function structural example of the 1st desired audio | voice estimation means 121. FIG. The figure which shows an evaluation experiment result. The figure which shows the function structure of the conventional noise suppression apparatus 900. The figure which shows the function structure of the gain calculation part.

    Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

[Basic idea of the present invention]
An object of the present invention is to provide a method for obtaining a speech existence probability robust to a local estimation error of a noise spectrum by calculating a speech existence probability in consideration of a general structural feature of speech. is there.

  First, the concept of speech existence probability will be described with reference to FIG. FIG. 1 illustrates the spectrum of desired speech and degraded speech in a voiced section of speech. In the frequency bin included in the range indicated by the arrow, the difference between the deteriorated voice and the desired voice is extremely small. On the other hand, in other frequency bins, the difference is extremely large. This is due to the sparseness of speech. Sparsity means that voice energy tends to concentrate on some frequency bins. In such frequency bins, the difference between degraded speech and desired speech is small. On the other hand, in other frequency bins, the degraded speech contains almost no noise. Therefore, if a frequency bin in which voice energy indicated by an arrow is concentrated can be identified, noise can be greatly suppressed in other frequency bins while keeping the voice distortion in such frequency bins small. The voice presence probability is defined for each frequency bin and represents the probability that the energy of the voice is concentrated in the frequency bin.

  In the present invention, the speech existence probability is calculated using the general structural features of two types of speech. One is that frequency bins in which voice energy is concentrated tend to appear in the vicinity of an integral multiple of the fundamental frequency. This is a feature known as the harmonic structure of speech. Therefore, if a prototype of a harmonic structure for each fundamental frequency as shown in FIG. 2 is prepared in advance and it is determined which prototype fits the observed spectrum, the probability of speech existence considering the harmonic structure is obtained. It is done.

  Another feature relates to the logarithmic spectral envelope of speech. The logarithmic spectral envelope of speech represents phonemes. Therefore, it can be approximately expressed by a finite number of prototypes as shown in FIG. Specifically, in the embodiment described below, a logarithmic envelope model of speech expressed by a mixed normal distribution is used instead of a set of discrete prototypes. Each element distribution of the mixed normal distribution corresponds to one prototype of the logarithmic spectral envelope. When calculating the degree of fit between the observed spectrum and the prototype of the harmonic structure, by using this mixed normal distribution at the same time, it is possible to obtain the speech existence probability considering both the logarithmic spectrum envelope and the harmonic structure of the speech. . Since a pre-defined harmonic structure and logarithmic spectrum envelope model are used, the speech existence probability can be calculated robustly against a local estimation error of the noise spectrum.

  FIG. 4 shows a functional configuration example of the noise suppression apparatus 100 of the present invention. The operation flow is shown in FIG.

  The noise suppression apparatus 100 includes a spectrum calculation unit 90, a logarithm calculation unit 10, a noise statistic estimation unit 11, a gain calculation unit 12, a filter calculation unit 93, and a filter application unit 94. The spectrum calculation unit 90, the filter calculation unit 93, and the filter application unit 94 are the same as the noise suppression device 900 described in the related art. The noise suppression apparatus 100 is realized by a predetermined program being read into a computer including, for example, a ROM, a RAM, and a CPU, and the CPU executing the program.

Spectrum calculation unit 90, the observation signals of short time frame _{y T = {y (1)} , ..., y (N)} is its power spectrum as an input the observed spectrum Y = _{a, {Y 1, ..., Y} L} Is calculated (step S90). Here, N represents the frame width, and L represents the number of frequency bins. Calculation of the observed spectra Y, after decomposition into components of each frequency bin of the observation signal y _T by fast Fourier transform, it is performed by taking the square of the absolute value of the amplitude of each frequency component. Instead of taking the square of the absolute value of the amplitude, the absolute value of the amplitude may be ζ power (ζ is an arbitrary positive real number). Also, any spectral analysis method may be used instead of the fast Fourier transform.

The logarithmic calculator 10 receives the observed spectrum Y and calculates an observed logarithmic spectrum y = {y ₁ ,..., Y _L } as a logarithmic value (step S10).

