KR100901367B1 - Speech enhancement method based on minima controlled recursive averaging technique incorporating conditional map - Google Patents

Abstract

A speech enhancement method using a condition post maximum probability base minimum value control recursion average techniques capable of improving the quality of the voice is provided to modify the speech signal probability by granting a condition about the presence of the voice at the previous frame and member. A threshold value is changed according to whether to exist a previous frame(S10). A critical value is compared with a ratio of the minimum value of the local energy in a signal mixing the voice and noise. It determines whether to exist the voice signal in the current frame(S20). According to existence and nonexistence of the determined speech signal, the voice probability of being is presumed(S30). The noises are presumed by using the presumed voice probability(S40). The tone quality of the voice in which the noises is mixed is improved by using the presumed noise.

Description

Condition Enhancement Based Minimum Probability Based Minimum Control Recursive Average Technique Speech Enhancement Method {SPEECH ENHANCEMENT METHOD BASED ON MINIMA CONTROLLED RECURSIVE AVERAGING TECHNIQUE INCORPORATING CONDITIONAL MAP}

The present invention relates to a speech enhancement method, and more particularly, by adding a post-condition maximum probability to a minimum control recursive average technique, i.e., in a manner in which a signal entering a current frame imposes conditions on the presence and absence of speech in a previous frame. The present invention relates to a speech enhancement method capable of significantly improving speech quality by modifying a signal existence probability.

Accurate estimation of noise in a practical speech enhancement system is one of the most important factors, especially the ability to handle abnormal noise signals. Since the estimation of the noise signal has a large influence on the quality of the speech enhancement system, when the estimated noise signal is too small, unnatural residual noise is detected, and when it is too large, the speech signal becomes dull and intelligibility is reduced. Traditional noise estimates are most commonly averaged in the absence of speech, and such estimates that rely on Voice Activity Detection (VAD) are difficult to adjust and are distorted when used in applications with low signal-to-noise ratios. There is a problem that the signal is output. Recent studies are mainly applying the soft decision method to estimate the power spectrum of the noise signal in the speech domain, and the Minimum Statistics Noise Estimation (MSNS) technique that is an alternative to the method of using the speech detector. Has originality without using VAD. Specifically, it is based on optimally smoothing the signal's power spectrum and applying a minimum probability. The power spectrum smoothing algorithm uses a first-order regression system as a smoothing parameter that is time and frequency dependent. The smoothing parameter tracks the abnormal noise signal by optimizing it to minimize the conditional mean square error. However, these noise estimates are sensitive only to outliers and have a variance that is twice as large as traditional methods. Moreover, this method has the disadvantage of making the phone with less energy weaker, especially if the minimum search window is too small. Although a small amount of computation and an efficient minimum tracking method have been proposed, when the energy of the noise signal rises, the noise estimation becomes slow and the signal disappears.

On the other hand, Cohen's recently proposed minima controlled recursive averaging (MCRA) compensates for these shortcomings by averaging the power spectrum using smoothing parameters that are controlled by the signal presence probability in subbands. The presence of the signal in each subband is determined by the ratio between the local energy of the noisy signal and the minimum in a given window. Compare this ratio to a specific threshold, and if the ratio value is less than the threshold, use a method that assumes no speech signal, and averages the time base to reduce variations that occur between and without the speech signal. Take This is because there is a high correlation between frames in the presence of voice signals.

However, there are some problems with this MCRA algorithm. That is, the computational complexity is reduced by using efficient local minimum tracking techniques, but delays occur in the presence of sudden noise, and the presence of the signal in each subband is also the ratio between the local energy of the noisy signal and the minimum in a given window. It is not reliable because it compares only the ratio and the specific threshold. In general, voice activity is strongly correlated with adjacent frames, so the frame immediately preceding or following the frame in which the voice is active is likely to be active, and vice versa. have.

Considering these results, based on the existing statistical assumptions, the noise-overlaid speech is applied by combining the MC with a hangover algorithm using the Hidden Markov Model (HMM) to the maximum condition A posteriori (MAP). There is a need to try to improve the quality.

The present invention is derived from the recognition of the necessity as described above, and instead of the conventional MCRA method of estimating the presence or absence of a speech signal in each subband using a specific threshold value, a condition for the presence or absence of a speech signal in a previous subband is added. To improve the performance of speech presence probability through a more reliable comparison, to derive an excellent noise estimation method, and to propose a speech enhancement method that can estimate the noise spectrum and improve the quality of the speech mixed noise. For that purpose.

