Prototype waveform phase modeling for a frequency domain interpolative speech codec system

[0001] This application claims benefit under 35 U.S.C. § 119(e) from U.S. Provisional Patent Application Ser. No. 60/268,327 filed on Feb. 13, 2001, and from U.S. Provisional Patent Application Ser. No. 60/314,288 filed on Aug. 23, 2001, the entire contents of both of said provisional applications being incorporated herein by reference.

[0002] 1. Field of the Invention

[0003] The present invention relates to a method and system for coding low bit rate speech for a communications system. More particularly, the present invention relates to a method and apparatus for encoding perceptually important information about the phase components of a prototype waveform.

[0004] 2. Background of the Invention

[0005] Currently, various speech encoding techniques are used to process speech. These techniques do not adequately address the need for a speech encoding technique that improves the modeling and quantization of a speech signal, specifically, the spectral characteristics of a speech prediction residual signal which includes a prototype waveform (PW) gain vector, a PW magnitude vector, and a PW phase information.

[0006] In particular, prior art techniques are representative but not limited to the following see, e.g., L. R. Rabiner and R. W. Schafer, “Digital Processing of Speech Signals” Prentice-Hall 1978 (hereinafter known as reference 1), W. B. Klejin and J. Haagen, “Waveform Interpolation for Coding and Synthesis”, in Speech Coding and Synthesis, Edited by W. B. Klejin, K. K. Paliwal, Elsevier, 1995 (hereinafter known as reference 2); F. latakura, “Line Spectral Representation of Linear Predictive Coefficients of Speech Signals”, Journal of Acoustical Society of America, vol 4. 57, no. 1, 1975 (hereinafter known as reference 3); P. Kabal and R. P. Ramachandran, “The Computation of Line Spectral Frequencies Using Chebyshev Polybimials”, IEEE Trans. On ASSP, vol. 34, no. 6, pp. 1419-1426, December 1986 (hereinafter known as reference 4); W. B. Klejin, “Encoding Speech Using Prototype Waveforms” IEEE Transactions on Speech and Audio Processing, Vol. 1, No. 4, 386-399, 1993 (hereinafter known as reference 5); and W. B. Kleijn, Y. Shoman, D. Sen and R. Hagen, “A Low Complexity Waveform Interpolation Coder”, IEEE International Conference on Acoustics, Speech and Signal Processing, 1996 (hereinafter known as reference 6). All of the references 1 through 6 are herein incorporated in their entirety by reference.

[0007] The prototype waveforms are a sequence of complex Fourier transforms evaluated at pitch harmonic frequencies, for pitch period wide segments of the residual, at a series of points along the time axis. Thus, the PW sequence contains information about the spectral characteristics of the residual signal as well as the temporal evolution of these characteristics. A high quality of speech can be achieved at low coding rates by efficiently quantizing the important aspects of the PW sequence.

[0008] In PW based coders, the PW is separated into a shape component and a level component by computing the RMS (or gain) value of the PW and normalizing the PW to a unity RMS value. As the pitch frequency varies, the dimensions of the PW vectors also vary, typically in the range of 11-61.

[0009] A PW magnitude vector sequence contains the evolving spectral characteristics of a linear predictive (LP) excitation signal and therefore is important in signal compression. Prior art techniques separate the PW sequence into slowly evolving and rapidly evolving components. This results in three disadvantages.

[0010] First the algorithmic delay of the prior art coding schemes are significantly increased and requires linear low pass and high pass filtering to separate the SEW and REW components. This delay can be noticeable in telephone conversations.

[0011] Second, the signal processing process used in the prior art is complicated due to the filters that are involved. This increases the cost and time to process the signal.

[0012] Third, performance of the prior art is poor at low coding rates. This is due to the fact that only SEW and REW magnitudes are coded in the prior art. Specifically, at the decoder phase models are used to obtain SEW and REW phases. Therefore, even if the SEW and REW magnitude spectra were accurately encoded, the magnitude of the sum of the complex SEW and REW vectors cannot come close to the original PW magnitude spectrum because the phases are estimated in the case of the prior art.

[0013] In addition, some prior art methods, references 2-6, employ a binary model based on a periodic phase or a random phase to encode SEW and REW phases. This results in poor performance because it is based on a binary voicing decision with only two states.

[0014] In some cases of prior art, the SEW phase is obtained at the receiver by a fixed phase model. The REW phase is obtained at a receiver using random phase models. The use of fixed and random phase models results in reconstructed speech that is excessively rough or excessively periodic due to the approximations made.

[0015] In prior art, at the receiver, the PW phase is determined by a vector addition of the SEW and REW vectors. Even if the SEW and REW magnitudes are preserved exactly, the PW magnitude cannot be accurately reproduced at the receiver.