The noise statistic estimation unit 11 receives the observed logarithmic spectrum y as an input, and the noise average logarithmic spectrum μ ^N = {μ ₁ ^N ,..., Μ _L which is an average estimated value of the logarithmic spectrum of noise included in the observed signal y _T. ^N } and the noise variance logarithm spectrum σ ^N = {σ ₁ ^N ,..., Σ _L ^N }, which is an estimated value of the variance, is calculated (step S11). A specific method for calculating the noise average logarithmic spectrum μ ^N and the noise variance logarithmic spectrum σ ^N will be described later.

The gain calculation unit 12 calculates the noise suppression gain g = {g ₁ ,..., G _L } using the observed log spectrum y, the noise average log spectrum μ ^N and the noise variance log spectrum σ ^N as inputs (step S12).

The filter calculation unit 93 converts the noise suppression gain g into the noise suppression filter g _T = {g (−J),..., G (J)} by inverse discrete cosine transform (step S93). Here, J is an integer such that 2J + 1 is the order of the noise suppression filter. Instead of applying an inverse discrete cosine transform to g, it was folded _{g {g 1, ..., g} L-1, g L, ¯g L-1, ..., ¯g 2} opposite to that A discrete Fourier transform may be applied, or the Parks-McClellan algorithm may be used. Here, ￣x represents a complex conjugate of x, and ￣ is described correctly on a variable.

The filter application unit 94 receives the observation signal y _T and the noise suppression filter g _T as inputs, and calculates an enhancement signal u _T = {u (1),. (Step S94).

Here, it is assumed that the value of y (n) for n ≦ 0 and n> N is zero. Instead of obtaining the enhanced signal u _T using the filter calculation unit 93 and the filter applying unit 94, it may be obtained emphasis signal u _T from enhancement spectrum obtained by multiplying the noise suppression gain g to the observed spectrum Y.

  The noise statistic estimation unit 11 and the gain calculation unit 12, which are the main parts of the present invention, will be described in more detail with reference to more specific functional configuration examples.

[Noise statistics estimation unit]
FIG. 6 shows a functional configuration example of the noise statistic estimation unit 11. The noise statistic estimation unit 11 includes a speech section detection unit 110, a noise logarithmic spectrum recording unit 111, and a statistic calculation unit 112. In the configuration of the noise statistic estimation unit 11, a silence interval that does not include speech is detected from the observed signal, and a noise statistic is calculated from the interval.

  The speech section detection means 110 receives the observed log spectrum y as an input and calculates a speech section flag z indicating whether or not the current short time frame belongs to the speech section. The voice section flag z is a binary variable. If z = 1, the current short frame belongs to the voice section, and if z = 0, the current short frame does not belong to the voice section. Any known voice detection section detection means can be used as the voice section detection means 110. Therefore, the detailed operation description is omitted.

  The logarithmic spectrum recording means 111 records the logarithmic spectrum of B noises together with the recording time. The logarithmic spectrum recording means 111 receives the observed logarithmic spectrum y and the speech interval flag z, does nothing if z = 1, and observes the oldest logarithmic spectrum of B noises if z = 0. Replace with logarithmic spectrum y.

Statistic calculation unit 112, the noise log spectrum recording unit 111 from the read log spectrum of B-number of the noise, calculating its average and is the mean noise is a logarithmic spectrum mu ^N and variance noise variance logarithmic spectrum sigma ^N. Note that the configuration of the noise statistic estimation unit 11 can be arbitrarily changed within the constraint that the noise logarithm spectrum μ ^N and the noise variance log spectrum σ ^N are calculated with at least the observed log spectrum as an input.

For example, the noise statistics estimator 11 inputs the observed signal y _T in addition to the observed log spectrum y, voice activity detection means 110 may be using the observed signal y _T as an input. Alternatively, instead of using the speech interval detection means 110, reference literature ("Noise spectrum estimation in adverse environments: improved minima controlled recursive averaging," IEEE Trans. SAP, vol. 11, no. 5, pp.466-475, 2003 .) May be used to estimate the logarithmic spectrum of noise and calculate the mean and variance thereof.

