KR20150078831A - Method and system forspeech enhancement using non negative matrix factorization and basis matrix update - Google Patents

Abstract

The present invention relates to a method and a system for speech enhancement. More specifically, the method includes the steps of: (1) drawing a pre-enhanced signal of a complex value converted from a sound signal where a noise and a voice are mixed; (2) estimating a signal-to-noise ratio and gaining a MMSE-LSA gain function to draw a second signal; and (3) updating a basis matrix by using the drawn second signal.

Description

TECHNICAL FIELD [0001] The present invention relates to a method and system for enhancing sound using nonlinear matrix factorization and base matrix update,

The present invention relates to an acoustic enhancement method and system, and more particularly, to an acoustic enhancement method and system using factorial matrix factorization and base matrix update.

Generally, the received input signal includes various noise other than a sound that a person is interested to hear. When such a variety of signals are mixed, it is difficult to perceive a sound that a person intends to listen to, which degrades perceptibility and clarity. Therefore, a method of improving acoustic quality of a noise-mixed acoustic signal is being developed. The statistical model based acoustic enhancement technique is based on statistical modeling of the target speech and other noise to be removed from the input signal, and voice activity detection (noise activity tracking), noise power tracking (See Patent Application No. 10-2006-0095820). The improvement of this method proceeds under the assumption that the noise slowly changes with the passage of time (stationary). The statistical model-based acoustic enhancement technique estimates the statistical characteristics of noise and speech from the noise-mixed input speech signal, so that information of the previously-trained data is not required. However, since it is a method that is based on the assumption that the characteristics of the noise change slowly, the sound improvement performance drops sharply in a non-stationary environment in which the ambient noise changes abruptly.

On the other hand, the template-based acoustic enhancement method is a method of using a pattern of pre-trained speech and noise or statistical data information. This is an acoustic improvement method that removes noise from an input signal through previously known speech characteristics and noise information through training. The template-based sound improvement technique has a robust sound enhancement performance even in an environment where the ambient noise is rapidly changed. However, it is necessary to have data information of pre-training voice and noise in advance. If the noise level of the training noise model and the noise There is a problem that the improvement performance is lowered.

Non-negative matrix factorization (NMF), on the other hand, separates the common basis of individual information in a particular information set. If the actual information group is V, and the matrix to be separated is W, H, then V = WH is satisfied. W denotes a base matrix, and H denotes an encoding matrix. V can be restored to the sum of the columns of W, bases. That is, it is useful for extracting data features because it is possible to approximate the entire data by linear combination of the optimal base patterns from a plurality of input data. In the present invention, a method of deriving a sound enhancement performance by solving the disadvantages of each technique by combining a statistical model-based sound enhancement technique and a template-based sound enhancement technique is proposed.

The present invention has been proposed in order to solve the above-mentioned problems of the previously proposed methods. The present invention proposes a method in which a sound signal mixed with noise and speech is converted into a complex number by using a statistical model- (SNR) value and a post-SNR value based on the noise and the noise (value obtained from the first signal) estimated from the first signal based on the non-sound number matrix factorization (NMF) A method and system for improving sound using non-tone number matrix factorization and basis matrix updating, which have a high performance sound enhancement function by obtaining an SNR value of a sound signal and deriving a second signal using an MMSE-LSA gain function For that purpose.

Further, the present invention can maintain the correct noise model as the initial value by updating the base matrix to be used for the factorization of the non-sound number matrix performed in the next time frame using the second signal, And the update rate is automatically calculated and applied according to the change rate of the noise environment so as to prevent the adverse effects such as overfitting due to unnecessary updates from being adversely affected. It is another object to provide a sound improvement method and system using decomposition and basis matrix update.

In addition, since the present invention utilizes the MMSE-LSA gain function, it is possible to obtain a stable performance rather than a conventional gain function using a Weiner filter type, and the size of voice and noise can be estimated separately. Which can further improve the sound improvement effect by using individual powers using a simple smoothing technique rather than estimating the power of noise and speech using the DD technique, It is another object to provide a sound improvement method and system using matrix updating.