[0016] Thus, a need exists for a system and method that provides information about the PW phase such that the characteristics of the PW phase can be reproduced at the decoder. Furthermore, a need exists for a system and method that provides for reproducing the phase characteristics of the PW phase without compromising the accuracy of the reproduction of the PW magnitude information.

[0017] An object of the present invention is to provide a system and method for providing encoding information related to the PW phase that can recreate characteristics of the PW phase at a decoder. Another object of the present invention is to provide a system and method that provides for reproducing the phase characteristics of the PW phase without compromising the accuracy of the reproduction of the PW magnitude information.

[0018] These and other objects are substantially achieved by a system and method employing a frequency domain interpolative CODEC system for low bit rate coding of speech. The CODEC comprises a linear prediction (LP) front end adapted to process an input signal that provides LP parameters which are quantized and encoded over predetermined intervals and used to compute a LP residual signal. An open loop pitch estimator adapted to process the LP residual signal, a pitch quantizer, and a pitch interpolator and provide a pitch contour within the predetermined intervals is also provided. Also provided is a signal processor responsive to the LP residual signal and the pitch contour and adapted to perform the following: provide a voicing measure, where the voicing measure characterizes a degree of voicing of the input speech signal and is derived from several input parameters that are correlated to degrees of periodicity of the signal over the predetermined intervals; extract a prototype waveform (PW) from the LP residual and the open loop pitch contour for a number of equal sub-intervals within the predetermined intervals; normalize the PW by a gain value of the PW; encode a magnitude of the PW; and separate stationary and nonstationary components of the PW using a low complexity alignment process and a filtering process that introduce no delay. The ratio of the energy of the nonstationary component of the PW to that of the stationary component of the PW is averaged across 5 subbands to compute the nonstationarity measure as a frequency dependent vector entity. A measure of the degree of voicing of the residual is also computed using openloop pitchgain, pitch variance, relative signal power, PW correlation and PW nonstationarity in low frequency subbands. The nonstationarity measure and voicing measure are encoded using a 6-bit spectrally weighted vector quantization scheme using a codebook partitioned based on a voiced/unvoiced decision. At the decoder, a stationary component of PW is reconstructed as a weighted combination of the previous PW phase vector, a random phase perturbation and a fixed phase vector obtained from a voiced pitch pulse.

[0029] Throughout the drawing figures, like reference numerals will be understood to refer to like parts and components.

[0030] FIGS. 1A and 1B are block diagrams of a Frequency Domain Interpolative (FDI) coder/decoder (CODEC) 100 for performing coding and decoding of an input voice signal in accordance with an embodiment of the present invention. The FDI CODEC 100 comprises a coder portion 100A which computes prototype waveforms (PW) and a decoder portion 100B which reconstructs the PW and speech signal.

[0031] Specifically, the coder portion 100A illustrates the computation of PW from an input speech signal. Voice activity detection (VAD) 102 is performed on the input speech to determine whether the input speech is actually speech or noise. The VAD 102 provides a VAD flag which indicates whether the input signal was noise or speech. The detected signal is then provided to a noise reduction module 104 where the noise level for the signal is reduced and provided to a linear predictive (LPC) analysis filter module 106.

[0032] The LPC module 106 provides filtered and residual signals to the prototype extraction module 108 as well as LPC parameters to decoder 100B. The pitch estimation and interpolation module 110 receives the LPC filtered and residual signals from the LPC analysis filter module 106 and pitch contours from the prototype extraction module 108 and provides a pitch and a pitch gain.

[0033] The extracted prototype waveform from prototype extraction module 108 is provided to compute prototype gain module 112, PW magnitude and computation and normalization module 114, compute subband nonstationarity measure module 116 and compute voicing measure module 118. Compute voicing measure (VM) module 118 also receives the pitch gain from pitch estimation and interpolation module 110 and computes a voicing measure.

[0034] The compute prototype gain module 112 computes a prototype gain and provides the PW gain value to decoder portion 100B. PW magnitude computation and normalization module 114 computes the PW magnitude and normalizes the PW magnitude.

[0035] Compute subband nonstationarity measure module 116 computes a subband nonstationarity measure from the extracted prototype waveform. The computed subband nonstationarity measure and computed voicing measure are provided to a subband nonstationarity measure—Vector quantizer (VQ) module 122 which processes the received signals.

[0036] A PW magnitude quantization module 120 receives the computed PW magnitude and normalized signal along with the VAD flag indication and quantizes the received signal and provides a PW magnitude value to the decoder 100B.

[0037] The decoder 100B further includes a periodic phase model module 124 and aperiodic phase model module 126 which receive the PW magnitude value and subband nonstationarity measure-voicing measure value from coder 100A and compute a periodic phase and an aperiodic phase, respectively, from the received signal. The periodic phase model module 124 provides a complex periodic vector having a periodic component level and the aperiodic phase model module 126 provides a complex aperiodic vector having an aperiodic component level to a summer which provides a complex PW vector to a normalize PW gain module 128. The normalize PW gain module also receives the PW gain value from coder 100A.

