Method and apparatus for CELP coding an audio signal while distinguishing speech periods and non-speech periods

BACKGROUND OF THE INVENTION

Field of the Invention

0001The present invention relates to a CELP (Code Excited Linear Prediction) coder and, more particularly, to a CELP coder giving consideration to the influence of an audio signal in non-speech signal periods.

Description of the Background Art

0002It has been customary with coding and decoding of speeches to deal with speech periods and non-speech periods equivalently. Non-speech periods will often be referred to as noise periods hereinafter simply because noises are conspicuous, compared to speech periods. A speech decoding method is disclosed in, e.g., Gerson and Jasiuk "VECTOR SUM EXCITED LINEAR PREDICTION (VSELP) SPEECH CODING AT 8 kbps", Proc. IEEE ICASSP, 1990, pp. 461-464. This document pertains to a VSELP system which is the standard North American digital cellular speech coding system. Japanese digital cellular speech coding systems also adopt a system similar to the VSELP system.

0003However, a CELP coder has the following problem because it attaches importance to a speech period coding characteristic. When a noise is coded by the speech period coding characteristic of the CELP coder and then decoded, the resulting synthetic sound sounds unnatural and annoying. Specifically, codebooks used as excitation sources are optimized for speeches. In addition, a spectrum estimation error derived from LPC (Linear Prediction Coding) analysis differs from one frame to another frame. For these reasons, the noise periods of synthetic sound coded by the CELP coder and then decoded are much removed from the original noises, deteriorating communication quality.

SUMMARY OF THE INVENTION

0004It is therefore an object of the present invention to provide a method and a device for CELP coding an audio signal and capable of reducing the influence of an audio signal (noises including one ascribable to revolution and one ascribable to vibration) on a coded output, thereby enhancing desirable speech reproduction.

0005In accordance with the present invention, a method of CELP coding an input audio signal begins with the step of classifying the input acoustic signal into a speech period and a noise period frame by frame. A new autocorrelation matrix is computed based on the combination of an autocorrelation matrix of a current noise period frame and an autocorrelation matrix of a previous noise period of frame. LPC analysis is performed with the new autocorrelation matrix. A synthesis filter coefficient is determined based on the result of the LPC analysis, quantized, and then sent. An optimal codebook vector is searched for based on the quantized synthetic filter coefficient.

0006Also, in accordance with the present invention, a method of CELP coding an input audio signal begins with the step of determining whether the input audio signal is a speech or a noise subframe by subframe. An autocorrelation matrix of a noise period is computed. LPC analysis is performed with the autocorrelation matrix. A synthesis filter coefficient is determined based on the result of the LPC analysis, quantized, and then sent. An amount of noise reduction and a noise reducing method are selected on the basis of the speech/noise decision. A target signal vector is computed by the noise reducing method selected. An optimal codebook vector is searched for by use of the target signal vector.

0007Further, in accordance with the present invention, an apparatus for CELP coding an input audio signal has an autocorrelation analyzing section for producing autocorrelation information from the input audio signal. A vocal tract prediction coefficient analyzing section computes a vocal tract prediction coefficient from the result of analysis output from the autocorrelation analyzing section. A prediction gain coefficient analyzing section computes a prediction gain coefficient from the vocal tract prediction coefficient. An autocorrelation adjusting section detects a non-speech signal period on the basis of the input audio signal, vocal tract prediction coefficient and prediction gain coefficient, and adjusts the autocorrelation information in the non-speech signal period. A vocal tract prediction coefficient correcting section produces from the adjusted autocorrelation information a corrected vocal tract prediction coefficient having the corrected vocal tract prediction coefficient of the non-speed signal period. A coding section CELP codes the input audio signal by using the corrected vocal tract prediction coefficient and an adaptive excitation signal.

0008Furthermore, in accordance with the present invention, an apparatus for CELP coding an input audio signal has an autocorrelation analyzing section for producing autocorrelation information from the input audio signal. A vocal tract prediction coefficient analyzing section computes a vocal tract prediction coefficient from the result of analysis output from the autocorrelation analyzing section. A prediction gain coefficient analyzing section computes a prediction gain coefficient from the vocal tract prediction coefficient. An LSP (Linear Spectrum Pair) coefficient adjusting section computes an LSP coefficient from the vocal tract prediction coefficient, detects a non-speech signal period of the input audio signal from the input audio signal, vocal tract prediction coefficient and prediction gain coefficient, and adjusts the LSP coefficient of the non-speech signal period. A vocal tract prediction coefficient correcting section produces from the adjusted LSP coefficient a corrected vocal tract prediction coefficient having the corrected vocal tract prediction coefficient of the non-speed signal period. A coding section CELP codes the input audio signal by using the corrected vocal tract coefficient and an adaptive excitation signal.

0009Moreover, in accordance with the present invention, an apparatus for CELP coding an input audio signal has an autocorrelation analyzing section for producing autocorrelation information from the input audio signal. A vocal tract prediction coefficient analyzing section computes a vocal tract prediction coefficient from the result of analysis output from the autocorrelation analyzing section. A prediction gain coefficient analyzing section computes a prediction gain coefficient from the vocal tract prediction coefficient. A vocal tract coefficient adjusting section detects a non-speech signal period on the basis of the input audio signal, vocal tract prediction coefficient and prediction gain coefficient, and adjusts the vocal tract prediction coefficient to thereby output an adjusted vocal tract prediction coefficient. A coding section CELP codes the input audio signal by using the adjusted vocal tract prediction coefficient and an adaptive excitation signal.

