Apparatus for expanding speech bandwidth

0001The present invention relates to an apparatus for producing wideband speech signals from narrowband speech signals and in particularly relates to an apparatus for producing wideband speech from telephone-band speech.

0002Among prior methods of expanding speech bandwidth, there are the method described in Y. Yoshida, T. Abe, et. al. "Recovery of wideband speech from narrowband speech by codebook mapping", Denshi Joho Tsushin Gakkai Shingakuho SP 93-61 (1993-08) (in Japanese language) and the method described in Y. Cheng, D. O'Shaughnessy, P. Mermelstein, "Statistical recovery of wideband speech from narrowband speech", Proceed. ICSLP 92 (1992), pp. 1577-1580.

0003According to the method by Yoshida et. al. a great number of code words, for instance 512 codes, have been necessary for faithfully expanding speech bandwidth, since the method relies on codebook mapping. On the other hand, the method of Cheng et. al. had a problem in quality of the synthesized speech, since white noise, which is not correlated to the original speech, is added.

0004The article "An algorithm to reconstruct wideband speech from narrowband speech based on codebook mapping", by Yoshida et al, ICSLP 1994, pp 1591 - 1594, also discloses the use of codebook mapping in generating wideband speech from narrowband speech in association with Linear Predictive Coding (LPC). This article also discloses filtering the LPC synthesised wideband speech and summing together an "up-sampled" version of the input narrowband speech with the filtered (and power modified) synthesised wideband speech signal to generate an output wideband speech signal.

0005An object of the present invention is therefore to produce a wideband speech signal from a narrowband speech signal using a small number of codes.

0006Another object of the present invention is to produce a wideband speech signal from a telephone-band speech signal.

0007A further object of the present invention is to produce a clear wideband speech signal from a narrowband speech signal.

0008In order to achieve the aforementioned objects, the present invention obtains a wideband speech signal from a narrowband speech signal by adding thereto a signal of a frequency range outside the bandwidth of the narrowband speech signal. More particularly, present invention consists in a bandwidth expansion apparatus for recovering wideband speech from narrowband speech comprising: <ul id="ul0001" list-style="none" compact="compact"><li>a bandwidth expansion means for extracting feature amounts from a narrowband input digital speech signal and generating a wideband digital speech signal based on said feature amounts, the bandwidth expansion means including</li><li>a linear predictive coding (LPC) analyzer for performing an LPC analysis on said narrowband input digital speech signal to obtain spectral envelope parameters and a residual signal,</li><li>a spectral envelope converter for converting said spectral envelope parameters into those of wideband,</li><li>a residual converter for converting said residual signal into that of wideband, and</li><li>an LPC synthesizer for synthesizing an output from said spectral envelope converter and an output from said. residual converter to output a wideband digital speech signal;</li><li>a filter means for extracting frequency components of said wideband digital speech signal output from said bandwidth expansion means not contained in the bandwidth of said narrowband input digital signal; and</li><li>a signal adder means for adding said narrowband input digital speech signal and an output signal of said filter means and outputting a synthesized wideband digital speech signal.</li></ul>

0009By means of the above composition, the present invention expands the bandwidth of a speech signal without altering the information contained in the narrowband speech signal. Further, the present invention can produce a synthesized signal having a great correlation with the narrowband speech signal. Still further, the present invention can freely vary the precision of the system by clarifying the process of expanding the bandwidth.