According to a feature of the present invention for achieving the above object, a method for improving speech using a post-condition maximum probability based minimum value control recursive average technique,

(1) a first step of changing a threshold in accordance with the presence or absence of a voice signal of a previous frame;

(2) a second step of determining the presence or absence of a voice signal in the current frame by comparing a ratio of the minimum value of the local energy and the local energy in a signal mixed with voice and noise with a threshold obtained in the first step;

(3) a third step of estimating a voice presence probability according to the presence or absence of the voice signal determined in the second step;

(4) a fourth step of estimating noise using the voice presence probability estimated in the third step; And

(5) The fifth aspect of the present invention is characterized by including a fifth step of improving the sound quality of the speech-mixed speech using the noise estimated in the fourth step.

Preferably, in the first step, the threshold value is changed using the following equation according to the presence or absence of the voice signal of the previous frame.

If there is no speech section in the previous frame in the above equation,

,

,

If there is a voice segment in the previous frame,

.

.

In short, the above equation is given by the following equation.

.

.

로서 정의됨.here,

Defined as

According to the speech enhancement method of the present invention, the presence of the speech signal in such a manner that the maximum value of the post-condition condition is added to the minimum value control recursive average technique, that is, the signal entering the current frame gives a condition for the presence or absence of the speech in the previous frame. By modifying the probability, the quality of speech can be greatly improved.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a view showing the configuration of a voice enhancement method according to an embodiment of the present invention. As shown in FIG. 1, the voice enhancement method according to an embodiment of the present invention includes a threshold value updating step S10, a voice signal presence determination step S20, a voice presence probability estimation step S30, A noise estimation step S40, and a sound quality improvement step S50.

First, in the threshold value updating step S10, the threshold value is changed depending on the presence or absence of the audio signal of the previous frame. Specifically, by what principle, which equation is applied to change the threshold will be described in detail later.

Next, in the step S20 of determining the presence or absence of a voice signal in the current frame, the presence or absence of a voice signal in the current frame is determined by comparing the ratio of the minimum value of the local energy and the local energy to the threshold obtained in step S10 in a signal mixed with voice and noise. In the speech presence probability estimation step S30, the speech presence probability is estimated according to the presence or absence of the speech signal determined in step S20. In the noise estimation step S40, the noise is estimated using the speech presence probability estimated in step S30. Estimate.

Finally, in the sound quality improvement step S50, the sound quality of the voice mixed with noise is improved by using the noise estimated in step S40.

Hereinafter, a process for obtaining a time-varying smoothing parameter adjusted to speech presence probability to estimate noise will be described first, and then the ratio of the minimum value of local energy and local energy for obtaining the speech presence probability is determined by the presence or absence of the voice of the previous frame. The process of comparing with the threshold value according to this will be described in detail.

1. Estimation of Noise by Minimum Controlled Recursive Average Technique

Let x (n) and d (n) be an additive noise signal that has no correlation with the audio signal, respectively. Where n represents a discrete time. The observed signal y (n) is given by y (n) = x (n) + d (n), partially divided into overlapping windows, and then using a short-time Fourier transform (STFT). It can be expressed as Equation 1 below.

Y (k, l) = X (k, l) + D (k, l)

Here, Y (k, l) becomes the k-th frequency component in the l-th frame. If the basic hypothesis used in the speech enhancement technique is H ₀ (k, l) and H ₁ (k, l) for each of absence and presence of speech, it can be expressed as Equation 2 below.

H ₀ (k, l): Y (k, l) = D (k, l)

H ₁ (k, l): Y (k, l) = X (k, l) + D (k, l)

Where X (k, l) and D (k, l) represent the Fourier transform coefficients of the original speech signal and the noise signal, respectively. Here, when is the variance of the noise signal in the th subband, it can be expressed as Equation 3 by applying the iterative smoothing of time to the signal observed in the absence of the speech signal to estimate.

Here, α _d (0 <α _d <1) is a smoothing parameter, and H ′ ₀ and H ′ ₁ respectively represent the absence and presence of a speech signal for noise power update based on the hypothesis. The hypothesis in Equation 2 used to estimate the speech signal and Equation 3 used to adjust the spectral update of the noise signal should be distinguished. In other words, determining that there is no voice signal H ₀ when the voice signal is present (H ₁ ) is more dangerous when estimating the voice signal than when estimating the noise signal. Therefore, different decision rule is used, in general, and with a higher confidence in the H ₁ rather than _{1, P (H 1 | Y} ) ≥P (H 'H | is the Y _1).

If p (k, l) = P (H ' ₁ (k, l) | Y (k, l)) represents the conditional probability of the presence of speech, Equation 3 is expressed by Equation 4 below.

는 음성 존재 확률로 조정하는 시변 스무딩 매개변수로서 다음 수학식 5와 같이 주어진다.here,

Is a time-varying smoothing parameter that adjusts to a voice presence probability as given by Equation 5 below.