0010In addition, in accordance with the present invention, an apparatus for CELP coding an input audio signal has an autocorrelation analyzing section for producing autocorrelation information from the input audio signal. A vocal tract prediction coefficient analyzing section computes a vocal tract prediction coefficient from the result of analysis output from the autocorrelation analyzing section. A prediction gain coefficient analyzing section computes a prediction gain coefficient from the vocal tract prediction coefficient. A noise cancelling section detects a non-speech signal period on the basis of bandpass signals produced by bandpass filtering the input audio signal and the prediction gain coefficient, performs signal analysis on the non-speech signal period to thereby generate a filter coefficient for noise cancellation, and performs noise cancellation with the input audio signal by using said filter coefficient to thereby generate a target signal for the generation of a synthetic speech signal. A synthetic speech generating section generates the synthetic speech signal by using the vocal tract prediction coefficient. A coding section CELP codes the input audio signal by using the vocal tract prediction coefficient and target signal.

BRIEF DESCRIPTION OF THE DRAWINGS

0011The objects and features of the present invention will become more apparent from the consideration of the following detailed description taken in conjunction with the accompanying drawings in which: <ul id="ul0001" list-style="none" compact="compact"><li>FIGS. 1 and 2 are schematic block diagrams showing, when combined, a CELP coder embodying the present invention;</li><li>FIG. 3 is a block diagram schematically showing an alternative embodiment of the present invention, particularly a part thereof alternative to the circuitry of FIG. 2;</li><li>FIG. 4 is a block diagram schematically showing another alternative embodiment of the present invention, particularly a part thereof alternative to the circuitry of FIG. 2; and</li><li>FIG. 5 is a block diagram schematically showing a further alternative embodiment of the present invention, particularly a part thereof alternative to the circuitry of FIG. 2.</li></ul>

DESCRIPTION OF THE PREFERRED EMBODIMENTS

0012Preferred embodiments of the method and apparatus for the CELP coding of an audio signal in accordance with the present invention will be described hereinafter. Briefly, in accordance with the present invention, whether an input signal is a speech or a noise is determined frame by by frame. Then, a synthesis filter coefficient is adjusted on the basis of the result of decision and by use of an autocorrelation matrix, an LSP (Linear Spectrum Pair) coefficient or a direct prediction coefficient, thereby reducing unnatural sounds during noise or unvoiced periods as distinguished from speech or voiced periods. Alternatively, in accordance with the present invention, whether an input signal is a speech or a noise is determined on a subframe-by-subframe basis. Then, a target signal for the selection of an optimal codevector is filtered on the basis of the result of decision, thereby reducing noises.

0013Referring to FIGS. 1 and 2, a CELP coder embodying the present invention is shown. This embodiment is implemented as an CELP speech coder of the type reducing unnatural sounds during noise or unvoiced periods. Briefly, the embodiment classifies input signals into speeches and noises frame by frame, calculates a new autocorrelation matrix based on the combination of the autocorrelation matrix of the current noise frame and that of the previous noise frame, performs LPC analysis with the new matrix, determines a synthesis filter coefficient, quantizes it, and sends the quantized coefficient to a decoder. This allows a decoder to search for an optimal codebook vector using the synthesis filter coefficient.

0014As shown in FIGS. 1 and 2, the CELP coder directed toward the reduction of unnatural sounds receives a digital speech signal or speech vector signal S in the form of a frame on its input terminal 100. The coder transforms the speech signal S to a CELP code and sends the CELP code as coded data via its output terminal 150. Particularly, this embodiment is characterized in that a vocal tract coefficient produced by an autocorrelation matrix computation 102, a speech/noise decision 110, an autocorrelation matrix adjustment 111 and an LPC analyzer 103 is corrected. A conventional CELP coder has coded noise periods, as distinguished from speech or voiced periods, and eventually reproduced annoying sounds. With the above correction of the vocal tract coefficient, the embodiment is free from such a problem.

0015Specifically, the digital speech signal or speech vector signal S arrived at the input port 100 is fed to a frame power computation 101. In response, the frame power computation 101 computes power frame by frame and delivers it to a multiplexer 130 as a frame power signal P. The frame-by-frame input signal S is also applied to the autocorrelation matrix computation 102. This computation 102 computes, based on the signal S, an autocorrelation matrix R for determining a vocal tract coefficient and feeds it to the LPC analyzer 103 and autocorrelation matrix adjustment 111.

0016The LPC analyzer 103 produces a vocal tract prediction coefficient <i>a</i> from the autocorrelation matrix R and delivers it to a prediction gain computation 112. Also, on receiving an autocorrelation matrix Ra from the adjustment 111, the LPC analyzer 103 corrects the vocal tract prediction coefficient <i>a</i> with the matrix Ra, thereby outputting an optimal vocal tract prediction coefficient <i>aa</i>. The optimal prediction coefficient <i>aa</i> is fed to a synthesis filter 104 and an LSP quantizer 109.

