CN101030380B - Method for estimating fractional base sound of code excitation linear predicted speech encoder - Google Patents

Abstract

The invention relates to a fractional pitch estimation method of a code-excited linear predictive speech coder. The fractional pitch estimation is obtained by directly estimating the peak position through polynomial interpolation, and the fractional pitch estimation value is obtained through the following steps: (1) first performing open-loop pitch estimation on the target signal, Calculate the open-loop integer pitch estimate value T _op ; (2) then carry out the open-loop pitch estimate, and calculate the closed-loop correlation sequence C (k) in the given neighborhood of the open-loop integer pitch estimate value T _op ; (3) if the closed-loop correlation sequence If the peak point T _cl of C(k) appears at both ends of a given neighborhood, the fractional pitch estimate T _fr is zero; if the peak point T _cl appears in a given neighborhood, then according to the peak point T _cl and its Perform polynomial fitting on the values of T _cl -1 at the previous point and T _cl +1 at the next point to obtain the peak position T _r , and linearly quantize the peak position T _r according to the interpolation factor D to obtain the fractional pitch estimated value T _fr , which is linearly quantized The formula is T _fr =round(T _r *D). On the premise of obtaining accurate fractional pitch estimation value and improving prediction gain, the present invention effectively reduces computational complexity and system overhead.

Description

A Fractional Pitch Estimation Method for Code Excited Linear Predictive Speech Coders

technical field

The invention belongs to the field of voice communication, in particular to a fractional pitch estimation method for a code-excited linear prediction speech coder.

Background technique

Code Excited Linear Prediction (CELP) comprehensively utilizes the short-term prediction characteristics, long-term prediction characteristics and vector quantization method of speech to obtain coding gain, and has been widely used in low-bit-rate speech communication. Pitch period estimation is one of the key factors affecting prediction gain in long-term prediction. In the digitization process of the speech signal, since the pitch period of the speech signal is usually not an integer multiple of the sampling interval T, if the integer pitch prediction is directly used, the pitch period obtained in the discrete domain will be -0.5T~ 0.5T mismatch. If fractional pitch estimation is used, the gain reduction caused by this mismatch can be better resolved.

The existing fractional pitch estimation method is to upsample and low-pass filter the correlation sequence of the target signal (speech signal or its residual signal after weighted linear prediction) near the peak value to obtain an interpolated correlation sequence, and the interpolation multiple depends on the fraction The estimated accuracy required by the pitch, and then calculate its peak value at the fractional point to obtain the fractional pitch estimate, and the final pitch period is synthesized by the integer pitch estimate and the fractional pitch estimate. However, the existing fractional pitch estimation methods have the following defects: the fractional pitch estimation value required for the long-term prediction of the first-order fraction only indicates the position of the peak value of the interpolation sequence, rather than the peak value; on the other hand, the interpolation composed of upsampling and low-pass filtering The filter cannot accommodate variable interpolation factors. Therefore, in the prior art, interpolation and comparison operations need to be solved to realize fractional pitch estimation, the algorithm is complicated and the memory consumption is large, and the efficiency of calculating the fractional pitch estimation value is low. How to improve the fractional pitch estimation method to overcome the above defects has become a technical problem to be solved urgently in this field.

Contents of the invention

The purpose of the present invention is to solve the deficiencies of the prior art, provide a fractional pitch estimation method for a code-excited linear predictive speech encoder, and improve the efficiency of calculating the fractional pitch estimation value under the premise of ensuring the accuracy of the fractional pitch estimation result.