Equation 6 is used to estimate p (k, l).

Where α _p (0 <α _p <1) is the smoothing parameter and δ is the threshold of speech signal presence.

In particular, the local energy is represented by the following equation (7).

S (k, l) = α _S S (k, l-1) + (1-α _S ) | Y (k, l) | ²

Where α _S (0 <α _S <1) is a parameter.

The minimum value of local energy S _min (k, l) and the temporary variable S _tmp (k, l) are given by S _min (k, 0) = S (k, 0) and S _tmp (k, 0) = S (k, 0 Initialize to). As shown in Equation 8, the minimum value of the current frame is obtained by comparing the local energy with the minimum value of the previous frame.

S _min (k, l) = min {S _min (k, l-1), S (k, l)}

S _tmp (k, l) = min {S _tmp (k, l-1), S (k, l)}

After all L frames have been processed, Equation 9 is initialized.

S _min (k, l) = min {S _tmp (k, l-1), S (k, l)}

S _tmp (k, l) = S (k, l)

By repeating the processes of Equations 8 and 9, the minimum value is searched as shown in Equation 10 below.

이다.here,

to be.

2. Conditional Post-Probability Based Minimum Value Control Recursive Average Technique

The traditional MAP criterion follows a determination method as in Equation 11 below.

Where H _n represents the exact assumption of the nth frame. According to Bayes' law, the criterion of Equation 11 can be changed as Equation 12 below.

In order to compensate for the bias, the equation (12) may be modified as shown in the following equation (13).

Where α ≧ 1.

In the present invention, when estimating the presence of a signal in each subband using only one threshold of the MCRA, two thresholds are defined by this condition by adding a condition as to whether or not the signal of the previous subband contains voice. Value to make the comparison more reliable than ever before. The following is a process of deriving the definition of a new voice signal existence probability by adding a condition of whether or not a signal of a previous subband includes voice.

Equation (13) is replaced with the Bayes' rule and is expressed by the following equation (14).

In general, voice activity is strongly correlated with adjacent frames. In other words, the frame immediately preceding the frame in which the voice is active or the frame immediately following the voice is likely to be active, and vice versa. Based on this correlation, the use of Hover with HMM can effectively reduce the error of VAD based on statistical model. Due to the strong interrelationship of the frames before and after the voice activity, the posterior probability P (H _n | X) is used instead of the posterior probability P (H _n | X). _n | X, H _{n -1} ) can be calculated. After the condition of the previous frame is added to Equation 14, it can be changed into Equation 15 by Bayes' law.

Equation 15 may be changed to Equation 16 below.

Although the speech activity of the current frame depends on the previous frame, the speech activity of the current frame is dominantly influenced by the distribution of DFT coefficients of the noisy speech signal observed in the current frame. Furthermore, the parameters distributed in p (S _r | H _n = H ₁ , H _{n -1} = H ₀ ) and p (S _r | H _n = H ₀ , H _{n -1} = H ₁ ) are obtained from the sample data. Lack of reliable estimates is not possible. For this reason, we can make simple assumptions such as

p (S _r | H _n = H _j , H _{n -1} = H _i ) = p (S _r | H _n = H _j ), i = 0,1 j = 0,1

By using this, Equation 16 may be simplified as in Equation 18 below.

In Equation 18, if there is no speech section in the previous frame, the following Equation 19 is obtained.

If there is a voice section in the previous frame, the following equation (20) is obtained.

This increase in the number of thresholds provides additional degrees of freedom for improving the performance of the VAD. Due to the strong correlation of voice activity between adjacent frames, Equation 21 is generally established.

Equation (21) shows that if the voice activity is observed in the previous frame, the probability that the voice activity is observed in the current frame is high, the probability of not being observed is low, and vice versa. It can be seen that.

In summary, the results can be expressed as in Equation 22 below.

Is given by Equation 23 below.

As described above, according to the present invention, a more accurate noise estimation can be performed by varying the threshold value according to the presence or absence of the voice signal of the previous frame and fluidly estimating the voice presence probability.

FIG. 2 is a diagram illustrating a comparison of voice presence probability through a threshold test based on MAP, for example, in the case of F16 noise (SNR = 10dB). FIG. 2A is a diagram showing a clean voice waveform, and FIG. 2B is a diagram showing voice presence probability in a real time frame. From FIG. 2, the voice enhancement method proposed in the present invention finds that the voice starts at a faster rate than the conventional MCRA method, and determines that the voice is not a voice even though it does not drop sharply at the end of the voice. It can be clearly seen that it reduces the loss of voice information.

