TW425542B - Kalman filter for speech enhancement - Google Patents

Abstract

The present invention relates to a Kalman filter for speech enhancement that, in a low operation volume, can effectively enhance the speech quality by eliminating the speech noises. The Kalman filter mainly consists of a secondary band analysis filter bank, several secondary band frame parameter estimators, several Kalman filters, and a secondary band synthesis filter bank. The secondary band analysis filter bank can separate the speech signal into several secondary band speeches. The secondary band frame parameter estimators can separate a secondary band speech into several continuous frames and calculate the autoregressive parameter of each frame. The Kalman filters can then use the autoregressive parameter of each frame to enhance the sound quality of the respective secondary band speech. Finally, the secondary band synthesis filter bank can synthesize the enhanced secondary band speeches into an enhanced speech.

Description

425 ^ 42 V. Description of the invention (1) The present invention relates to a Speech Enhancement device ', and in particular to a Carmen wave reducer for speech enhancement. Carmen (Ka 1 ma η) filtering technology is well known in the field of digital signal processing. Proceeding IEEE International Conference On Acoustic, Speech, published in April 1987,

Signa1, Processing, pp. 177 to 1 80 "A Speech Enhancement Method Based On Ka 1 man F i 1 te r | 'ng 丨 丨 The authors of the paper Pai i wal, Basu use Carmen filter to enhance the white noise ( White Noise) contaminated voice signals. Gibson, Koo, and Gray, authors of the article "Filtering Of Colored Noise For Speech Enhancement And Coding" in IEEE Transactions On Signal Processing, page 1.32 to 1741, published in August 1991, aim at The speech signal polluted by colored noise is discussed. In this example, Gibson 'Koo' Gray assumes the speech sH number and colored noise as an Autoregressive Model (AR Model). The Carmen filtering method based on scalar and vector is established. In order to estimate the AR Coefficients, the EM-based tooling method is used. Published in Signal Processing, September 1995, pages 1 to 14 The author of 'An EM-Base Approach For Parameter Enhancement With An Application To Speech Signals1' Lee 'Ann proposed non-Gaussian (Non-Ga ussian) automatic regression model of the speech signal. In this example, Lee and Ann assumed that the distribution of driving noise was a Gaussian Mixture model, and applied decision-making

Page 4

Five 'Invention Note ¢ 2) (Decision-Directed) nonlinear Carmen filter. In addition, IEEE Transactions on Signal Processing, pages 528 to 537, published in March 1996 " Adaptive Scheme For

Elimination Of Broadband Noise And Impulsive Disturbance From AR And ARMA Signals "-The authors of the text, Niedzwiecki and Cisowki, assume that the speech signal is a Nonstat ionary autoregressive model, and assume that the autoregressive coefficient is a Random-Walk model. . And the extended Carmen filter. Wave filter (Extended

Kalman Fi 1 ter) estimates both the speech signal and the autoregressive coefficient. When the above-mentioned Carmen filter is actually applied to enhance the sound quality of a voice, the problem of excessive calculation amount may occur. This is because the speech signal is usually assumed to be an autoregressive model of high order (Order), so a high-powered Carmen filter is required to consider the wave. Estimating the parameters of high-powered autoregressive molding and the application of high-powered Carmen filters require a large amount of calculation. In view of this, the main purpose of the present invention is to provide a kind of Carmen filter that is suitable for strengthening a language to reduce the parameters for estimating high-power autoregressive models (referred to as autoregressive parameters) and the application of high-power Carmen filters A lot of calculations. In order to achieve the above object, the present invention utilizes the nature of the autoregressive model in which the sub-band t audio signal can be set to a lower power, and proposes a Carmen for speech enhancement; a wave filter. This type of Carmen filter is mainly composed of a sub-band analysis filter, an Analysis Filter iank, a plurality of sub-band sound box parameter estimators, a plurality of Carmen filters, and a sub-band wave (㈣heSlsFllterBank). Among them, the secondary frequency Y analysis filter