According to an aspect of the present invention, there is provided an acoustic enhancement method using factor matrix nonparametric factorization and base matrix update,

(1) deriving a pre-enhanced signal obtained by converting a sound signal mixed with noise and speech into a complex number using a statistical model-based sound improvement technique;

(2) estimating a signal-to-noise ratio (SNR) value using a value obtained from the first signal based on a non-sound number matrix factorization (NMF), and using the estimated SNR value to calculate an MMSE- Deriving a second signal by obtaining a function; And

(3) updating a basis matrix to be used for factorizing the non-noise matrix of the step (2) performed in the next time frame using the second signal derived in the step (2) And is characterized by its constitution.

Preferably, the step (2)

(2-1) obtaining an absolute value of the first signal, and estimating an encoding matrix through a previously pre-trained noise base matrix and a speech base matrix;

(SNR) value and a post-SNR (SNR) value using the encoding matrix estimated in the step (2-1) and the base matrix updated in the step (3) of the previous time frame, Estimating a value; And

(2-3) deriving the second signal using the SNR values estimated in step (2-2) and the MMSE-LSA gain function.

Preferably, the step (2)

A smoothing technique may be used to estimate the SNR value.

Preferably, the step (3)

Both the noise and speech models can be updated in each frame at the same time, and can be individually calculated and updated for each frequency.

Preferably, the step (3)

(3-1) estimating a voice presence probability value (SPP) in a predetermined frequency bin using the second signal; And

(3-2) determining a rate of updating the base matrix using the voice presence probability value.

More preferably, the step (3-2)

A reconstruction error can be used as an indicator and computed using a sigmoid function.

According to an aspect of the present invention, there is provided an acoustic enhancement system using factor matrix non-noise matrix factorization and base matrix update,

A first signal derivation module for deriving a pre-enhanced signal obtained by converting an acoustic signal in which noise and speech are mixed into a complex number using an acoustic improvement technique based on a statistical model;

Noise ratio (SNR) value using a value obtained from the first signal based on a non-sound number matrix factorization (NMF), and calculates an MMSE-LSA gain function using the estimated SNR value A second signal derivation module for deriving a second signal; And

And a base matrix updating module for updating a basis matrix to be used for factoring non-noise factors performed in the next time frame using the second signal.

Advantageously, the base matrix update module comprises:

An SPP estimation module for estimating a voice presence probability value (SPP) in a predetermined frequency bin using the second signal; And

And an update rate determination module that determines a rate of updating the base matrix using the voice presence probability value.

According to the method and system for improving sound using non-sound number matrix factorization and base matrix update proposed in the present invention, a sound signal in which noise and speech are mixed is converted into a complex number using a statistical model- Noise ratio (SNR) value based on the speech and noise (value obtained from the first signal) estimated from the first signal based on the non-sound number matrix factorization (NMF) after deriving the pre-enhanced signal And a post-SNR (Signal-to-Noise Ratio) value, and derives a second signal using the MMSE-LSA gain function.

Further, according to the present invention, a correct noise model can be maintained at a supposed value by updating the base matrix to be used for the factorization of the non-sound number matrix performed in the next time frame using the second signal, and the speech presence probability value (SPP) It is possible to prevent an adverse effect such as overfitting due to unnecessarily large number of updates from being incurred by automatically calculating and applying the update rate according to the change rate of the noise environment.

In addition, according to the present invention, by using the MMSE-LSA gain function, it is possible to derive a stable performance rather than using the gain function of the conventional Weiner filter type, and the magnitude of the voice and the noise can be estimated, By using individual power using a simple smoothing technique rather than using the Decision Directe (DD) technique to estimate noise and speech power, the sound improvement effect can be further enhanced.

Brief Description of the Drawings Fig. 1 is a flowchart illustrating a method for improving sound using non-tone number matrix factorization and base matrix update according to an embodiment of the present invention.
FIG. 2 is a flowchart illustrating a method of improving sound using non-tone number factorization and base matrix update according to another embodiment of the present invention. FIG.
FIG. 3 is a diagram illustrating a flow of an acoustic enhancement method using factor matrix factorization and base matrix update according to an embodiment of the present invention; FIG.
FIG. 4 illustrates a sound improvement system using a factorial matrix factorization and a base matrix update according to an embodiment of the present invention. FIG.
FIG. 5 illustrates a sound improvement system using a factorial matrix factorization and a base matrix update according to another embodiment of the present invention. FIG.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. In the following detailed description of the preferred embodiments of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. The same or similar reference numerals are used throughout the drawings for portions having similar functions and functions.