0017The prediction gain computation 112 transforms the vocal tract prediction coefficient <i>a</i> to a reflection coefficient, produces a prediction gain from the reflection coefficient, and feeds the prediction gain to the speech/noise decision 110 as a prediction gain signal <i>pg</i>. A pitch coefficient signal <i>ptch</i> is also applied to the speech/noise decision 110 from an adaptive codebook 105 which will be described later. The decision 110 determines whether the current frame signal S is a speech signal or a noise signal on the basis of the signal S, vocal tract prediction coefficient <i>a</i>, and prediction gain signal <i>pg</i>. The decision 110 delivers the result of decision, i.e., a speech/noise decision signal <i>v</i> to the autocorrelation matrix adjustment 111.

0018The autocorrelation matrix adjustment 111, among the others, is the essential feature of the illustrative embodiment and implements processing to be executed only when the input signal S is determined to be a noise signal. On receiving the decision signal <i>v</i> and vocal tract prediction coefficient <i>a</i>, the adjustment 111 determines a new autocorrelation matrix Ra based on the combination of the autocorrelation matrix of the current noise frame and that of the past frame determined to be a noise. The autocorrelation matrix Ra is fed to the LPC analyzer 103.

0019The adaptive codebook 105 stores data representative of a plurality of periodic adaptive excitation vectors beforehand. A particular index number Ip is assigned to each of the adaptive excitation vectors. When an optical index number Ip is fed from a weighting distance computation 108, which will be described, to the codebook 105, the codebook 105 delivers an adaptive excitation vector signal <i>ea</i> designated by the index number Ip to a multiplier 113. At the same time, the codebook 105 delivers the previously mentioned pitch signal <i>ptch</i> to the speech/noise decision 110. The pitch signal <i>ptch</i> is representative of a normalized autocorrelation between the input signal S and the optimal adaptive excitation vector signal <i>ea</i>. The vector data stored in the codebook 105 are updated by an optimal excitation vector signal exOP derived from the excitation vector signal <i>ex</i> output from an adder 115.

0020The illustrative embodiment includes a noise codebook 106 storing data representative of a plurality of noise excitation vectors beforehand. A particular index number Is is assigned to each of the noise excitation vector data. The noise codebook 106 produces a noise excitation vector signal <i>es</i> designated by an optimal index number Is output from the weighting distance computation 108. The vector signal <i>es</i> is fed from the codebook 106 to a multiplier 114.

0021The embodiment further includes a gain codebook 107 storing gain codes respectively corresponding to the adaptive excitation vectors and noise excitation vectors beforehand. A particular index Ig is assigned to each of the gain codes. When an optimal index number Ig is fed from the weighting distance computation 108 to the codebook 107, the codebook 107 outputs a gain code signal <i>ga</i> for an adaptive excitation vector signal or feeds a gain code signal <i>gs</i> for a noise excitation vector signal. The gain code signals <i>ga</i> and <i>gs</i> are fed to the multipliers 113 and 114, respectively.

0022The multiplier 113 multiplies the adaptive oscillation vector signal <i>ea</i> and gain code signal <i>ga</i> received from the adaptive codebook 105 and gain codebook 107, respectively. The resulting product, i.e., an adaptive oscillation vector signal with an optimal magnitude is fed to the adder 115. Likewise, the multiplier 114 multiplies the noise excitation vector signal <i>es</i> and gain code signal <i>gs</i> received from the noise code book 106 and gain codebook 107, respectively. The resulting product, i.e., a noise excitation vector signal with an optimal magnitude is also fed to the adder 115. The adder 115 adds the two vector signals and feeds the resulting oscillation vector signal <i>ex</i> to the synthesis filter 104. At the same time, the adder 115 feeds back the previously mentioned optimal excitation vector signal exOP to the adaptive codebook 105, thereby updating the codebook 105. The above vector signal exOP makes a square sum to be computed by the weighting distance computation 108 minimum.

0023The synthesis filter 104 is implemented by an IIR (Infinite Impulse Response) digital filter by way of example. The filter 104 generates a synthetic speech vector signal (synthetic speech signal) Sw from the corrected optimal vocal tract prediction coefficient <i>aa</i> and excitation vector (excitation signal) <i>ex</i> received from the LPC analyzer 103 and adder 115, respectively. The synthetic speech vector signal Sw is fed to one input (-) of a subtracter 116. Stated another way, the IIR digital filter 104 filters the excitation vector signal <i>ex</i> to output the synthetic speech vector signal Sw, using the corrected optimal vocal tract prediction coefficient <i>aa</i> as a filter (tap) coefficient. Applied to the other input (+) of the subtracter 116 is the input digital speech signal S via the input port 100. The subtracter 116 performs subtraction with the synthetic speech vector signal Sw and audio signal S and delivers the resulting difference to the weighting distance computation 108 as an error vector signal <i>e</i>.

0024The weighting distance computation 108 weights the error vector signal <i>e</i> by frequency conversion and then produces the square sum of the weighted vector signal. Subsequently, the computation 108 determines optimal index numbers Ip, Is and Ig respectively corresponding to the optimal adaptive excitation vector signal, noise excitation vector signal and gain code signal and capable of minimizing a vector signal E derived from the above square sum. The optimal index numbers Ip, Is and Ig are fed to the adaptive codebook 105, noise codebook 106, and gain codebook 107, respectively.