[Gain calculation section]
FIG. 7 shows a functional configuration example of the gain calculation unit 12. The operation flow is shown in FIG. The gain calculation unit 12 includes a speech presence probability calculation unit 120, a first desired speech estimation unit 121, a second desired speech estimation unit 122, and a gain determination unit 123.

The speech existence probability calculation means 120 receives the observed logarithmic spectrum y, the noise average logarithmic spectrum μ ^N, and the noise variance logarithmic spectrum σ ^N as input, and the speech existence probability vector r = {r ₁ ,..., R _L } and KH states. A posteriori probability {w _1,1 ,..., W _{1, H} ,..., W _{K, H} } is calculated (step S120).

First desired speech estimation means 121, observations and log spectrum y, the set of KH number of states posterior probability _{W = {w 1,1, ...,} w 1, H, ..., w K, H} and the noise mean log Using the spectrum μ ^N and the noise variance logarithmic spectrum σ ^N as inputs, the first enhanced logarithmic spectrum ^ x ⁽¹⁾ = {^ x ₁ ⁽¹⁾ ,..., ^ _L ⁽¹⁾ } is calculated (step S121). . W is called a state posterior probability set. A method of calculating the first emphasized logarithmic spectrum ^ x ⁽¹⁾ will be described later.

The second desired speech estimation means 122 receives the observed logarithmic spectrum y as an input and uses the expression (2) to obtain the second enhanced logarithmic spectrum ^ x ⁽²⁾ = {^ x ₁ ⁽²⁾ , ..., ^ x _L ⁽²⁾ }. Calculate (step S122).

Here, g _FLR is a given constant representing the noise suppression gain when it is assumed that there is no speech in the observed logarithmic spectrum y. Instead of calculating the second enhanced logarithmic spectrum ^ x ⁽²⁾ using equation (2), any method for estimating the logarithmic spectrum of the desired speech when no speech is present in the observed logarithmic spectrum y May be used. For example, the method disclosed in the reference (M. Fujimoto, et al., “Study of integration of statistical model-based voice activity detection and noise suppression,” in Proc. ICSLP, 2008, pp. 2008-2011.) Can also be used.

The gain determining means 123 receives the observed logarithmic spectrum y, the first emphasized logarithmic spectrum ^ x ⁽¹⁾ , the second emphasized logarithmic spectrum ^ x ^(2), and the speech existence probability vector r as noises according to the following equation. The suppression gain g = {g ₁ ,..., G _L } is calculated (step S123).

  FIG. 9 shows a more specific functional configuration example of the voice presence probability calculation means 120, and the operation will be described in more detail. The operation flow of the speech existence probability calculation means 120 is shown in FIG. The speech existence probability calculation unit 120 includes an undegraded speech log spectrum envelope model recording unit 1200, a model synthesis unit 1201, a degraded speech log spectrum envelope model recording unit 1202, a harmonic structure model recording unit 1203, and a conditional speech presence. Probability calculation means 1204, state posterior probability calculation means 1205, and speech presence probability determination means 1206 are provided.

The undegraded speech logarithmic spectrum envelope model recording means 1200 records a logarithmic spectrum envelope model of undegraded speech. Specifically, assuming that the model of the non-degraded speech logarithmic spectrum envelope is expressed in the form of a mixed normal distribution, the weight, average spectrum, and variance spectrum of each element distribution are recorded. The average spectrum of each element distribution is referred to as an undegraded speech average logarithmic spectrum, and its dispersion spectrum is referred to as an undegraded speech distributed logarithmic spectrum. In this embodiment, the number of element distributions is K, and the k-th non-degraded speech average logarithmic spectrum envelope is μ _k ^X = {μ _{k, 1} ^X ,..., Μ _{k, L} ^X }, k-th non-degraded speech. The variance logarithmic spectrum envelope is expressed as σ _k ^X = {σ _{k, 1} ^X ,..., Σ _{k, L} ^X }, and the k-th weight is expressed as π _k . The non-degrading speech average logarithmic spectrum envelope μ _k ^X , the non-degrading speech variance logarithmic spectrum envelope σ _k ^X, and the weight π _{k of} each element distribution are calculated in advance from the database of undegraded speech by the EM algorithm.