In addition, in the entire specification, when a part is referred to as being 'connected' to another part, it may be referred to as 'indirectly connected' not only with 'directly connected' . Also, to "include" an element means that it may include other elements, rather than excluding other elements, unless specifically stated otherwise.

FIG. 1 is a flowchart illustrating a method of improving sound using non-tone number matrix factorization and basis matrix updating according to an embodiment of the present invention. Referring to FIG. As shown in FIG. 1, the acoustic enhancement method using the factorial matrix factorization and the base matrix update according to an embodiment of the present invention uses an acoustic enhancement technique based on a statistical model, (SNR) value by using a value obtained from the first signal based on the non-noise matrix factorization (NMF) (step S100), deriving a pre-enhanced signal converted to a complex number value (S200) deriving a second signal by obtaining an MMSE-LSA gain function using an estimated signal-to-noise ratio (SNR) value, and using the second signal derived in step S200, (Step S300) of updating the basis matrix to be used in the factorization of the non-sound number matrix of step S200 performed in step S300.

In step S100, the sound signal mixed with the noise and voice is a pre-enhancement process for primarily improving the sound using the existing statistical model-based sound improvement technique. (t) is transformed into a complex value through a statistical model-based acoustic enhancement technique, and the signal with the thus transformed noise is called a first signal (Y '(t)). The absolute value of the first signal Y '(t) becomes V, and V can be expressed as V = [Ws, Wn] H through the base matrix Wn (t), Ws (t).

In step S200, a signal-to-noise ratio (SNR) value can be estimated using the values (V, H) obtained from the first signal Y '(t) based on the non-sound number matrix factorization NMF. The gain function of the MMSE-LSA type can be obtained based on the estimated SNR values, and a second signal (enhanced speech) can be obtained using the gain function. Nominal Matrix Factorization (NMF) is the separation of common parts of a set of information in a particular set of information. If the actual information group is V, and the matrix to be separated is W, H, then V = WH is satisfied. W denotes a base matrix, and H denotes an encoding matrix. V can be restored to the sum of the columns of W, bases. V is an (n * m) matrix, W is (n * r) m, and H is a matrix of (r * m) size. n is the number of frequency axes, m is the number of time frames, and r is the number of bases. In a noise mixed sound, W is composed of a speech base matrix Ws (M * rs) and a noise base matrix Wn (M * rn). That is, the speech base matrix obtained through training is denoted by Ws and the noise base matrix is denoted by Wn. The speech basis Ws (t) and the noise basis Wn (t) in the t-frame are analyzed from the pre-enhanced signal of the t-th frame. In this manner, steps S100 and S200 are a process for obtaining a noise-removed speech signal from a noise-mixed input acoustic signal.

In step S300, the final enhancement result through step S200 may be used to update the basis matrix to be used for the factorization of the non-note number matrix performed in the next time frame. Specifically, the probability that speech exists in the frame may be determined, and the SPP value may be used to determine a base matrix update rate or to derive an update value. Steps S100 to S200 are performed based on the base matrix Wn, Ws obtained in step S300, so that the correct noise model can be maintained as the initial value.

FIG. 2 is a flowchart illustrating a method for improving sound using non-tone number matrix factorization and base matrix update according to another embodiment of the present invention. Referring to FIG. As shown in FIG. 2, according to another embodiment of the present invention, in step S200, an absolute value of a first signal is obtained, and a pre- Estimating an encoding matrix through a noise base matrix and a speech base matrix at step S210, using a coding matrix estimated at step S210 and a base matrix updated at step S300 of a previous time frame, Estimating a post-SNR value (S220), and deriving a second signal (S230) using the estimated SNR values and the MMSE-LSA gain function in step S220 . That is, H is initially estimated on the basis of previously pre-trained noise and speech base matrices Wn, Ws, and thereafter, based on the noise and speech base matrix Wn, Ws updated through step S300, H (See Equation 1 below). V (t) is constructed by Y '(t) and is used to improve the factorization matrix of the nonnegative matrix in the t-th frame (V (t) = Y' In this case, the encoding matrix Ht is given randomly at every frame. However, at this time, the difference between voice and noise is different. The speech is comparable in value to the value of the previous frame in the frequency domain. Therefore, it is good for performance improvement to give the initial value of Hs (t) as Hs (t-1) which is the value of the previous frame. On the contrary, in case of noise, performance is degraded when given to the previous value. Thus, only Hn (t) sets the frame every frame at random.