IH Page 5 425542 Explanation of the Five 'Invention (3) The wave module can separate the voice signal into a plurality of sub-band voices; the sub-band voice frame parameter estimator can divide the sub-band voice signal into consecutive frames, and For each frame, an innovative autocorrelation function subtraction method is used to calculate the autocorrelation function of the sub-frequency speech signal without noise, which is then solved by the conventional Yuler-Walker equation. Regression parameters (s. Hayk in 'Adaptive Filter Theory' Englewood Clitts, NJ: Prentice-Ha 11 1991); the Carmen filter can use the autoregressive parameters of each sound box to enhance the sound quality of the sub-band speech; and the sub-band synthesis filter module is The enhanced sub-band speech can be synthesized into enhanced speech. In this Carmen's wave filter for speech enhancement, the sub-band analysis filter module can be composed of several bandpass filters (Bandpass Fi Iter) and several downsampling devices (Downsampiing). Among them, several band-pass filters can respectively filter the corresponding components of the voice signal to the sub-bands; and several down-sampling samplers can respectively multiply reduce the corresponding components of the sub-bands to obtain the sub-band voices. Signal. In this type of Carmen filter used for speech enhancement, the 'sub-band synthesis filter module' may be composed of several upsampling filters, several bandpass filters, and an adder. Among them, several multiplier samplers can multiply the sampled and strengthened sub-band voice signals to obtain the corresponding components of the sub-bands; several band-pass filters can filter the corresponding components of the sub-bands separately; The output of these band-pass filters for enhanced speech.

IHI I Page 6/4 4 2 t V. Description of the invention (4) Detailed design method of sub-band analysis filter module and sub-band synthesis filter module 'can refer to: pLVaidyanthan' Mutirate System And Filter Bank, Ciitts, N I: Prentice-HalI, 1 993 ° In order to make the above and other objects, features, and advantages of the present invention more comprehensible, an example is given below, and it is described in detail with the accompanying drawings as follows: Figure The first diagram is a context diagram of a Carmen filter for speech enhancement according to the present invention; and the second diagram is a detailed block diagram of the first diagram. Example First, it is explained how a conventional Carmen filter is used to enhance speech disturbed by white or colored noise. For example, based on Short-Time Basis, the sequence of speech signals without noise (Sequence), +) 'can be expressed as an autoregressive model, that is: jf i-1 v ^ where ~ is autoregressive T < «) is a white Gaussian (wh. T) with a mean of zero (Zero-Mean) and a variance (Var iance) of 4

Gaussian) drive signal. And the voice signal office) you can add y ~ J from the assumption of X⑻

_ 4255 4 2 V. Description of the invention (5) Add a white noise V («), that is: s (n) = H ~ v («) (2) where ν («) is zero and the variance is < White Gaussian Noise. Let X (n) ^ [x (n) x (nl) A x (np + \)] r, then the expressions (1) and (2) can be rearranged into state equations, that is: --FX {n- 1) + gw (n) (3) .s («) = hT x (ή) + v (n) ⑷ α 2 a S-1 1 0 A 0 0 F = 0 1 A 0 0 (5) Μ Μ 0 MM 0 0 Λ 1 0 ~ 1 Production? G = k-= [10 Λ 〇L ⑹ According to (3) and (4), Gibson, Koo, and Gray can be obtained by using a Carmen waver 4 «) The best estimate of is enhanced speech. The Carmen filter can be expressed recursively in the following four forms: X (^) = FX (^ ~ V) s (n)-hTF X (n -1) (7) k (n) = P {n \ i-\ ) h [al + / 2r ^ («|«-(8) Ρ (? ί ^ -1) = FP (? i-\) FT + gcF ^ gr (9) P (n) = [l-k { n) hT | p («|« -1) (10)