After the noise average logarithmic spectrum μ ^N and the noise variance logarithmic spectrum σ ^N are input, the model synthesizing unit 1201 receives the non-degraded speech average logarithmic spectrum envelope of each element distribution k from the non-degraded speech logarithmic spectrum envelope model recording unit 1200. Read μ _k ^X , undegraded speech variance logarithmic spectrum envelope σ _k ^X and weight π _k , and degraded speech average logarithmic spectrum envelope μ _k ^Y = {μ _{k, 1} ^Y ,..., μ _{k, L} ^Y } and degraded speech variance Logarithmic spectral envelope σ _k ^Y = {σ _{k, 1} ^Y ,..., Σ _{k, L} ^Y } is calculated. The calculated degraded speech average logarithmic spectrum envelope μ _k ^Y and degraded speech variance logarithmic spectrum envelope σ _k ^Y are recorded in the degraded speech logarithmic spectrum envelope model recording means 1202 together with the weight π _k . The weight of each element distribution, the average logarithmic spectrum envelope of degraded speech, and the logarithmic envelope of degraded speech variance define a logarithmic spectrum envelope model of degraded speech.

The degraded speech average logarithmic spectrum envelope μ _k ^Y is _defined by equation (5), and the non-degraded speech variance logarithmic spectrum envelope σ _k ^Y is defined by equation (6).

  Equations (5) and (6) are based on the reference (A. Acero, L. Deng, T. Kristjansson, and J. Zhang, “HMM adaptation using vector Taylor series for noisy speech recognition,” in Proc. Int '. l Conf. Spoken Lang. Process., vol.3, 2000, pp. 869-872.), but based on VTS (Vector Taylor Series), such as PMC (Parallel Model Combination) A similar model synthesis method may be used.

As a method for calculating the degraded speech average logarithmic spectrum envelope μ _k ^{Y of} each element distribution recorded in the degraded speech logarithmic spectrum envelope model recording means 1202, the degraded speech variance logarithmic spectrum envelope σ _k ^Y and the weight π _k , these are not used. Although an example of calculation using the degraded speech log spectrum envelope model recording unit 1200 and the model synthesis unit 1201 has been described, the present invention is not limited to this example. For example, the degraded speech average logarithmic spectrum envelope μ _k ^Y , the degraded speech variance logarithmic spectrum envelope σ _k ^Y and the weight π _k of each element distribution may be calculated in advance from a database of degraded speech using an EM algorithm.

The harmonic structure model recording unit 1203 records a model of a speech presence prior probability vector that is a vector of a prior probability for calculating a speech presence probability vector. Specifically, it is assumed that the speech presence prior probability vector is expressed in the form of a codebook, and the speech presence prior probability vector corresponding to each code and its weight are recorded. In this embodiment, the codebook size is H, the speech existence prior probability vector corresponding to the h-th code is expressed as q _h = {q _{h, 1} ,..., Q _{h, L} }, and the weight is expressed as θ _h. To do. Codebook, for example, the fundamental frequency o _h corresponding to h-th code is calculated by Equation (7).

Here, H _L and H _H are given constants representing the lower and upper limits of the range in which the fundamental frequency exists, respectively. Using this fundamental frequency, the speech presence prior probability vector q _h is calculated by Equation (8).

Here, f _l represents the center frequency of the frequency bin l and is calculated by f _l = lf _S / 2L. f _S is a sampling frequency. β _h is a constant for limiting the speech presence prior probability vector q _h to a value in the range of 0-1. Γ and δ are given constants, and f _G is a probability density function of a generalized normal distribution. Note that the method of creating the H speech existence prior probability vectors recorded in the harmonic structure model recording unit 1203 is not limited to the above-described method, and the value of the speech existence prior probability vector q _h is set to 0 to 1. Any method may be used as long as it can fall within the above range.