0025The two outputs <i>ga</i> and <i>gs</i> of the gain codebook 107 are connected to the quantizer 117. The quantizer 117 quantizes the gain code <i>ga</i> or <i>gs</i> to output a gain code quantized signal gain and feeds it to the multiplexer 130. The illustrative embodiment has another quantizer 109. The quantizer 109 LSP-quantizes the vocal tract prediction coefficient <i>aa</i> optimally corrected by the noise cancelling procedure, thereby feeding a vocal tract prediction coefficient quantized signal 〈aa〉 to the multiplexer 130.

0026The multiplexer 130 multiplexes the frame power signal P, gain code quantized signal gain, vocal tract prediction coefficient quantized signal 〈aa〉, index Ip for adaptive excitation vector selection, index number Ig for gain code selection, and index number Is for noise excitation vector selection. The multiplexer 130 sends the mutiplexed data via the output 150 as coded data output from the CELP coder.

0027In operation, the frame power computation 101 determines on a frame-by-frame basis the power of the digital speech signal arrived at the input terminal 100, while delivering the frame power signal P to the multiplexer 130. At the same time, the autocorrelation matrix computation 102 computes the autocorrelation matrix R of the input signal S and delivers it to the autocorrelation matrix adjustment 111. Further, the speech/noise decision 110 determines whether the input signal S is a speech signal or a noise signal, using the pitch signal <i>ptch</i>, voice tract prediction coefficient <i>a</i>, and prediction gain signal <i>pg</i>.

0028The LPC analyzer 103 determines the vocal tract prediction coefficient <i>a</i> on the basis of the autocorrelation matrix R received from the autocorrelation matrix computation 102. The prediction gain computation 112 produces the prediction gain signal <i>pg</i> from the prediction coefficient <i>a</i>. These signals <i>a</i> and <i>pg</i> are applied to the speech/noise decision 110. The decision 110 determines, based on the pitch signal <i>ptch</i> received from the adaptive codebook 105, vocal tract prediction coefficient <i>a</i>, prediction gain signal <i>pg</i> and input speech signal S, whether the signal S is a speech or a noise. The decision 110 feeds the resulting speech/noise signal <i>v</i> to the autocorrelation matrix adjustment 111.

0029On receiving the autocorrelation matrix R, vocal tract prediction coefficient <i>a</i> and speech/noise decision signal <i>v</i>, The autocorrelation matrix adjustment 111 produces a new autocorrelation matrix Ra based on the combination of the autocorrelation matrix of the current frame and that of the past frame determined to be a noise. As a result, the autocorrelation matrix of a noise portion which has conventionally been the cause of an annoying sound is optimally corrected.

0030The new autocorrelation matrix Ra is applied to the LPC analyzer 103. In response, the analyzer 103 produces a new optimal vocal tract prediction coefficient <i>aa</i> and feeds it to the synthesis filter 104 as a filter coefficient for an IIR digital filter. The synthesis filter 104 filters the excitation vector signal <i>ex</i> by use of the optimal prediction coefficient <i>aa</i>, thereby outputting a synthetic speech vector signal Sw.

0031The subtracter 116 produces a difference between the input audio signal S and the synthetic speech vector signal Sw and delivers it to the weighting distance computation 108 as an error vector signal <i>e</i>. In response, the computation 108 converts the frequency of the error vector signal <i>e</i> and then weights it to thereby produce optimal index numbers Ia, Is and Ig respectively corresponding to an optimal adaptive excitation vector signal, noise excitation vector signal and gain code signal which will minimize the square sum vector signal E. The optimal index numbers Ip, Is and Ig are fed to the multiplexer 130. At the same time, the index numbers Ip, Is and Ig are applied to the adaptive codebook 105, noise codebook 106 and gain codebook 107 in order to obtain optimal excitation vectors <i>ea</i> and <i>es</i> and an optimal gain code signal <i>ga</i> or <i>gs</i>.

0032The multiplier 113 multiplies the adaptive excitation vector signal <i>ea</i> designated by the index number Ip and read out of the adaptive codebook 105 by the gain code signal <i>ga</i> designated by the Index number Ig and read out of the gain codebook 107. The output signal of the multiplier 113 is fed to the adder 115. On the other hand, the multiplier 114 multiplies the noise excitation vector signal <i>es</i> read out of the noise codebook 106 in response to the index number Is by the gain code <i>gs</i> read out of the gain codebook 107 in response to the index number Ig. The output signal of the multiplier 114 is also fed to the adder 115. The adder 115 adds the two input signals and applies the resulting sum or excitation vector signal <i>ex</i> to the synthesis filter 104. As a result, the synthesis filter outputs a synthetic speech vector signal Sw.

0033As stated above, the synthetic speech vector signal Sw is repeatedly generated by use of the adaptive codebook 105, noise codebook and gain codebook 107 until the difference between the signal Sw and the input speech signal decreases to zero. For periods other than speech or voiced periods, the vocal tract prediction coefficient <i>aa</i> is optimally corrected to produce the synthetic speech vector signal Sw.