0010These and other objects and features of the present invention will become clear from the following description taken in conjunction with the preferred embodiments thereof with reference to the accompanying drawings throughout which like parts are designated by like reference numerals, and in which: <ul id="ul0002" list-style="none" compact="compact"><li>Fig. 1 is a block diagram illustrating the apparatus for expanding speech bandwidth of an embodiment in accordance with the present invention;</li><li>Fig. 2 is a block diagram illustrating a spectral envelope converter shown in Fig. 1;</li><li>Fig. 3 is a block diagram illustrating another spectral envelope converter of the embodiment in accordance with the present invention;</li><li>Fig. 4 is a block diagram illustrating another spectral envelope converter of the embodiment in accordance with the present invention;</li><li>Fig. 5 is a block diagram illustrating another spectral envelope converter of the embodiment in accordance with the present invention;</li><li>Fig. 6 is a block diagram illustrating the residual converter shown in Fig. 1;</li><li>Fig. 7 is a block diagram illustrating the apparatus for expanding speech bandwidth of another embodiment in accordance with the present invention;</li><li>Fig. 8 is a schematic drawing illustrating the waveform smoother shown in Fig. 1;</li><li>Figs. 9 and 10 illustrate a graph of the number of subspaces and mean distance between original word speeches and word speeches synthesized according to the present invention, in which Fig. 9 shows the results obtained by male speeches and Fig. 10 shows those obtained by female speeches; and</li><li>Fig. 11 illustrates results of a subjective test for evaluating the present invention</li></ul>

0011The preferred embodiments according to the present invention will be described below with reference to the attached drawings.

0012Fig. 1 is a block diagram illustrating the apparatus for expanding speech bandwidth of an embodiment in accordance with the present invention. In Fig. 1, 101 is an A-D converter that converts an original narrowband speech analog signal input thereto to a digital speech signal. The output of A-D converter 101 is fed to a signal adder 103 and an addition signal generator 102. Addition signal generator 102 extracts features from the output signal of A-D converter 101 to output a signal having frequency characteristics of a bandwidth wider than the bandwidth of the input signal. Signal adder 103 algebraically adds the output of A-D converter 101 and the output of addition signal generator 102 to output the resulting signal. A D-A converter 104 converts the digital signal output from signal adder 103 into an analog signal to output. The present embodiment generates an output signal of a bandwidth wider than that of the original signal by this composition.

0013Next, the composition of addition signal generator 102 is described in the following. A bandwidth expander 106 reads the output signal of A-D converter 101 to generate a signal of a bandwidth wider than that of the read signal. Addition signal generator 102 comprises bandwidth expander 106 and filter section 105. The output signal of bandwidth expander 106 is fed to a filter section 105. Filter section 105 extracts frequency components outside the bandwidth of the original signal. For example, if the original signal has frequency components of 300 Hz to 3,400 Hz, then the bandwidth of the components extracted by filter section 105 is the band below 300 Hz and the band above 3,400 Hz.

0014However, it is not necessary to extract all components outside the bandwidth of the original signal. Filter section 105 is preferably configured with a digital filter, which may be either an FIR filter or an IIR filter. FIR and IIR filters are well known and can be realized, for example, by the compositions described in Simon Haykin, "Instruction to adaptive filters", (MacMillan).