S (Speech) and N (Noise) values estimated using the estimated H and the updated noise and voice basis matrix Wn, Ws can be obtained (see Equation 2 below). The restoration formula at time t is shown as V (t) = [Ws, Wn] [Hs; Hn] to perform in real time every frame. In this case, Hs and Hn are (r * 1) matrices and H (t) = [Hs; Hn]. When the estimation process is completed in this frame, the estimation value of voice and noise size is shown as below.

(SNR) value can be obtained using S (Speech) and N (Noise) values (see Equation 3 below). After this process, the actual noise depends on S to some extent. Therefore, there is a performance problem in using it as a result, and a gain function is obtained and used. In the present invention, the MMSE-LSA gain function is used (see Equation 4 below).

ξ is the priori SNR, and γ is the posteriori SNR. Originally, the power used in the preceding signal-to-noise ratio or the post-signal-to-noise ratio is the expectation in a particular time interval. Preferably, a Smoothing technique may be employed to estimate the SNR value. Since the power used for the propor- tional signal-to-noise ratio or the post-signal-to-noise ratio in the original formula is the expected value at a certain time interval, ξ and γ are obtained using a forgetting factor. Ps (m, t) and Pn (m, t) represent the power spectral density for the m-th frequency axis in the t frame. Ts and Tn represent the forgetting rate. As a result, the improved speech spectrum in the t-th frame is obtained according to X = GY '. Y '(m, t) means the Mth element of Y'.

Conventionally, in connection with the NMF-based sound enhancement technique, a stable Weiner filter gain function is used. However, in the present invention, by using the MMSE-LSA algorithm as a gain function, a more effective sound enhancement effect is exhibited have. In addition, conventionally, it has been common to estimate the power of noise and speech using a Decision Directed (DD) technique. However, this technique has a problem in that it does not show high performance in NMF based enhancement. In the present invention, by using the NMF-based enhancement, unlike the statistical basis improvement using the conventional DD technique, the magnitude of the voice and the noise are separately estimated, so that the individual power can be used by using simple smoothing And it is confirmed that it is much more effective for acoustic improvement.

Step S300 further includes a step S310 of estimating a voice existence probability value SPP in a predetermined frequency bin using the second signal S310 and a step S310 of updating the base matrix using the voice existence probability value And determining a speed (S320). That is, in step S300, the NMF basis can be updated. Specifically, the final result of step S200 may be used to determine the probability (SPP) that speech exists in the frame, and the SPP value may be used for base matrix update rate determination and base matrix update (see Equations 5 through 7 below).

The conventional template-based sound improvement technique has a problem in that the improvement performance is remarkably degraded if the training noise basis and the noise in the actual environment are different from each other. In order to solve this problem, a method of updating only one voice or noise model has been developed. However, in an interval where a voice exists, the voice model and the update are performed. However, the NMF update method is used to update the noise to affect the noise. Only updating the noise model with the same technique in the absence of the noise is very limited and its effect is also not significant. In the present invention, both the noise and the speech model are simultaneously updated in every frame, but they can be separately calculated and updated for each frequency, thereby achieving higher performance. SPP (Voice Presence Probability) was used to reduce the influence of noise bases on the noise basis, or conversely, the noise basis in the updating process. SPP (voice presence probability) means the probability of voice activity on a particular frequency axis for voice and noise base control. This can be derived using adaptive interpolation as shown in Equations (6) and (7) above.