0034The multiplexer 130 multiplexes the frame power signal P, gain code quantized signal gain, vocal tract prediction coefficient quantized signal 〈aa〉, index number Ip for adaptive excitation vector selector, index number Ig for gain code selection and index number Is for noise excitation vector selection every moment, thereby outputting coded data.

0035The speech/noise decision 110 will be described in detail. The decision 110 detects noise or unvoiced periods, using a frame pattern and parameters for analysis. First, the decision 110 transforms the parameters for analysis to reflection coefficients r[i] where i = 1, ..., Np which is the degree of the filter. With a stable filter, we have the condition, -1.0 &lt; r[i] &lt;1.0. By using the reflection coefficients r[i], a prediction gain RS may be expressed as:<maths id="math0001" num="Eq. (1)"><math display="block"><mrow><msup><mrow><mtext>RS = Π(1.0 - r[i]</mtext></mrow><mrow><mtext>2</mtext></mrow></msup><mtext>)</mtext></mrow></math><img file="EP0762386A2_D0001.tif" /></maths> where i = 1, ..., Np.

0036The reflection coefficient r[0] is representative of the inclination of the spectrum of an analysis frame signal; as the absolute value |r0| approaches zero, the spectrum becomes more flat. Usually, a noise spectrum is less inclined than a speech spectrum. Further, the prediction gain RS is close to zero in speech or voiced periods while it is close to 1.0 in noise or unvoiced periods. In addition, in a handy phone or similar apparatus using the CELP coder, the frame power is great in voiced periods, but small in unvoiced periods, because the user's mouth or speech source and a microphone or signal input section are close to each other. It follows that a speech and a noise can be distinguished by use of the following equation:<maths id="math0002" num="Eq. (2)"><math display="block"><mrow><mtext>D = Pow · [r[0]| / Rs</mtext></mrow></math><img file="EP0762386A2_D0002.tif" /></maths> A frame will be determined to be a speech if D is greater than Dth or determined to be a noise if D smaller than Dth.

0037The autocorrelation matrix adjustment 111 will be described in detail. The adjustment 111 corrects the autocorrelation matrix R when the past <i>m</i> consecutive frames were continuously determined to be noise. Assume that the current frame and the frame occurred <i>n</i> frames before the current frame have matrices R[0] and R[n], respectively. Then, the noise period has an adjusted autocorrelation matrix Radj given by:<maths id="math0003" num="Eq. (3)"><math display="block"><mrow><mtext>Radj = Σ(Wi · R[i])</mtext></mrow></math><img file="EP0762386A2_D0003.tif" /></maths> where i = 0 through m-1, ΣWi = 1.0, and Wi ≥ W_i+1 > 0.

0038The adjustment 111 computes the autocorrelation matrix Radj with the above Eq. (3) and delivers it to the LPC analyzer 103.

0039The illustrative embodiment having the above configuration has the following advantages. Assume that an input signal other than a speech signal is coded by a CELP coder. Then, the result of analysis differs from the actual signal due to the influence of frame-by-frame vocal tract analysis (spectrum analysis). Moreover, because the degree of difference between the result of analysis and the actual signal varies every frame, a coded signal and a decoded signal each has a spectrum different from that of the original speech and is annoying. By contrast, in the illustrative embodiment, an autocorrelation matrix for spectrum estimation is combined with the autocorrelation matrix of the past noise frame. This successfully reduces the degree of difference between frames as to the result of analysis and thereby obviates annoying synthetic sounds. In addition, because a person is more sensitive to varying noises than to constant noises due to the inherent orditory sense, perceptual quality of a noise period can be improved.

0040Referring to FIG. 3, an alternative embodiment of the present invention will be described. FIG. 3 shows only a part of the embodiment which is alternative to the embodiment of FIG. 2. The alternative part is enclosed by a dashed line A in FIG. 3. Briefly, in the embodiment to be described, the synthesis filter coefficient of a noise period is transformed to an LSP coefficient in order to determine the spectrum characteristic of the synthesis filter 104. The determined spectrum characteristic is compared with the spectrum characteristic of the past noise period in order to compute a new LSP coefficient having reduced spectrum fluctuation. The new LSC coefficient is transformed to a synthesis filter coefficient, quantized, and then sent to a decoder. Such a procedure also allows the decoder to search for an optimal codebook vector, using the synthesis filter coefficient.

0041As shown in FIG. 3, the characteristic part A of the alternative embodiment has an LPC analyzer 103A, a speech/noise decision 110A, a vocal tract coefficient/LSP converter 119, an LSP/vocal tract coefficient converter 120 and an LSP coefficient adjustment 121 in addition to the autocorrelation matrix computation 102 and prediction gain computation 112. The circuitry shown in FIG. 3, like the circuitry shown in FIG. 2, is combined with the circuitry shown in FIG. 1. Hereinafter will be described how the embodiment corrects a vocal tract coefficient to obviate annoying sounds ascribable to the conventional CELP coding of the noise periods as distinguished from speech periods, concentrating on the unique circuitry A. In FIG. 3, the same circuit elements as the elements shown in FIG. 2 are designated by the same reference numerals.