In order to achieve the above object, the fractional pitch estimation method provided by the present invention obtains the fractional pitch estimation value by directly estimating the peak position through polynomial interpolation, and the fractional pitch estimation value is obtained through the following steps,

(1) First, open-loop pitch estimation is performed on the target signal, and an open-loop integer pitch estimation value T _op is calculated;

(2) Then perform open-loop pitch estimation, and calculate the closed-loop correlation sequence C(k) in a given neighborhood of the open-loop integer pitch estimate value T _op , k∈[T _op -l, T _op +l], l is based on The signal frequency is a given integer, and the given field is the allowable closed-loop pitch search range;

(3) If the peak point T _cl of the closed-loop correlation sequence C(k) appears at both ends of a given neighborhood, the fractional pitch estimate T _fr is zero; if the peak point T _cl of the closed-loop correlation sequence C(k) appears in a given neighborhood, polynomial fitting is performed according to the peak point T _cl and the values of its previous point T _cl -1 and subsequent point T _cl +1 to obtain the peak position T _r , and the peak position T _r is calculated by the interpolation factor D performs linear quantization to obtain fractional pitch estimation value T _fr , and the linear quantization formula is T _fr =round(T _r *D).

And, the polynomial fitting described in step (3) adopts the quadratic polynomial fitting method, and its computing formula is

${\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; *</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$

Among them, C(T _cl ) represents the value of T _cl at the peak point in the closed-loop correlation sequence C(k), C(T _cl -1) represents the value of T _cl -1 a point before the peak point, and C(T _cl +1) represents The value of T _cl +1 a little after the peak point.

The present invention creatively proposes to obtain the peak position directly without relying on the interpolation factor, and then obtain the estimated value of the fractional pitch. The new method does not need to save the interpolation filter coefficient table, and the complexity has nothing to do with the interpolation factor. When obtaining the accurate estimated value of the fractional pitch, On the premise of improving the prediction gain, the computational complexity and system overhead are effectively reduced. It can be seen that the method provided by the present invention can realize high-efficiency and high-precision fractional pitch estimation, provide reliable support for the application of code-excited linear predictive speech coders, and have great practical value in the field of low-bit-rate speech communication.

Description of drawings

Fig. 1 is a flowchart of fractional pitch estimation according to an embodiment of the present invention.

Detailed ways

The invention provides a fractional pitch estimation method for a code-excited linear predictive speech coder. After the open-loop and closed-loop pitch estimation, the estimation of the peak position is obtained through a polynomial interpolation method, and then quantized according to the interpolation ratio to obtain a fractional pitch estimation value. The steps to obtain the fractional pitch estimate are: (1) first perform open-loop pitch estimation on the target signal, and calculate the open-loop integer pitch estimate T _op ; (2) then perform open-loop pitch estimation, and calculate the open-loop integer pitch estimate T op Calculate the closed-loop correlation sequence C(k) in the given neighborhood of _op ; (3)

If the peak point T _cl of the closed-loop correlation sequence C(k) appears at both ends of the given neighborhood, the fractional pitch estimate T _fr is zero; if the peak point T _cl of the closed-loop correlation sequence C(k) appears at the given neighborhood In the fixed neighborhood, polynomial fitting is performed according to the peak point T _cl and the values of its previous point T _cl -1 and the latter point T _cl +1 to obtain the peak position T _r , and the peak position T _r is linearized according to the interpolation factor D Quantization to obtain fractional pitch estimated value T _fr , the linear quantization formula is T _fr =round(T _r *D). The method provided by the present invention directly estimates the peak position through polynomial interpolation to obtain the estimated value of the fractional pitch without saving the interpolation filter coefficient table, and the complexity has nothing to do with the interpolation factor D, so the computational complexity and system overhead can be effectively reduced.