는 스무딩된 복원 에러 값을 의미한다. 노잡음만을 위한 복원 에러값을 계산하는 것은 어려우므로 음성 존재 인터벌(speech presence intervals)에서는

는

와 같은 것으로 볼 수 있다. 수학식 7에서 보여지는 잡음 기저 업데이트 비율

는 수학식 10 및 11에 따라 구해질 수 있고,

의 non decreasing function로서 확정될 수 있다.

는 업데이트 비율을 위한 가장 큰 값이고, 수학식 6에서 음성 기저 업데이트 비율

는 상수

로 정정될 수 있다.The update rate decision is a step of determining the rate at which to update the basis matrix for use in the factoring of the non-note number matrix performed in the next time frame. If the environment of the actual noise changes rapidly, the base matrix update must be performed at a high speed. On the contrary, the base matrix update should be performed at a slow rate as the noise environment changes slowly. If the model you use is appropriate, unnecessarily many updates can have an adverse effect, such as overfitting. In the present invention, the updating rate or speed is automatically calculated and applied in consideration of this. Preferably, when determining the rate of updating the base matrix using the presence probability value, a reconstruction error can be used as an indicator and computed using a sigmoid function 8 to 10). Equation (8) is an equation for obtaining a reconstruction error index, and Equation (9) is an equation for smoothing a reconstruction error index. In Equation (9), Te denotes a smoothing constant,

Denotes a smoothing restoration error value. Since it is difficult to calculate the reconstruction error value for noise only, speech presence intervals

The

And so on. The noise base update rate < RTI ID = 0.0 >

Can be obtained according to Equations 10 and 11,

Can be determined as a non-decreasing function of.

Is the largest value for the update rate, and the voice base update rate

Constant

Lt; / RTI >

FIG. 3 is a diagram illustrating a flow of an acoustic enhancement method using a factorial matrix factorization and a base matrix update according to an embodiment of the present invention. As shown in FIG. 3, the acoustic enhancement method using the factor matrix factorization and the base matrix update according to an embodiment of the present invention is characterized in that an acoustic signal (Y (t), noise spectrum) (NMF process), SNF estimation (SNF estimation), and MMSE-LSA gain function are obtained by pre-enhancement and Y '(t) (Second signal) may be estimated. In addition, SPP estimation is performed by estimating the probability that a speech exists in a frame using the final improvement result, and the SPP value p (t) is calculated based on a decision rate of the update rate and a base matrix Wn, and Ws) and providing them to factorization of the non-sound number matrix of the next frame (On-line bases update). As described above, the present invention combines the statistical model-based acoustic enhancement technique and the template-based acoustic enhancement technique to solve the disadvantages of each technique, thereby achieving higher speech enhancement performance than the conventional methods. More specifically, a sound SNR and post-SNR (Equation 3) are obtained based on the estimated voice and noise (Equation 2) using the NMF technique, and the MMSE-LSA gain function (Equation 4) And the update technique can maintain the correct noise model as the initial value.

4 is a diagram illustrating an acoustic enhancement system using factor matrix non-noise matrix factorization and base matrix update according to an embodiment of the present invention. As shown in FIG. 4, the sound improvement system using factor matrix non-noise matrix factorization and base matrix update according to an embodiment of the present invention uses a statistical model-based sound improvement technique (SNR) using a value obtained from the first signal based on a non-noise matrix factorization (NMF), a first signal derivation module (100) for deriving a first signal (pre-enhanced signal) A second signal derivation module 200 for deriving a second signal by obtaining an MMSE-LSA gain function using an estimated signal-to-noise ratio (SNR) value, And a base matrix update module 300 for updating a basis matrix to be used for factoring the non-note number matrix to be performed.

FIG. 5 is a diagram illustrating a sound improvement system using a factorial matrix factorization and a base matrix update according to another embodiment of the present invention. Referring to FIG. 5, the second signal derivation module 200 includes a non-noise matrix factorization operation module 210, an SNR estimation module 220, and an MMSE-LSA gain function operation module 230, And the base matrix update module 300 may include an SPP estimation module 310 and an update rate determination module 320. Since each configuration is similar to that described above with reference to Figs. 1 to 3, detailed description thereof will be omitted.