0042The vocal tract coefficient/LSP converter 119 transforms a vocal tract prediction coefficient <i>a</i> to an LSP coefficient <i>l</i> and feeds it to the LSP coefficient adjustment 121. In response, the adjustment 121 adjusts the LSP coefficient <i>l</i> on the basis of a speech/noise decision signal v received from the speech/noise decision 110 and the coefficient <i>1</i>, thereby reducing the influence of noise. An adjusted LSP coefficient <i>la</i> output from the adjustment 121 is applied to the LSP/vocal tract coefficient converter 120. This converter 120 transforms the adjusted LSP coefficient <i>la</i> to an optimal vocal tract prediction coefficient <i>aa</i> and feeds the coefficient <i>aa</i> to the synthesis filter 104 as a digital filter coefficient.

0043The LSP coefficient adjustment 121 will be described in detail. The adjustment 121 adjusts the LSP coefficient only when the past <i>m</i> consecutive frames were determined to be noises. Assume that the current frame has an LSP coefficient LSP-0[i], that the frame occurred <i>n</i> frames before the current frame has a noise period LSP coefficient LSP-n[i], and that the adjusted LSP coefficient is i = 1, ..., Np where Np is the degree of the filter. Then, there holds an equation:<maths id="math0004" num="Eq. (4)"><math display="block"><mrow><msub><mrow><mtext>LSP</mtext></mrow><mrow><mtext>adj</mtext></mrow></msub><msub><mrow><mtext>[i] = ΣW</mtext></mrow><mrow><mtext>k</mtext></mrow></msub><mtext> · LSP-k[i]</mtext></mrow></math><img file="EP0762386A2_D0004.tif" /></maths> where k = 0 through m - 1, ΣW_k = 1.0, i = 0 through Np - 1, and Wk ≥ Wk+1 ≥ 0.

0044LSP coefficients belong to the cosine domain. The adjustment 121 produces an LSP coefficient <i>la</i> with the above equation Eq. (4) and feeds it to the LSP/vocal tract coefficient converter 120.

0045The operation of this embodiment up to the step of computing the optical vocal tract prediction coefficient <i>aa</i> will be described because the subsequent procedure is the same as in the previous embodiment. First, the autocorrelation matrix computation 102 computes an autocorrelation matrix R based on the input digital speech signal S. On receiving the autocorrelation matrix R, the LPC analyzer 103A produces a vocal tract prediction coefficient <i>a</i> and feeds it to the prediction gain computation 112, vocal tract coefficient/LSP converter 119, and speech/noise decision 110.

0046In response, the prediction gain computation 112 computes a prediction gain signal <i>pg</i> and delivers it to the speech/noise decision 110. The vocal tract coefficient/LSP converter 119 computes an LSP coefficient <i>1</i> from the vocal tract prediction coefficient <i>a</i> and applies it to the LSP coefficient adjustment 121. The speech/noise decision 110 outputs a speech/noise decision signal <i>v</i> based on the input vocal tract prediction coefficient <i>a</i>, speech vector signal S, pitch signal <i>ptch</i>, and prediction gain signal <i>pg</i>. The decision signal <i>v</i> is also applied to the LSP coefficient adjustment 121. The adjustment 121 adjusts the LSP coefficient <i>l</i> in order to reduce the influence of noise with the previously mentioned scheme. An adjusted LSP coefficient <i>la</i> output from the adjustment 121 is fed to the LSP/vocal tract coefficient converter 120. In response, the converter 120 transforms the LSP coefficient <i>1a</i> to an optimal vocal tract prediction coefficient <i>aa</i> and feeds it to the synthesis filter 104.

0047As stated above, the illustrative embodiment achieves the same advantages as the previous embodiment by adjusting the LSP coefficient directly relating to the spectrum. In addition, this embodiment reduces computation requirements because it does not have to perform LPC analysis twice.

0048Referring to FIG. 4, another alternative embodiment of the present invention will be described. FIG. 4 shows only a part of the embodiment which is alternative to the embodiment of FIG. 2. The alternative part is enclosed by a dashed line B in FIG. 4. Briefly, in the embodiment to be described, the noise period synthesis filter coefficient is interpolated with the past noise period synthesis filter coefficient in order to directly compute the new synthesis filter coefficient of the current noise period. The new coefficient is quantized and then sent to a decoder, so that the decoder can search for an optimal codebook vector with the new coefficient.

0049As shown in FIG. 4, the characteristic part B of this embodiment has an LPC analyzer 103A and a vocal tract coefficient adjustment 126 in addition to the autocorrelation matrix computation 102, speech/noise decision 110, and prediction gain computation 112. The circuitry shown in FIG. 3 is also combined with the circuitry shown in FIG. 1. The vocal tract coefficient adjustment 126 adjusts, based on the vocal tract prediction coefficient <i>a</i> received from the analyzer 103A and the speech/noise decision signal <i>v</i> received from the decision 110, the coefficient <i>a</i> in such a manner as to reduce the influence of noise. An optical vocal tract prediction coefficient <i>aa</i> output from the adjustment 126 is fed to the synthesis filter 104. In this manner, the adjustment 126 determines a new prediction coefficient <i>aa</i> directly by combining the prediction coefficient <i>a</i> of the current period and that of the past noise period.