0015Next, the composition and operation of bandwidth expander 106 are described in the following. In bandwidth expander 106, LPC (Linear Predictive Coding) analyzer 107 first reads the output signal of A-D converter 101 to perform linear predictive coding (LPC) analysis. LPC analysis is well known and can be realized, for example, by the methods described in Lawrence. R. Rabiner, "Digital processing of speech signals", (Prentice-Hall). LPC analyzer 107 obtains LPC coefficients, which are also called linear predictive codings. The number P of LPC coefficients, i.e. dimension P of feature vector extracted by LPC analyzer is chosen in relation to the sampling frequency and is selected at ten or sixteen since the sampling frequency is 16kHz in the speech analysis. LPC analyzer 107 then obtains other sets of feature amounts from LPC coefficients by transformations. These feature amounts are reflection coefficients, PARCOR (partial correlation) coefficients, Cepstrum coefficients, LSP (line spectrum pair) coefficients and other, and they are all spectral envelope parameters obtained by LPC coefficients. Further, LPC analyzer 107 obtains a residual signal from the LPC coefficients. The residual signal is the difference between the output signal of A-D converter 101 and the predicted signal output from an FIR filter having filter coefficients given by the LPC coefficients. That is, if the output signal of A-D converter 101 is denoted by r(t_n) wherein tn denotes a present sampling time and t_n-1 (i = 1, 2, ..., p) denotes a sampling time i times before, and LPC coefficients are denoted by a_i, i = 1, 2, ..., p, then the residual signal r(t_n) is<maths id="math0001" num="(1)"><math display="block"><mrow><msub><mrow><mtext>r(t</mtext></mrow><mrow><mtext>n</mtext></mrow></msub><msub><mrow><mtext>) = y(t</mtext></mrow><mrow><mtext>n</mtext></mrow></msub><msub><mrow><mtext>) - a</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext>y(t</mtext></mrow><mrow><mtext>n-1</mtext></mrow></msub><msub><mrow><mtext>) - a</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><msub><mrow><mtext>y(t</mtext></mrow><mrow><mtext>n-2</mtext></mrow></msub><msub><mrow><mtext>) - ... - a</mtext></mrow><mrow><mtext>p</mtext></mrow></msub><msub><mrow><mtext>y(t</mtext></mrow><mrow><mtext>n-p</mtext></mrow></msub><mtext>)</mtext></mrow></math><img file="EP0732687B2_D0001.tif" /></maths> The spectral envelope parameters output from LPC analyzer 107 are converted by a spectral envelope converter 109 into spectral envelope parameters of a bandwidth wider than the bandwidth of the IIR filter constructed with the spectral envelope parameters output from LPC analyzer 107. On the other hand, the residual signal output from LPC analyzer 107 is converted by a residual converter 110 into a residual signal of a bandwidth wider than that of the residual signal output from LPC analyzer 107. An LPC synthesizer 108 synthesizes a digital speech signal from the output of spectral envelope converter 109 and the output of residual converter 110.

0016The spectral envelope converter 109 converts the input spectral envelope parameters into spectral envelope parameters of a wider bandwidth as follows. Namely, assuming â and <img file="EP0732687B2_D0002.tif" /> denote an input feature vector having p elements comprising the input spectral envelope parameters and an output or converted feature vector obtained by k th linear mapping function of matrix B_k = (b^k_i j) (i,j = 1, ..., p, k = 1, ..., M ; the number of linear mapping functions), respectively, <img file="EP0732687B2_D0003.tif" /> is given by the following equation.<maths id="math0002"><img file="EP0732687B2_D0004.tif" /></maths>

0017Spectral envelope converter 109 can also be realized by a composition shown in Fig. 2. In this composition, spectral envelope converter 109 comprises a spectral envelope codebook 201 that has a M spectral envelope codes, for instance sixteen codes, each of which is representative of a set of spectral envelope parameters, and a linear mapping function codebook 202 that has M linear mapping functions, each of which corresponds to a spectral envelope code of spectral envelope codebook 201 one to one. The spectral envelope codes are created by dividing a multidimensional space of the spectral envelope parameters into M subspaces and by averaging the spectral envelope parameter vectors belonging to each subspace. For example, if the <i>j</i>th feature value of the ith spectral envelope parameter vector belonging to a subspace is a_ij, then the jth feature value c_j of the spectral envelope code corresponding to that subspace is<maths id="math0003"><img file="EP0732687B2_D0005.tif" /></maths> where <i>R</i> is the number of spectral envelope parameter vectors (feature vectors) belonging to the subspace.

0018The spectral envelope parameters obtained by LPC analyzer 107 are fed to a distance calculator 203, and a linear mapping function calculator 205. Distance calculator 203 calculates the distance between the spectral envelope parameters a(j),j = 1, ... , p output from LPC analyzer 107 and each spectral envelope code stored in spectral envelope codebook 201. If thejth feature value of the ith spectral envelope code is c_ij, then the distance is obtained by the equation<maths id="math0004"><img file="EP0732687B2_D0006.tif" /></maths> where i = 1, ... , M, and M is the number of spectral envelope codes which is equal to the number of the divided subspaces. The calculated results of distance calculator 203 are input to a comparator or selector 204. Comparator 204 selects the minimum distance of the input multiple distances and outputs, into linear mapping function calculator 205, a linear mapping function stored in linear transformation codebook 202 and corresponding to the linear spectral code that gives the selected minimum distance. Linear mapping function calculator 205 performs computation similar to the equation (2) based on the spectral envelope parameters output from LPC analyzer 107 and the linear transformation output from comparator 204. The output of linear mapping function calculator 205 is the converted spectral envelope parameters in the present composition.