Hereinafter, the effects of the present invention will be described in more detail through experimental examples, but the scope of the present invention is not limited by the following experimental examples.

Experimental Example  1. Comparison experiment of sound improvement effect

는 0.03,

는 0.2, Ts는 0.5, Tn은 0.9, Te는 0.98로 하였다. 또한, 결과는 ITU-T 권고안 P.862 Perceptual evaluation of speech quality (PESQ) 점수(Perceptual evaluation of speech quality (PESQ): An objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs, Tech. Rep. ITU-T P.862, 2001. 참조)로 비교하였다.
In order to evaluate the sound improvement effect of the present invention, a sound improvement comparative experiment was conducted. As a first-stage SE (statistical model-based sound improvement technique) module, S. Rangachari and PC Loizou, " A noise-estimation algorithm for highly non-stationary environments, " 48, pp. 220-231, 2006. < / RTI > This provides SPP, PSD estimation in each frame as well as an improved spectrum. TIMIT and NOISEX-92 databases (DBs) were selected as the voice and noise materials, respectively, and the sampling rate was 16 kHz. A 512-point fast Fourier transform with 75% overlap was used. Each noise criterion was composed of a 16-second-long noise signal that was not included in the test database. The number of speech and noise is 40 (rn = rs = 40). With respect to the base update and smoothing related parameter values,

0.03,

0.2, Ts was 0.5, Tn was 0.9, and Te was 0.98. The results are also presented in the ITU-T Recommendation P.862 Perceptual evaluation of speech quality (PESQ) score: An objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs, Tech. Rep. ITU-T P.862, 2001).

The performance of the five methods was compared by the above method. The five methods are NMF-based acoustic enhancement method using only the statistical model-based acoustic enhancement method (SE), NMF-based enhancement method (NMF), Cabras et al. (Pro < SEP > w / o < / RTI > update) proposed by the present invention, except for updating the base matrix, (Proposed).

SE & Cabras updates the noise baseline only in the presence of speech, while conversely, the voice basis updates in the presence of acoustic signals. G. Cabras, S. Canazza, PL Montessoro and R. Rinaldo, "Restoration of audio documents with low SNR : a NMF parameter estimation and perceptually motivated Bayesian suppression rule, in Proc. Sound and Music Computing Conference, pp. 314-321, 2010. < / RTI >

The experiment was performed under three different conditions. The three different conditions are: stationary noise environment with matched noise bases (Table 1), stationary noise environment with stationary noise environment with mismatched noise bases, 2), non-stationary and stationary noise environments, and non-stationary and stationary noise environments with mismatched noise bases (Table 3). If the noise base is matched, it means that the base derived from the noise signal is the same as that used for the test database. For example, the factory floor 2 (factory2) noise base is trained only in the factory2 noise database. Conversely, if the noise base does not match, the noise base is created by a database that is different from the actual noise of the test database. Glass used white noise in this experiment.

Table 1 shows the experimental results when the stationary noise environment coincides with the noise base. Table 2 shows the experimental results when the stationary noise environment does not coincide with the noise base. Table 3 shows the results of the non-stationary and non- This is a table showing experimental results when the stationary noise environment does not coincide with the noise base.

As shown in Table 1, it can be confirmed that the sound improvement method according to the present invention is more effective than the other techniques. Also, as shown in Table 2, when the noise bases do not coincide with each other, the performance of the NMF-based technique deteriorates. However, it can be confirmed that the effect of the base matrix update increases greatly. Also, as shown in Table 3, it can be confirmed that the method according to the present invention is very effective also in the method according to the present invention when the characteristic of the noise changes slowly or rapidly. As described above, the present invention, which is a combination of a statistical model-based sound improvement technique and a negative matrix factorization and base matrix update method, is excellent in sound improvement effect even in training data that does not match actual noise and rapidly changing noise environment, (factory2, F-16, M109, Lepard), it shows remarkable sound improvement effect compared to other methods.

The present invention may be embodied in many other specific forms without departing from the spirit or essential characteristics of the invention.