\ 425542 V. Description of the invention (6) where 'is the estimated value of ⑻, ⑻ is the Carmen gain, and (4 * -1) is the State Prediction-Error Covariance Matrix. () State Filtration Error Covariance Matrix. And enhanced, speech can be obtained. Filtering for colored noise 'First assume that colored noise is Stationary and can be expressed as an autoregressive model with a power of Q, that is: ν («) = ρν (" ϋ «) (n) if The self-returning operation calculates where Τ7 («) is a white Gaussian signal with a mean value of zero and a variance σ [. The reduction coefficients h λ and < can be expressed in a non-speech interval (Non-Speech Interval) estimation formula (11). Is the equation of state. Γ («) = FvV {n -1) + gvi7 («) v {n) = klV (n) and can be assumed to be known (12) (13) where ❿) = [々》- :!) '... ,, Ί, Αϊ are the same as expressions (5) and (6), but f P is replaced by BU. The combined expressions (12) ^' (3), (4) can be get:

5. Explanation of the invention (7) X (n)-FX {n-1) + Gw («) s («) = kr X (η) where (14) (15) X {n) = F =, W («), ⑻.» _ (Γ, σ = g 0 — Fy, 0 Sv. (16) (17) F h

Ton) variance matrix is defined as: & la⑻. According to (1 4) and (1 5), the Carmen's expression can be set by Θ = 〇, and f⑻, corpse, A4, g in Λ (7) to (10) are expressed by 歹 ⑻, 歹, 、, 0 '-Λ r soil σ instead. (Enhanced voice * ⑻ can be obtained from coffee) = 0

Go to D. Next, the method of the Carmen filter for speech enhancement according to the present invention will be described as follows. Please refer to FIG. 1, which is a block diagram of a Carmen filter for speech enhancement according to the present invention. In the figure, such a filter is mainly composed of a sub-band analysis filter module 10, a Carmen filter 20 ^ 2 (^, a sub-band sound frame parameter estimator 25i to 25M, and a band synthesis filter module 30. The sub-band The analysis and filtering module 10 receives the voice signal (c), and separates it into a plurality of (eg, M) sub-band voice signals ♦), where I = 1 ~ M. And M Carmen filters 2 ~ 2 0M

V. Description of the invention (8) Receiving M sub-boats ^. Ν e e this frequency band 1 frequency ▼ θ, and perform a filtering action to strengthen the strongest two-tone signal 'and the enhanced sub-band voice signal 1 _ | The in-band synthesis filter module 30 estimates the parameters of the enhanced sub-band voice frequency sub-frames. All the Carmen wave filters are combined-required parameters. The calculator is 25m ~, which is used to supply the Carmen filter. In this embodiment, any sub-band analysis and filtering by noise is performed blindly by the guilt M group. M sub-band voice signals

Si (n), τ = \, ..., Μ. -a This peach 1 ^-human band audio signal can be expressed as: A («) = 々〇!) + _), j = 1; ^ Μ) ί 二 :) ㈣ and)) subband signals, respectively . If the sound is', then the state can also assume white noise. Conversely, if V⑷ is colored noise, then it will sound) It can also be assumed to be colored noise. In this way, the coffee can also be used as a self-return model. #Based on a short time, X plus can be regarded as frequency 1 of the band voice signal. 'Because the frequency spectrum of the band 0 signal is lower than its full-time turn-back; ^ band voice signals can be used and can be simplified. . The 'Qing Qing / Xia Xia' Carmen waver also made the watchman express the self-winged bismuth 9 of the power attacker, and the type of 'shell 0 («) can be expressed as: (w) = 2 ai Xi (n ~ j) + W; («) J-1 (19) where '% («) means zero and square — white lancet driven signal by l *.

iH IHI page 11 425542 V. Description of the invention ¢ 9) (1 8), (1 9) can be rewritten into state equations, so the Carmen filter 2 ~ 2 0M can be used to strengthen the sub-band voice signal. The enhanced sub-band speech signal, which is expressed as', is then reconstructed by the sub-band synthesis filter module to strengthen the speech signal, which is shown as male). To use the Carmen filter, it is necessary to estimate the autoregressive parameters of the sub-band speech signal ^ ⑻ and noise κ⑻; including the autoregressive coefficient and the drive noise variance. The autoregressive parameters of '⑻ can be estimated from the interval without speech (the interval containing only noise): first estimate the autocorrelation function of ~ ⑻, and then solve the autoregressive coefficient by the Ulle ~ Walk equation.