0019In the following, a learning method for determining spectral envelope codes and corresponding linear mapping functions is explained. <ul id="ul0003" list-style="none" compact="compact"><li>(a) A plurality of word speech samples of a wideband are prepared.</li><li>(b) Each of these word speech samples is LPC analyzed to obtain LPC parameters of the wideband.</li><li>(c) Each of these word speech samples is transformed to a corresponding word speech sample of a narrowband by filtering each original speech using low frequency cut filter and high frequency cut filter. Then, each word speech sample of the narrowband is LPC analyzed to obtain LPC parameters of the narrowband.</li><li>(d) Next, a multi-dimension space of feature vectors thus obtained regarding word speech samples of the narrowband is divided into subspaces of an appropriate number. This is done so as to satisfy the following conditions: <ul id="ul0004" list-style="none" compact="compact"><li><d1> Consider M subspaces and calculate a mean value of feature vectors belonging to one of M subspaces. A central value obtained by mean values of M subspaces is as close as possible to a central value obtained by averaging all feature vectors now considered.</li><li><d2> The number of feature vectors belonging to each subspace is substantially equal to each other. Namely, feature vectors are uniformly distributed over all subspaces.</li></ul></li><li>(e) When the division into M subspaces is achieved, linear mapping functions are sought for M subspaces. Since the relationship between each original word speech and corresponding narrowband word speech has been obtained, each linear mapping function is determined so that a distance between the original word speech of the wideband and a word speech mapped into the corresponding subspace by that linear mapping function can be minimized.</li></ul>

0020Figs. 9 and 10 illustrate a graph of the number of subspaces versus mean distance between original word speeches and word speeches synthesized according to the present invention. Figs. 9 illustrates results obtained regarding male speech and Fig. 10 illustrates those regarding female speech.

0021It is to be noted that the mean distance is minimized at 16 subspaces when 100 word speech samples have been used for learning. In other words, an enough learning with an enough number of word speech samples does not necessitate a plenty of subspaces more than 16. This fact indicates that the method of the present invention can simplify the expansion operation from narrowband to wideband resulting in a quick response.

0022Fig. 3 shows another composition of spectral envelope converter 109. In the composition of Fig. 3, the compositions of spectral envelope codebook 201, linear mapping function codebook 202, distance calculator 203, linear mapping function calculator 205 are the same as in Fig. 2. The spectral envelope parameters output from LPC analyzer 107 are input to distance calculator 203 and linear transformation calculator 205. Distance calculator 203 calculates the distance between the spectral envelope parameters output from LPC analyzer 107 and each spectral envelope code stored in spectral envelope codebook 201. The results are input to weights calculator 301. Weights calculator 301 calculates a weight corresponding to each spectral envelope code by the following equation (5).<maths id="math0005"><img file="EP0732687B2_D0007.tif" /></maths> where w_i is the weight corresponding to the ith spectral envelope code, and d_i is the distance to the ith spectral envelope code calculated by distance calculator 203. On the other hand, linear mapping function calculator 205 reads the spectral envelope parameters â output from LPC analyzer 107 and each linear mapping function B_i (i = 1, ..., M) stored in linear mapping function codebook 202 to transform the former into spectral envelope parameters <img file="EP0732687B2_D0008.tif" /> by a method similar to the equation (2). The output of weights calculator 301 and the output of linear mapping function calculator 205 are input to a linear transformation results adder 302. Linear transformation results adder 302 calculates the converted spectral envelope parameters <img file="EP0732687B2_D0009.tif" /> by the following equation (6).<maths id="math0006"><img file="EP0732687B2_D0010.tif" /></maths>