3. Experimental Results

In the present invention, in the existing MCRA algorithm, the presence of the signal in each subband is determined as the ratio between the local energy of the noisy signal and the minimum value in a given window, and the ratio is compared with only a specific threshold. By applying the MAP, the threshold is continuously updated to make the speech presence probability more reliable, thereby inducing speech enhancement. ITU-T P.862 PESQ and subjective sound quality evaluation, which are widely applied to evaluate the sound quality of the proposed speech enhancement algorithm, were performed. Table 1 and Table 2 show the extracted PESQ values and the subjective sound quality evaluations, respectively. .

Noise type Method SNR (dB) 5 10 15 White noise MCRA Proposed 1.936 2.050 2.296 2.391 2.641 2.715 Babble noise MCRA Proposed 2.379 2.397 2.697 2.724 2.972 3.004 F16 noise MCRA Proposed 2.056 2.152 2.467 2.548 2.776 2.846

For the PESQ test of Table 1, the sample added four types of noise to data sampled from 10ms to 8kHz in size for a voice that allowed each male and female speaker to pronounce 100 sentences. Noise was measured using white noise, babble noise, and F16 noise from the NOISEX-92 database. SNRs were tested at 5, 10, and 15 dB. The PESQ values were expressed as average values for these samples, and the weighting parameters were set to α _d = 0.95, α _p = 0.2, α _S = 0.45, respectively, for PESQ by the existing MCRA, and the threshold values β ', β''. Was set to be β '= 9 and β''= 3.5.

From Table 1, it can be seen that the proposed method shows a significant performance improvement over 5dB and 10dB of white noise and F16 noise, compared with the conventional method. It can be seen that the proposed algorithm shows an improved speech enhancement by better estimating the presence of speech even at low SNR as shown in FIG. In the case of babble noise, the voice probability of the proposed algorithm is similar to the existing one, but shows a slightly improved result.

Noise type Method SNR (dB) 5 10 15 White noise MCRA Proposed 2.17 2.23 2.56 2.65 3.03 3.07 Babble noise MCRA Proposed 2.93 2.95 3.32 3.42 3.69 3.76 F16 noise MCRA Proposed 2.23 2.33 2.77 2.90 3.20 3.46

In the subjective sound quality evaluation of Table 2, 20 listeners evaluated noise signals with white, babble, and F16 noise added with SNR of 5, 10, and 15 dB, respectively, for male and female speakers to pronounce 10 sentences. . From Table 2, it can be seen that the proposed method shows improved results for all three noise environments and each SNR compared to the conventional method. In conclusion, we can confirm that the proposed MCRA algorithm based on conditional MAP is excellent in various noise environments.

The present invention described above may be variously modified or applied by those skilled in the art, and the scope of the technical idea according to the present invention should be defined by the following claims.

1 is a view showing the configuration of a voice enhancement method according to an embodiment of the present invention.

FIG. 2 is a diagram showing a comparison of probabilities in F16 noise (SNR = 10dB), and FIG. 2A is a diagram of a clean speech waveform, and FIG. 2B is a diagram of speech presence probability in a real time frame.

<Explanation of symbols for main parts of the drawings>

S10: threshold update step

S20: Step of determining whether or not a voice signal of the current frame

S30: speech existence probability estimation step

S40: Noise Estimation Step

S50: Sound Quality Step

Claims (2)

(1) a first step of changing a threshold in accordance with the presence or absence of a voice signal of a previous frame; (2) a second step of determining the presence or absence of a voice signal in the current frame by comparing the ratio of the minimum value of the local energy and the local energy in the signal mixed with the voice and noise with the threshold obtained in the first step; (3) a third step of estimating a voice presence probability according to the presence or absence of the voice signal determined in the second step; (4) a fourth step of estimating noise using the voice presence probability estimated in the third step; And (5) a fifth step of improving the sound quality of the voice-mixed voice using the noise estimated in the fourth step Voice enhancement method comprising a. The method of claim 1, In the first step, according to the presence or absence of the audio signal of the previous frame, the voice enhancement method characterized in that the threshold value is changed using the following equation a. <Formula a>

If there is no speech section in the previous frame in Equation a, it is given as Equation b below. Equation b

, If there is a voice interval in the previous frame, it is given by Equation c below. <Equation c>

. When the equation b and the equation c are briefly expressed as one equation, the following equation d is given. <Equation d>

here,

S _r = S _r (k, l) is defined as S (k, l) / S _min (k, l) as the ratio of the minimum value of local energy and local energy, and S (k, l) And S _min (k, l) represent the local energy and the minimum of the local energy, respectively. In addition, H _n represents the exact assumption of the n-th frame, and H ₀ , H ₁ represent the assumption of the absence of speech and the presence of speech, respectively. In addition, α is a coefficient for compensating for the biased, and the range is given as α ≧ 1.

Images