100: first signal derivation module 200: second signal derivation module
210: non-sound number matrix factorization operation module 220: SNR estimation module
230: MMSE-LSA gain function calculation module 300: Base matrix update module
310: SPP estimation module 320: update rate determination module
S100: deriving a pre-enhanced signal obtained by converting a sound signal mixed with noise and speech into a complex number using a statistical model-based sound improvement technique
S200: estimates the signal-to-noise ratio (SNR) value using the value obtained from the first signal based on the nonnegative matrix factorization (NMF), and calculates the MMSE-LSA gain function using the estimated SNR Thereby deriving a second signal,
S210: Estimating an absolute value of the first signal, and estimating an encoding matrix through a previously pre-trained noise base matrix and a speech base matrix
S220: Estimating the SNR value and the SNR value using the encoding matrix estimated in step S210 and the base matrix updated in step S300 of the previous time frame
S230: deriving a second signal using the SNR values estimated in step S220 and the MMSE-LSA gain function
S300: Updating a basis matrix to be used for factoring the non-note number matrix of step S200 performed in the next time frame using the second signal derived in step S200
S310: estimating a speech presence probability value (SPP) in a predetermined frequency bin using the second signal
S320: determining a rate of updating the base matrix using the probability of voice presence value

Claims (8)

As an acoustic improvement method using non-sound number matrix factorization,
(1) deriving a pre-enhanced signal obtained by converting a sound signal mixed with noise and speech into a complex number using a statistical model-based sound improvement technique;
(2) estimating a signal-to-noise ratio (SNR) value using a value obtained from the first signal based on a non-sound number matrix factorization (NMF), and using the estimated SNR value to calculate an MMSE- Deriving a second signal by obtaining a function; And
(3) updating a basis matrix to be used for factorizing the non-noise matrix of the step (2) performed in the next time frame using the second signal derived in the step (2) A method for acoustic enhancement using factor matrix factorization and base matrix update.
2. The method of claim 1, wherein step (2)
(2-1) obtaining an absolute value of the first signal, and estimating an encoding matrix through a previously pre-trained noise base matrix and a speech base matrix;
(SNR) value and a post-SNR (SNR) value using the encoding matrix estimated in the step (2-1) and the base matrix updated in the step (3) of the previous time frame, Estimating a value; And
(2-3) deriving a second signal using the SNR values estimated in step (2-2) and the MMSE-LSA gain function. A method of improving sound using a base matrix update.
2. The method of claim 1, wherein step (2)
And performing a smoothing technique to estimate the signal-to-noise ratio (SNR) value using the non-noise matrix factorization and the base matrix update.
2. The method of claim 1, wherein step (3)
Wherein the noise and speech models are simultaneously updated in each frame, and the noise and speech models are individually calculated and updated for each frequency, thereby improving the sound using the non-sound number matrix factorization and the base matrix update.
2. The method of claim 1, wherein step (3)
(3-1) estimating a voice presence probability value (SPP) in a predetermined frequency bin using the second signal; And
(3-2) determining a rate at which the base matrix is updated using the speech presence probability value. ≪ Desc / Clms Page number 19 >
6. The method according to claim 5, wherein the step (3-2)
Wherein a reconstruction error is used as an indicator and computed using a sigmoid function. ≪ Desc / Clms Page number 20 >
As an acoustic improvement system using the factorization of non-sound number matrix,
A first signal derivation module for deriving a pre-enhanced signal obtained by converting an acoustic signal in which noise and speech are mixed into a complex number using an acoustic improvement technique based on a statistical model;
(SNR) value using a value obtained from the first signal based on a non-sound number matrix factorization (NMF), and calculates an MMSE-LSA gain function using the estimated SNR A second signal derivation module for deriving a second signal; And
And a base matrix update module for updating the basis matrix for use in the factorization of the non-note number matrix performed in the next time frame using the second signal. ≪ RTI ID = 0.0 > Sound Enhancement System Using.
8. The apparatus of claim 7, wherein the base matrix update module comprises:
An SPP estimation module for estimating a voice presence probability value (SPP) in a predetermined frequency bin using the second signal; And
And an update rate determination module that determines a rate of updating the base matrix using the speech presence probability value. ≪ Desc / Clms Page number 19 >

Images