0023Another composition of spectral envelope converter 109 is shown in Fig. 4. In this composition, spectral envelope converter 109 has a narrowband spectral envelope codebook 401 that has a plurality of spectral envelope codes having narrowband spectral envelope information and a wideband spectral envelope codebook 402 that has spectral envelope codes having wideband spectral envelope information and one-to-one corresponding to the narrowband spectral codes. The spectral envelope parameters output from LPC analyzer 107 are input to the distance calculator 203 of Fig. 2. Using the equation (4), distance calculator 203 calculates the distance between the spectral envelope parameters output from LPC analyzer 107 and each narrowband spectral envelope code stored in narrowband spectral envelope codebook 401 to output the calculated results to comparator 403. Distance calculator 203 can use the following equation (7) in place of the equation (4).<maths id="math0007"><img file="EP0732687B2_D0011.tif" /></maths> where x may be other than 2. Preferably, x may be between 2 and 1.5. Comparator 403 extracts from wideband spectral envelope code book 402 the wideband spectral envelope code corresponding to the narrowband spectral envelope code that gives the minimum value of the distances calculated by distance calculator 203. The extracted wideband spectral envelope code is made to be the converted spectral envelope parameters in the present composition.

0024Another composition of spectral envelope converter 109 is described in Fig. 5. In this composition, a neural network is used to convert spectral envelope parameters. Neural networks are well-known techniques, and can be realized, for example, by the methods described in E.D. Lipmann, "Introduction to computing with neural nets", IEEE ASSP Magazine (1987.4), pp. 4-22. An example is shown in Fig. 5. The spectral envelope parameters output from LPC analyzer 107 are input to a neural network 501. If the input spectral envelope parameters are a(i) i = 1, ..., p, then the converted spectral envelope parameters in the present method, fa(k), are<maths id="math0008"><img file="EP0732687B2_D0012.tif" /></maths><maths id="math0009"><img file="EP0732687B2_D0013.tif" /></maths> where w_ij and w_jk are respectively the weights between the ith layer and the jth layer and the weights between the jth layer and the kth layer. Besides the three-layer composition shown in Fig. 5, the neural network may be constructed with a greater number of layers. Further, the equations for calculation may be different from (8) and (9).

0025Next, a preferred example of residual converter 110 is described in the following with reference to Fig. 6. The residual signal output from LPC analyzer 107 is fed to a power calculator 601 and a nonlinear processor 602. Power calculator 601 calculates the power of the residual signal by summing the powers of each value of the residual signal and dividing the result by the sample number. Specifically, the power g is calculated by<maths id="math0010"><img file="EP0732687B2_D0014.tif" /></maths> where r(i), i = 1, ..., p are the residual signal values. Nonlinear processor 602 performs nonlinear processing of the residual signal to obtain a processed residual signal. The processed residual signal is fed to a power calculator 603 and a gain controller 604. Gain controller 604 multiplies the processed residual signal output from nonlinear processor 602 by the ratio of the power obtained by power calculator 601 to the power obtained by power calculator 603. That is, if the residual signal values processed by nonlinear processor 602 are nr(i), i = 1, ..., p, then the residual signal values fnr(i), i = 1, ..., p output from gain controller are calculated by<maths id="math0011" num="(11)"><math display="block"><mrow><msub><mrow><mtext>fnr(i) = g</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext>/g</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><mtext> · nr(i),</mtext></mrow></math><img file="EP0732687B2_D0015.tif" /></maths> where g₁ is the power obtained by power calculator 601 and g₂ is the power obtained by power calculator 603. These fn(i) are the output of the residual converter 110 of the present example.

0026Nonlinear processor 602 can be realized using full-wave rectification or half-wave rectification. Alternatively, nonlinear processor 602 can be realized by setting a threshold value and fixing the residual signal values at the threshold value if the magnitude of the original residual signal values exceeds the threshold value. In this case, the threshold value is preferably determined based on the power obtained by power calculator 601. For example, the threshold value is set at 0.8·g₁, where g₁ is the power output from power calculator 601. Other methods of calculating the threshold value are also possible.

0027Another composition of nonlinear processor 602 can be realized using the multi-pulse method. The multi-pulse method is well known and described, for example, in B. S. Atal et al., "A new model of LPC excitation for producing natural sound speech at very low bit rates", Proceed. ICASSP (1982), pp. 614-617. In this compostion, nonlinear processor 602 generates multi-pulses to perform nonlinear processing of the residual signal obtained by LPC analyzer 107.