Conditional speech presence probability calculation unit 1204 is composed of a total of KH number of partial conditional speech presence probability calculator unit ₁₂₀₄ 11 ～1204 _KH to one Class 1 ~K species H class. The partial conditional speech existence probability calculation unit 1204 _kh receives the observed logarithmic spectrum y as an input, the degraded speech average logarithmic spectrum envelope μ _k ^{Y of the} k-th element distribution from the degraded speech logarithmic spectrum envelope model recording means 1202 and the degraded speech variance logarithm. The speech envelope prior probability vector q _h corresponding to the h-th code is read out from the harmonic structure model recording means 1203 as the spectrum envelope σ _k ^Y. Then, the speech existence probability vector r _{k, h} = {r _k, when the k th element distribution is selected as the envelope model of the observed log spectrum y and the h th code is selected as the harmonic structure model of the observed log spectrum y _{. h, 1} ,..., rk _{, h, L} } are calculated and output to the speech existence probability determining means 1206 (step S1204). r _{k, h} are referred to as k-type h-class conditional speech existence probability vectors. Also, KH conditional speech existence probability vectors are collectively written as R = {r _1,1 ,..., R _{K, H} }, and this is called a conditional speech existence probability vector set.

rk _{, h} called k-type h-class conditional speech existence probability vector is calculated by equation (9).

Here, p _N (x) and p _{Y | K} (x | k) are normal distributions corresponding to the normal distribution of the noise logarithmic spectrum and the kth element distribution of the model of the degraded logarithmic spectrum envelope, respectively. It is defined by the following formula.

Here, f _N is the probability density function of the normal distribution.

State posterior probability calculation unit 1205 is composed of a total of KH pieces of partial state posterior probability calculation unit ₁₂₀₅ 11 ～1205 _KH to one Class 1 ~K species H class. The partial state posterior probability calculation unit 1205 _kh receives the observed logarithmic spectrum y as an input, the degraded speech average logarithmic spectrum envelope μ _k ^{Y of the} k-th element distribution from the degraded speech logarithmic spectrum envelope recording means 1202, and the degraded speech variance logarithmic spectrum envelope. The σ _k ^Y and the weight π _k are read out from the harmonic structure model recording unit 1203 as the speech prior probability vector q _h and the weight θ _h corresponding to the h-th code, and the envelope model of the observed logarithmic spectrum y is the k-th model. A posteriori probabilities w _{k, h} that are element distributions and whose harmonic structure model of the observed log spectrum envelope is the h-th code are calculated and output (step S1205). Let w _{k, h} be k-type h-class state posterior probabilities. The k-type h-class state posterior probability w _{k, h} is calculated by the following equation.

Speech presence probability determination unit 1206, one first-type ~K species H such conditional speech presence probability vector _{{r 1,1, ..., r 1} , H, ..., r K, H} and, one first-type- K species H such state posterior probability _{{w 1,1, ..., w 1} , H, ..., w K, H} speech presence probability vector _{r l} = as input _{{r 1, ..., r J} } formula (14 ) And output to the gain determination means 123 (step S1206).

FIG. 11 shows a more specific functional configuration example of the first desired speech estimation means 121 and its operation will be described. The first desired speech estimation unit 121 includes a conditional desired speech estimation unit 1210 and a desired speech estimation value determination unit 1211. Conditional desired sound estimation unit 1210 is composed of the K total partial conditional desired sound estimation unit 1210 ₁ ～1210 _K to one ~K species.

The partially conditional desired speech estimation unit 1210 _k receives the observed logarithmic spectrum y as an input, and the undegraded speech average logarithmic spectrum of the K-th element distribution from the undegraded speech logarithmic spectrum envelope model recording unit 1200 of the speech existence probability calculating unit 120. An envelope μ _k ^X and an undegraded speech variance logarithmic spectrum envelope σ _k ^X are read out, and k type emphasis that is an estimate of the logarithmic spectrum of the desired speech when the k-th element distribution is selected as the envelope model of the observed logarithmic spectrum The logarithmic spectrum { _{circumflex over} (x) _{} k} = {ｘx _{k, 1} ,..., Ｘx _{k, L} } is calculated and output to the desired speech estimation value determining means 1211. The k-weighted logarithmic spectrum is calculated according to equation (15).