^ = h (κ), ν; («-1), Λ, vi (nq + l) f JL < = E (V, (n) V, («) Γ) 'E {Vi («+1 ) ^ («)} (20) where M ·) is the expected value operand, then the autoregressive coefficient of ^ ㈦ can be obtained from the solution of the Juler-Walker equation, that is: ^ = (^)-1 ^ (21) The corresponding driving noise variance can be expressed as: < = ^ (〇) -Σάχ〇) (22) J-1 Among them, 4 (〇 = run is “W” autocorrelation function. Value- It should be noted that 砣 and β are composed of the autocorrelation function 4 (0, where r = 0,1, Λ, q- \.

Page 12 425542 V. Description of the Invention (ίο) If it is based on a short time, \ ⑻ can be regarded as a steady state. Therefore, if the sub-band speech signal is cut into continuous frames, each frame can be assumed to be an autoregressive model. However, in fact, we only have noise in the sub-band speech α («) and no sub-band speech 4«) without noise, so the auto-correlation function and even the auto-regressive parameters cannot be obtained by conventional methods. The present invention proposes an autocorrelation function subtraction method to solve this problem. The autocorrelation function subtraction method is described as follows: The autocorrelation function of the subband speech without noise can be estimated by subtracting the autocorrelation function of the subband noise from the autocorrelation function of the subband speech with noise. The autocorrelation function subtraction method can be expressed mathematically as follows. First assume that noise-free sub-band speech & sub-band noise

News ν »is irrelevant. Let 4 (0 and 4⑺ denote the autocorrelation functions of 5» and a⑻, respectively, 1J Ο) = £ {si («+ r) si («)} = («+ r) + V ; («+ 〇] [^ («) + Vj («)]} (23) = + ΐ ·) ^ («)} + + r) Vj («)} = 4⑺ + 4⑺ So, the sub-band voice signal The autocorrelation function can be obtained by subtracting the autocorrelation function, that is: 4 (0 = Ca)-弋 (Γ) (24) where " wOO, and 4 (Γ) represent the autocorrelation functions of Vi⑻, and 5⑻, respectively. To make the estimation of 4 (0 more flexible), (24) can be rewritten as

Page 13 (25) V. Description of the invention (π) 4 (0 = 4 where α is a constant between 0 and 1. Let 40) = [七 〇), 七 〇-1), Λ, ' («+ L) f ", and (26) 4 = («) 4 ⑻7}, Θ = pair ... (《+1) 4 («)) Then 砼 and Θ can be combined by 4 (0. Similar expressions (2 1), the automatic return coefficient of 回 w, meaning = k, aUeUf, which can be obtained from the solution of the Juler-Walker equation, that is: and the corresponding driving noise variance is σίιν = 4 (〇) _Σ «⑺ (27) (28) In this example, matrix inversion is also included in the parameter estimation process. However, if the power of the autoregressive model is low, these actions can be easily achieved. In addition, the The correlation function can also be time averaged to get the estimated value. For example: (29) where # is the size of the sound frame and μ is the number of the specific sound frame. Please refer to Figure 2, which describes the sub-band in more detail than Figure 1. Internal block diagrams of the analysis filter module and the sub-band synthesis filter module. Among them, the sub-band analysis filter module 10 is obtained by MU band-pass filters 1, 2, 12