0050Specifically, the adjustment 126 performs the above adjustment only when the past <i>m</i> consecutive frames were determined to be noises. Assume that the synthesis filter coefficient of the current frame is 1-0[i], and that the synthetic filter coefficient of the frame occurred <i>n</i> frames before the current frame is a-n[i]. If i = 1, ..., Np where Np is the degree of the filter, then the adjusted filter coefficient is produced by:<maths id="math0005" num="Eq. (5)"><math display="block"><mrow><msub><mrow><mtext>a</mtext></mrow><mrow><mtext>adj</mtext></mrow></msub><msub><mrow><mtext>[i] = ΣW</mtext></mrow><mrow><mtext>k</mtext></mrow></msub><mtext> · (a - k)[i]</mtext></mrow></math><img file="EP0762386A2_D0005.tif" /></maths> where ΣW_k = 1.0, W_k ≥ W_k+1 ≥ 0, <i>k</i> = 0 through m-1, and <i>i</i> = 0 through N_p-1. At this instant, it is necessary to confirm the stability of the filter used the adjusted coefficient. Preferably, the filter determined to be unstable should be so controlled as not to execute the adjustment.

0051The operation of this embodiment up to the step of computing the optimal vocal tract prediction coefficient <i>aa</i> will be described because the subsequent procedure is also the same as in the previous embodiment. First, the autocorrelation matrix computation 102 computes an autocorrelation matrix R based on the input digital speech signal S. On receiving the autocorrelation matrix R, the LPC analyzer 103A produces a vocal tract prediction coefficient <i>a</i> and feeds it to the prediction gain computation 112, vocal tract coefficient adjustment 126, and speech/noise decision 110. The speech/noise decision 110 determines, based on the digital audio signal S, prediction gain coefficient <i>pg</i>, vocal tract prediction coefficient <i>a</i> and pitch signal <i>ptch</i>, whether the signal S is representative of a speech period or a noise period. A speech/noise decision signal <i>v</i> output from the decision 110 is fed to the vocal tract coefficient adjustment 126. The adjustment 126 outputs, based on the decision signal <i>v</i> and prediction coefficient <i>a</i>, an optimal vocal tract prediction coefficient <i>aa</i> so adjusted as to reduce the influence of noise. The optimal coefficient <i>aa</i> is delivered to the synthesis filter 104.

0052As stated above, the this embodiment also achieves the same advantages as the previous embodiment by combining the vocal tract coefficient of the current period with that of the past noise period. In addition, this embodiment reduces computation requirements because it can directly calculate the filter coefficient.

0053A further alternative embodiment of the present invention will be described with reference to FIG. 5. FIG. 5 also shows only a part of the embodiment which is alternative to the embodiment of FIG. 2. The alternative part is enclosed by a dashed line C in FIG. 5. This embodiment is directed toward the cancellation of noise. Briefly, in the embodiment to be described, whether the current period is a speech period or a noise period is determined subframe by subframe. A quantity of noise cancellation and a method for noise cancellation are selected in accordance with the result of the above decision. The noise cancelling method selected is used to compute a target signal vector. Hence, this embodiment allows a decoder to search for an optimal codebook vector with the target signal vector.

0054As shown in FIG. 5, the unique part C of the speech coder has a speech/noise decision 110B, a noise cancelling filter 122, a filter bank 124 and a filter controller 125 as well as the prediction gain computation 112. The filter bank 124 consists of bandpass filters <i>a</i> through <i>n</i> each having a particular passband. The bandpass filter <i>a</i> outputs a passband signal SDbp1 in response to the input digital speech signal S. Likewise, the bandpass filter <i>n</i> outputs a passband signal SbpN in response to the speech signal S. This is also true with the other bandpass filters except for the output passband signal. The bandpass signals Sbp1 through SbpN are input to the speech/noise decision 110B. With the filter bank 124, it is possible to reduce noise in the blocking frequency band and to thereby output a passband signal with an enhanced signal-to-noise ratio. Therefore, the decision 110B can make a decision for every passband easily.

0055The prediction gain computation 112 determines a prediction gain coefficient <i>pg</i> based on the vocal tract prediction coefficient <i>a</i> received from the LPC analyzer 103A. The coefficient <i>pg</i> is applied to the speech/noise decision 110B. The decision 110B computes a noise estimation function for every passband on the basis of the passband signals Sbp1-SbpN output from the filter bank 124, pitch signal <i>ptch</i>, and prediction gain coefficient <i>pg</i>, thereby outputting speech/noise decision signals v1-vN. The passband-by-passband decision signals v1-vN are applied to the filter controller 125.

0056The filter controller 125 adjusts a noise cancelling filter coefficient on the basis of the decision signals v1-vN each showing whether the current period is a voiced or speech period or an unvoiced or noise period. Then, the filter controller 125 feeds an adjusted noise filter coefficient <i>nc</i> to the noise cancelling filter 122 implemented as an IIR or FIR (Finite Impulse Response) digital filter. In response, the filter 122 sets the filter coefficient <i>nc</i> therein and then filters the input speech signal S optimally. As a result, a target signal <i>t</i> with a minimum of noise is output from the filter 122 and fed to the subtracter 116.