Desired speech estimation value determination means 1211 calculates the first enhanced logarithmic spectrum {^ x ₁ ⁽¹⁾ ,..., ^ X _L ⁽¹⁾ } according to the following equation, and outputs this to gain determination section 123.

  In addition, although the example using the model of a non-degraded speech logarithmic spectrum envelope and a degraded speech logarithmic spectrum envelope has been described in the first desired speech estimation unit 121, the configuration method of the first desired speech estimation unit 121 is not limited to this example. For example, the first desired speech estimation means of the gain calculator disclosed in Non-Patent Document 1 may be used without using the model of the non-degraded speech logarithmic spectrum envelope and the degraded speech logarithmic spectrum envelope.

[Evaluation experiment]
For the purpose of confirming the effect of the noise suppression device of the present invention, an evaluation experiment was performed to compare the signal-to-noise ratio obtained by the noise suppression device 100 and the noise suppression device disclosed in Non-Patent Document 1. The result is shown in FIG. The horizontal axis in FIG. 12 is time (seconds), and the vertical axis is the SN ratio (dB).

  For the observation signal used in the experiment, first, a desired voice signal of about 2 seconds for one woman who reads out three consecutive numbers is prepared, and bubble noise is superimposed on the signal so that the SN ratio is 5 dB. Created. Bubble noise is a typical example of non-stationary noise.

  A one-dot chain line in FIG. 12 indicates a time series of the S / N ratio before noise suppression. A thick solid line is a time series of the S / N ratio after noise suppression by the noise suppression apparatus 100 of the present invention. A thin solid line is a time series of the S / N ratio after noise suppression by the noise suppression device disclosed in Non-Patent Document 1.

  It can be seen that the noise suppression apparatus 100 of the present invention can improve the SN ratio by about 5 dB compared to the conventional apparatus. This result suggests that the noise suppression device 100 of the present invention is more robust against non-stationary noise than the conventional noise suppression device.

  As described above, the noise suppression apparatus 100 according to the present invention calculates the speech existence probability in consideration of the logarithmic spectrum envelope and the harmonic structure of the desired speech. As a result, the speech existence probability can be obtained robustly against a local estimation error of the noise spectrum that occurs when the degree of noise non-stationarity is large. Therefore, it is possible to realize a noise removal device that is robust against non-stationary noise.

  The present invention is not limited to the above-described embodiments. Various modifications can be made to the noise suppression device within the scope of the technical idea of obtaining the speech existence probability in consideration of the logarithmic spectrum envelope and harmonic structure of the desired speech of the present invention.

  Further, the processes described in the above method and apparatus are not only executed in time series according to the order of description, but also may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Good.

  When the processing means in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described as a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on any computer-readable recording medium. Examples of the computer-readable recording medium include a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory. More specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape, etc., and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read) Only Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording media, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  This program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each unit may be configured by executing a predetermined program on a computer, or at least a part of the processing contents may be realized as hardware.

Claims (5)