0028In the following is described a second embodiment in accordance with the present invention. As shown in Fig. 7, the present embodiment has a waveform smoother 111 between the bandwidth expander 106 and the filter section 105 of Fig. 1.

0029The composition of waveform smoother 111 is described in the following using its schematic illustration of Fig. 8. When the output signal of bandwidth expander 106 is obtained for each determined time period (frame length), there exists discontinuity between the subsequent frames, if the subsequent frame signals are simply connected to output to filter 105 as they are. In the composition of the second embodiment, the discontinuity between the frame signals is mitigated by waveform smoother 111. If bandwidth expander 106 is constructed so as to temporarily overlap the subsequent frame signals, then the output frame signals are overlapped as shown in (a) and (d) of Fig. 8. Waveform smoother 111 multiplies the output signals of bandwidth expander 106 by waveform smoothing functions to add them over the time domain, as shown in Fig. 8. Specifically, the output frame signals (a) and (d) of bandwidth expander 106 are respectively multiplied by the smoothing function (b) and (e) of Fig. 8. The resulting signals (c) and (f) are then added over the time domain to output the signal (g). Let the output of waveform smoother 111 and the output of bandwidth expander 106 be respectively D(N, x) and F(N, x), where N is the frame number and x is the time within each frame. Let the waveform smoothing weight functions for the past frame and the present frame be respectively CFB and CFF,<maths id="math0012" num="(12)"><math display="block"><mrow><mtext>D(N,x) = CFB(x)·F(N-1, x) + CFF(x)·F(N, x).</mtext></mrow></math><img file="EP0732687B2_D0016.tif" /></maths> Preferably, CFB and CFF are defined as<maths id="math0013" num="(13)"><math display="block"><mrow><mtext>CFB(x) = (-2 ·x + L)/L,</mtext></mrow></math><img file="EP0732687B2_D0017.tif" /></maths><maths id="math0014" num="(14)"><math display="block"><mrow><mtext>CFF(x) = 2·x/L,</mtext></mrow></math><img file="EP0732687B2_D0018.tif" /></maths> where L is the frame length.

0030Fig. 11 illustrates results of a subjective test for evaluating the present invention. Test conditions are as follows; <ul id="ul0005" list-style="none" compact="compact"><li>(a) Content of test Hearing test of an original speech of narrowband and corresponding speech of wideband recovered according to the present invention. </li><li>(b) Manner of evaluation Seven steps evaluation whether or not the synthesized speech has an expanded frequency range in comparison with the original speech of narrowband. <ul id="ul0006" list-style="none" compact="compact"><li>* 0 point : not distinguishable,</li><li>* 1 (-1) point : slightly distinguishable from the original speech (synthesized one),</li><li>* 2 (-2) point : distinguishable from the original speech (synthesized one), and</li><li>* 3 (-3) point : clearly distinguishable from the original speech (synthesized one)</li></ul></li><li>(c) Number of tested persons 12 persons including researchers of phonetics.</li><li>(d) Number of linear mapping functions used 16 linear mapping functions having been obtained by learning 100 word speech samples</li><li>(e) Sample data used for the test 10 sentences by a single speaker each having a length of about ten seconds</li><li>(f) Used speaker monoral speaker The test was done by making each person hear one set of original and synthesized speeches without noticing which is original one. Each person scored after hearing every one set. The axis of abscissa in Fig. 11 denotes values of seven steps evaluation and that of vertex denotes values of summation by 12 persons.</li></ul>

0031Fig. 11 indicates that speeches synthesized according to the present invention have a widely expanded sensation relative to an original narrowband speech.

0032It is to be noted that A/D converter and D/A converter are omittable in the case that the input speech signal is a digital speech signal for processing.

0033Although the present invention has been fully described in connection with the preferred embodiments thereof with reference to the accompanying drawings, it is to be noted that various changes and modifications are apparent to those skilled in the art. The scope of the invention is only limited by the appended claims.