Referring to the accompanying drawings, for the convenience of implementation, the fractional pitch estimation process of the embodiment provided by the present invention is as follows:

At first input speech signal S (n), can carry out first-order pre-emphasis filter processing to speech signal S (n) during implementation and obtain pre-emphasis speech signal S _emph (n), and carry out LPC analysis and obtain LPC filter A (z);

Then perform open-loop pitch estimation, pass _Senph (n) through the weighted filter W(z) to obtain the weighted signal S _w (n), and calculate the peak value of the weighted autocorrelation function of S _w (n) to obtain the open-loop An integer pitch estimation value T _op , where T _op represents an open-loop pitch period;

Then perform closed-loop pitch estimation, subtract the weighted synthesis filter W(z)/A(z) zero-input response from S _emph (n), and obtain the target signal x(n) of closed-loop search; and take T _op as the center, calculate The target signal x(n) and the integer delay k∈[T _op -l, T _op +l] (l is a given integer, given according to the signal frequency l belongs to the prior art in the art, and the calculation in this range can It is both accurate and efficient. For example, for a speech signal sampled at 16KHz, l can take a value of 8) to obtain the correlation between the synthesized signal y _k (n), and obtain the closed-loop correlation sequence C (k), so as to find the peak value of C (k) Point Tc _l , T _cl represents the closed-loop pitch period;

Judging whether T _cl appears in a given neighborhood [T _min , T _max ] (T _min and T _max are respectively the integer left and right boundaries of the allowed closed-loop pitch search range, and the values of T _min and T _max according to the signal frequency belong to this field Prior art, for example, to the speech signal of 16KHz sampling, T _min can take value 34, and T _max can take the value 231) on the boundary, if yes then fractional pitch estimated value T _fr is zero value; If not then to peak point T _cl and the values C(T _cl ), C(T _cl -1), and C(T _cl +1) of the previous point T _cl -1 and the latter point T _cl +1 are fitted with quadratic polynomials to obtain the peak position T _r ,

${\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; *</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; cl</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ;</span>$

After obtaining the peak position T _r , perform linear quantization on T _r with a quantization step of 1/D to obtain the fractional pitch estimated value T _fr , the formula is T _fr = round(T _r ^* D), where D is the interpolation factor , represents the precision of fractional pitch estimation, and round() represents rounding.

The pitch period is synthesized by an integer pitch estimate value and a fractional pitch estimate value, so the final pitch estimate value can be obtained as Tc _l +T _fr /D.

Claims (2)

1. the mark pitch estimation method of clep speech coder is characterized in that: obtain mark fundamental tone estimated value by polynomial interpolation direct estimation peak, mark fundamental tone estimated value is obtained by following steps,

(1) at first echo signal is carried out open-loop pitch and estimate, calculate open-loop integer fundamental tone estimated value T _Op

(2) then carry out closed loop pitch and estimate, at open-loop integer fundamental tone estimated value T _OpGiven neighborhood in calculate closed loop correlated series C (k), k ∈ [T _Op-l, T _Op+ l], l is according to the given integer of signal frequency, the closed loop pitch searcher scope of described given field for allowing;

(3) if the peak point T of closed loop correlated series C (k) _ClAppear at the two ends of given neighborhood, then mark fundamental tone estimated value T _FrBe null value; If the peak point T of closed loop correlated series C (k) _ClAppear in the given neighborhood, then according to peak point T among the closed loop correlated series C (k) _ClAnd preceding 1 T _Cl-1, back 1 T _Cl+ 1 respective value is carried out fitting of a polynomial, tries to achieve peak T _r, to peak T _rCarry out equal interval quantizing by interpolation factor D, obtain mark fundamental tone estimated value T _Fr, the equal interval quantizing formula is T _Fr=round (T _r* D), round () expression round off rounds.

2. mark pitch estimation method according to claim 1 is characterized in that: the described fitting of a polynomial of step (3) adopts quadratic polynomial to fit method, and its computing formula is

$T_r=\frac{C(T_{\mathrm{cl}}-1)-C(T_{\mathrm{cl}}+1)}{2*(C(T_{\mathrm{cl}}-1)+C(T_{\mathrm{cl}}+1)-2*C\left(T_{\mathrm{cl}}\right))}$

Wherein, C (T _Cl) the middle peak point T of expression closed loop correlated series C (k) _ClValue, C (T _Cl-1) preceding 1 T of expression peak point _Cl-1 value, C (T _Cl+ 1) 1 T behind the expression peak point _Cl+ 1 value.