Page 14 V. Description of the invention (12) Sampler 14 "14M composed of a band-pass filter 12 丨] 2 voice signal coffee), and will receive the input from Richmond and Ancient Eighth Class W / M knife J ', The filtering knife is divided into M sub-band components and ⑷. And | {subsampling samples are respectively multiplied by 2 to reduce the sampling and M to obtain the sub-band voice signal coffee described above). In addition — human band synthesis filter module 3 0 is caused by μ 彳 wailing due to 俾 俾 32 丨 ~ called, M band-pass filter turbid crying "^" In the multiplier sampling benefit, M multiplier increases' sample 3 "Second; ° Pirates 36 composition. Its device 3 2 ～ 3 2 Μ respectively receive the enhanced signal signal from the μ Carmen filter 2 0 ～ 2 0 Μ, and the private signal is drawn to obtain M sub-band components, and these signals are sampled and M Each band-pass filter 34 丨 ~ 34M filters these components separately, so as to obtain "另外. In addition, the adder 36 adds these signals (small), so as to obtain the enhanced voice office). 士 r Γ: η: M · 5 and a bandpass filter with a length of twenty sub-band filter 'wave module as an example' to illustrate the simulation results of the present invention. In this example, the voice signal is uttered by a woman '1 while considering white noise and colored Noise. In order to understand the filtering characteristics, colored noise uses locomotive noise and / speed noise Λ. The noise ratio of the right input is 5db, and the results of enhancement are shown in Table 1. The data in Table 1 are the enhancements of the noise-to-noise ratio. In Table 1, (p, Q) represents the power of the autoregressive model of the sub-band speech signal without noise and the sub-band noise, respectively. To simplify the problem, (P, Q) of all sub-band signals are the same. To show the present invention The effect of the speech, and the analog of the present invention and κ〇〇 Β · j · d. Gibs〇n mentioned full frequency

Carmen EM filter for comparison. Let e =, b κ〇ο and J D

Page 15 425542 V. Description of the invention (13) The performance method proposed by Gib s η first divides the voice signal into plural frames, and then repeats the following two steps for individual frames. Step 1. Use Carmen Filter. Step 2. Use the estimated value of) to calculate less w) where / is the number of repetitions. The results of this simulation are listed in Table 1 and are labeled as EM- / ', where 1,2,3. For the Carmen EM filtering algorithm, the speech is set to the fourth order, and the noise is set to the second order autoregressive model. It can be seen from the table that under the white or locomotive noise, except for the worst performance of EM-1, the results of all methods are similar. And (0, 2) of the present invention is better than EM-2 when applied to all noise 'and (2' 2) can achieve the best effect on automobile noise. Next, the amounts of the present invention (0, 2)) and EM-1 (4) are listed in Table 2. In Table 2, MPU, DVU, and ADU represent the amount of multiplication per unit time and the amount of division per unit time. The amount of increase per unit of time. Table 3 lists the calculated amounts of (0, 2), (2, 2), and ._ in this example. The amount of difference calculated for (2, 2) is only EM _ ^ I / 2, and (〇, 2) is only} / 6 The left is that the calculation amount of EM-2 is twice as much as ⑽— 丨, and the calculation amount of EM_3 is three 卞 1. & The calculation amount of the algorithm is far more than the full-frequency Carmen EM algorithm.

Descend according to the present invention's Carmen waver for speech enhancement can greatly coat the parameters of the auto-regressive model of the intermission # and use the power of the high-powered Carmen II / device to effectively achieve the sound quality of the voice signal strengthen. In addition, if the Noise 0 hole is a Wide Band noise ’, then the (0’0) model is sufficient to improve