A spectrum calculation unit that calculates an observation spectrum that is a power spectrum of an observation signal as an input,
A logarithmic calculator for calculating the observed logarithm spectrum as a logarithmic value with the observed spectrum as an input;
A noise statistic estimator that calculates the mean noise logarithm spectrum and the noise variance logarithm spectrum with the observed logarithm spectrum as an input;
A gain calculation unit for calculating a noise suppression gain using the noise average log spectrum, the noise variance log spectrum and the observed log spectrum as inputs;
A filter calculation unit for calculating a noise suppression filter using the noise suppression gain as an input;
A filter application unit that calculates the enhancement signal with the observation signal and the noise suppression filter as inputs; and
The gain calculation unit receives the observed logarithmic spectrum, the noise average logarithmic spectrum, and the noise variance logarithmic spectrum as inputs, and calculates a speech existence probability vector using a degraded speech logarithmic spectrum envelope model and a harmonic structure model. Existence probability calculation means,
First input logarithmic spectrum, which is an estimate of the logarithmic spectrum of a desired speech when speech is present in the observed spectrum, is input using the observed logarithmic spectrum, the noise average logarithmic spectrum, and the noise variance logarithmic spectrum as inputs. A desired speech estimation means;
As input the observed logarithmic spectrum, a second desired speech estimation means for calculating a second emphasis logarithmic spectrum is an estimate of the log spectrum of a desired sound when no speech is present in the observed spectrum,
Gain determining means for calculating the noise suppression gain using the observed log spectrum, the speech presence probability vector, the first emphasized log spectrum, and the second emphasized log spectrum as inputs;
A noise suppression device comprising: The noise suppression device according to claim 1,
The voice existence probability calculating means is:
Degraded speech logarithmic spectrum envelope model record that records the degraded speech average logarithmic envelope, degraded speech variance logarithmic envelope, and weights of each element distribution that defines the degraded speech mixed normal distribution, which is a model of the logarithmic spectrum envelope of degraded speech Means,
Harmonic structure model recording means for recording a codebook of a speech presence prior probability vector that is a harmonic structure model;
For each combination of the element distribution of the degraded speech mixed normal distribution and the codebook code with the observed logarithmic spectrum as an input, the conditional speech existence probability vector when the element distribution and the code are selected is calculated. Conditional speech existence probability calculation means,
State posterior probability calculating means for calculating the state posterior probability that the element distribution and the sign are selected for each of all the combinations using the observed log spectrum as an input;
Speech presence probability determining means for calculating the speech presence probability vector using the conditional speech presence probability vector and the state posterior probability for all the combinations as inputs;
A noise suppression device comprising: A spectrum calculation process for calculating an observation spectrum that is a power spectrum with an observation signal as an input,
A logarithmic calculation process of calculating the observed logarithm spectrum as a logarithmic value with the observed spectrum as an input;
A noise statistic estimation process for calculating a noise average logarithm spectrum and a noise variance logarithm spectrum with the observed logarithm spectrum as an input;
A gain calculation process for calculating a noise suppression gain using the noise average log spectrum, the noise variance log spectrum and the observed log spectrum as inputs;
A filter calculation process for calculating a noise suppression filter using the noise suppression gain as an input;
A filter application process for calculating an enhancement signal using the observation signal and the noise suppression filter as inputs,
The gain calculation process uses the observed logarithmic spectrum, the noise average logarithmic spectrum, and the noise variance logarithmic spectrum as inputs, and calculates a speech existence probability vector using a degraded speech logarithmic spectrum envelope model and a harmonic structure model. An existence probability calculation step;
First input logarithmic spectrum, which is an estimate of the logarithmic spectrum of a desired speech when speech is present in the observed spectrum, is input using the observed logarithmic spectrum, the noise average logarithmic spectrum, and the noise variance logarithmic spectrum as inputs. A desired speech estimation step;
As input the observed logarithmic spectrum, a second desired speech estimation step of calculating a second emphasis logarithmic spectrum is an estimate of the log spectrum of a desired sound when there is no voice in the observed spectrum,
A gain determining step of calculating the noise suppression gain by inputting the observed log spectrum, the speech existence probability vector, the first enhanced log spectrum, and the second enhanced log spectrum;
Including a noise suppression method. In the noise suppression method according to claim 3,
The speech existence probability calculation step includes
Using the observed logarithmic spectrum as an input, the element distribution of the degraded speech mixed normal distribution, which is a model of the logarithmic spectrum envelope of the degraded speech recorded in the degraded speech logarithmic envelope model recording means, and the modulation recorded in the harmonic structure model recording means. Conditional speech existence probability calculation that calculates a conditional speech presence probability vector when the element distribution and the code are selected for each combination of codes in the codebook of speech presence prior probability vectors that are wave structure models Steps,
A state posterior probability calculation step for calculating a state posterior probability that the element distribution and the code are selected for each of all the combinations, using the observed log spectrum as an input,
A speech presence probability determining step of calculating the speech presence probability vector by inputting the conditional speech presence probability vector and the state posterior probability for all the combinations;
Including a noise suppression method.   A program for causing a computer to function as the noise suppression device according to claim 1.

Images