0057The operation of this embodiment up to the step of producing the target signal <i>t</i> will be described because the optimal excitation vector signal <i>ex</i> is generated in the same manner as in FIG. 2. First, the autocorrelation matrix computation 102 computes an autocorrelation matrix R in response to the input speech signal S. The autocorrelation matrix R is fed to the LPC analyzer 103A. In response, the LPC analyzer 103A produces a vocal tract prediction coefficient <i>a</i> and delivers it to the prediction gain computation 112 and synthesis filter 104. The computation 112 computes a prediction gain coefficient <i>pg</i> corresponding to the input prediction coefficient <i>a</i> and feeds it to the speech/noise decision 110B.

0058On the other hand, the bandpass filters <i>a</i>-<i>n</i> constituting the filter bank 124 respectively output bandpass signals Sbp1-SbpN in response to the speech signal S. These filter outputs Sbp1-SbpN and the pitch signal <i>ptch</i> and prediction gain coefficient <i>pg</i> are applied to the speech/noise decision 110B. In response, the decision 110B outputs speech/noise decision signals v1-vN on a band-by-band basis. The filter controller 125 adjusts the noise cancelling filter coefficient based on the decision signals v1-vN and delivers an adjusted filter coefficient <i>nc</i> to the noise cancelling filter 122. The filter 122 filters the speech signal S optimally with the filter coefficient <i>nc</i> and thereby outputs a target signal <i>t</i>. The subtracter 116 produces a difference <i>e</i> between the target signal <i>t</i> and the synthetic speech signal Sw output from the synthesis filter 104. The difference is fed to the weighting distance computation 108 as the previously mentioned error signal <i>e</i>. This allows the computation 108 to search for an optimal index based on the error signal <i>e</i>.

0059With the above configuration, the embodiment reduces noise in noise periods, compared to the conventional speech coder, and thereby obviates coded signals which would turn out annoying sounds.

0060As stated above, the illustrative embodiment reduces the degree of unpleasantness in the auditory sense, compared to the case wherein only background noises are heard in speech periods. The embodiment distinguishes a speech period and a noise period during coding and adopts a particular noise cancelling method for each of the two different periods. Therefore, it is possible to enhance sound quality without resorting to complicated processing in speech periods. Further, effecting noise cancellation only with the target signal, the embodiment can reduce noise subframe by subframe. This not only reduces the influence of speech/noise decision errors on speeches, but also reduces the influence of spectrum distortions ascribable to noise cancellation.

0061In summary, it will be seen that the present invention provides provides a method and an apparatus capable of adjusting the correlation information of an audio signal appearing in a non-speech signal period, thereby reducing the influence of such an audio signal. Further, the present invention reduces spectrum fluctuation in a non-speech signal period at an LSP coefficient stage, thereby further reducing the influence of the above undesirable audio signal. Moreover, the present invention adjusts a vocal tract prediction coefficient of a non-speech signal period directly on the basis of a speech prediction coefficient. This reduces the influence of the undesirable audio signal on a coded output while reducing computation requirements to a significant degree. In addition, the present invention frees the coded output in a non-speech signal period from the influence of noise because it can generate a target signal from which noise has been removed.

0062While the present invention has been described with reference to the particular illustrative embodiments, it is not to be restricted by the embodiments. It is to be appreciated that those skilled in the art can change or modify the embodiments without departing from the scope and spirit of the present invention. For example, a pulse codebook may be added to any of the embodiments in order to generate a synthesis speech vector by using a pulse excitation vector as a waveform codevector. While the synthesis filter 104 shown in FIG. 2 is implemented as an IIR digital filter, it may alternatively be implemented as an FIR digital filter or a combined IIR and FIR digital filter.

0063A statistical codebook may be further added to any of the embodiments. For a specific format and method of generating a statistical codebook, a reference may be made to Japanese patent laid-open publication No. 130995/1994 entitled "Statistical Codebook and Method of Generating the Same" and assigned to the same assignee as the present application. Also, while the embodiments have concentrated on a CELP coder, the present invention is similarly practicable with a decoder disclosed in, e.g., Japanese patent laid-open publication No. 165497/1993 entitled "Code Excited Linear Prediction Coder" and assigned to the same assignee as the present application. In addition, the present invention is applicable not only to a CELP coder but also to a VS (Vector Sum) CELP coder, LD (Low Delay) CELP coder, CS (Conjugate Structure) CELP coder, or PSI CELP coder.

0064While the CELP coder of any of the embodiment is advantageously applicable to, e.g., a handy phone, it is also effectively applicable to, e.g., a TDMA (Time Division Multiple Access) transmitter or receiver disclosed in Japanese patent laid-open publication No. 130998/1994 entitled "Compressed Speech Decoder" and assigned to the same assignee as the present application. In addition, the present invention may advantageously be practiced with a VSELP TDMA transmitter.

0065While the noise cancelling filter 122 shown in FIG. 5 is implemented as an IIR, FIR or combined IIR and FIR digital filter, it may alternatively be implemented as a Kalman filter so long as statistical signal and noise quantities are available. With a Kalman filter, the coder is capable of operating optimally even when statistical signal and noise quantities are given in a time varying manner.