425542 V. Description of the invention (14) Provide satisfactory enhancement results. If the noise is a narrow-band Band noise, the higher power (2 '2) model is sufficient to provide satisfactory enhancement results. The card-door filter based on the sub-band sound frame can be used. Processor (Digital Signal Processor, DSP), special C / Algorithm Specific Integrated jrCUl t 'As 1C) or high-level programming languages, such as C, Fortran, Mat lab, etc. The above embodiments have been disclosed as above, but they are not used for the present invention. Any person skilled in the art can make changes and retouching within the scope of the present invention. Therefore, the protection scope of the present invention When < the scope of the patent application is defined. (P * Q) White noise locomotive hybrid pole automobile noise The present invention (0, 0) 5.39 5.81 3.53 The present invention (1, 0) 5.50 5.82 1 II 3.43 The present invention (0, 1) 5.40 5.81 5.7〇 The present invention ( 1,1) 5.49 5.S4 6.98 The present invention (2,0) 5.38 5.64 2.94 The present invention (0,2) 5.40 5.82 7.51 The present invention (2,2) 5.19 5.57 9.05 EM-1 (4 '2) 3.7 3.51 4.97 EM-2 (4, 2) 5.4 5.16 737 ~ EM-3 (4'2) 5.63 5.84 8 ^ 20 ~ Table 1

V. Description of the invention (15) Action EM · 1 (4, 2) The invention (2, 2) The invention (0, 2) MPU DVU ADU MPU DVU ADU MPU DVU ADU Carmen filter 120 6 111 56 4 51 16 2 15 Autocorrelation function calculation 5-5 3 3 1 1 Band filter module--4 4 4 4 Total 127 6 116 63 4 58 21 2 20 Table 2 SB (2, 2) SB (0, 2) Carmen EM -1 (4,2) 1/2 1/6 Carmen EM-2 (4,2) 1/4 1/12 Carmen EM-3 (4,2) 1/6 1/18 Table 3

IMII Page 18

Claims (1)

425542 6. Scope of patent application i. A kind of Carmen filter that is used to strengthen the speech enhancement λ μ \ \ Insectification, including the separation of speech signals into a plurality of-two-person band is an analysis filter module, the sub-band voice; complex number ^ Π ν ^ ^ \ ^ Most estimators' separate the sub-band speech into a plurality of continuous sound frames, respectively. * &Amp; & The multiple Carmen chirpers corresponding to the primary frequency band speech without noise are separately profitable; = the autoregressive parameters corresponding to the speech are used to strengthen the secondary frequency band speech; and the primary frequency band synthesis wave module is used to convert multiple The enhanced sub-band speech is combined into an enhanced speech. 2. The Carmen filter for speech enhancement described in item 1 of the scope of the patent application, where 'the sub-band analysis and filtering module includes: a plurality of band-pass filters', which respectively filter the voice signal into a plurality of sub-bands. Corresponding components; and a plurality of down-samplers' respectively down-sample the corresponding components of the sub-band to obtain the sub-band voice signal. 3. The Carmen wave filter for speech enhancement as described in item 1 of the scope of the patent application, wherein the sub-band synthesis filtering module includes: a plurality of multiplier samplers' multiples to increase sampling of the enhanced sub-band voice signal Obtain the corresponding components of a plurality of sub-bands;. A plurality of band-pass filters' respectively correspond to the corresponding components of the sub-band; and an adder that adds the multiple outputs of the band-pass filter to obtain A fortified voice. Page 19, 4255 4 2 ", patent application scope 4. Carmen filter for speech and tone enhancement as described in item 1 of the patent application scope, wherein the autoregressive parameter of the sub-band speech uses an autocorrelation function phase Subtraction calculates an autocorrelation function corresponding to the sub-frequency speech signal that contains no noise in each frame in the continuous frame, and then uses a Juler-Walker. Equation to find that each frame contains no noise. The autoregressive parameter corresponding to the sub-frequency voice signal of the sub-frequency voice signal. 5, as described in item 4 of the scope of the patent application, the auto-correlation function subtraction method, wherein each sub-frame of the sub-frequency voice signal that does not contain noise corresponds to the sub-frequency voice signal. The autocorrelation function is estimated by subtracting an autocorrelation function corresponding to α times the primary frequency band noise from an autocorrelation function corresponding to a subband speech containing noise, where α is between 0 and Constant between 1. ~ Page 20