[0038] A pitch interpolation module 130 performs pitch interpolation on a pitch period provided by encoder 100A. The normalize PW gain signal and interpolated pitch frequency contour signal is provided to an interpolative synthesis module 132 which performs interpolative synthesis to obtain a reconstructed residual signal from the previously mentioned signals.

[0039] The reconstructed residual signal is provided to an all pole LPC synthesis filter module 134 which processes the reconstructed residual signal and provides the filtered signal to an adaptive postfilter and tilt correction module 136. Modules 134 and 136 also receive the VAD flag indication signal and interpolated LPC parameters from the encoder 100A. A reconstructed speech signal is provided by the adaptive postfilter and tilt correction module 136.

[0040] Specifically, the FDI codec 100 is based on techniques of linear predictive (LP) analysis, robust pitch estimation and frequency domain encoding of the LP residual signal. The FDI codec operates on a frame size of preferably 20 ms. Every 20 ms, the speech encoder 100A produces 80 bits representing compressed speech. The speech decoder 100B receives the 80 compressed speech bits and reconstructs a 20 ms frame of speech signal. The encoder 100A preferably uses a look ahead buffer of at least 20 ms, resulting in an algorithmic delay comprising buffering delay and look ahead delay of 40 ms.

[0041] The speech encoder 100A is equipped with a built-in voice activity detector (VAD) 102 and can operate in continuous transmission (CTX) mode or in discontinuous transmission (DTX) mode. In the DTX mode, comfort noise information (CNI) is encoded as part of the compressed bit stream during silence intervals. At the decoder 100B, the CNI packets are used by a comfort noise generation (CNG) algorithm to regenerate a close approximation of the ambient noise. The VAD information is also used by an integrated front end noise reduction scheme that can provide varying degrees of background noise level attenuation and speech signal enhancement.

[0042] A single parity check bit is preferably included in the 80 compressed speech bits of each frame of the input speech signal to detect channel errors in perceptually important compressed speech bits. This enables the codec 100 to operate satisfactorily in links with a random bit error rate up to about 10⁻³. In addition, the decoder 100B uses bad frame concealment and recovery techniques to extend signal processing operations during frame erasures.

[0043] Additionally, in addition to the speech coding functions, the codec 100 also has the ability to transparently pass dual tone multifrequency (DTMF) and signaling tones.

[0044] As discussed above, the FDI codec 100 uses the linear predictive analysis technique to model the short term Fourier spectral envelope of the input speech signal. Subsequently, a pitch frequency estimate is used to perform a frequency domain prototype waveform analysis of the LP residual signal. Specifically, the PW analysis provides a characterization of the harmonic or fine structure of the speech spectrum. More specifically, the PW magnitude spectrum provides the correction necessary to refine the short term LP spectral estimate to obtain a more accurate fit to the speech spectrum at the pitch harmonic frequencies. Information about the phase of the signal is implicitly represented by the degree of periodicity of the signal measured across a set of subbands.

[0045] In a preferred embodiment of the present invention, the input speech signal is processed in consecutive non-overlapping frames of 20 ms duration, which corresponds to 160 samples at the sampling frequency of 8000 samples/sec. The encoder 100A parameters are quantized and transmitted once for each 20 ms frame. A look-ahead of 20 ms is used for voice activity detection, noise reduction, LP analysis and pitch estimation. This produces in an algorithmic delay which is defined as a buffering delay and a look-ahead delay of 40 ms.

[0046] Referring to FIG. 2 which illustrates the samples used for various functions at the encoder 100A, an estimated size of buffered samples for various frames is shown. For example, a VAD window 210 uses buffered samples from about 160 to 400 samples. A noise reduction window 220 uses about the same number of samples. Pitch estimation windows 230₁ up to 230₅ each uses about 240 samples The LP analysis window processes the signal in about 80 to 400 samples. A current frame being encoded is processed between 80 to 240 samples. A new input speech data 260 and look-ahead 280 are processed from about 240 to 400 samples while a past data is processed from zero to 80 samples. For the purposes of excitation modeling, each frame is further divided into 8 subframes preferably of duration 2.5 ms or 20 samples.

[0047] The invention will now be discussed in terms of front end processing, specifically input preprocessing. The new input speech samples are first scaled down by preferably 0.5 to prevent overflow in fixed point implementation of the coder 100A. In another embodiment of the present invention, the scaled speech samples can be high-pass filtered using an infinite impulse response (IIR) filter with a cut-off frequency of 60 Hz, to eliminate undesired low frequency components. The transfer function of the 2nd order high pass filter is given by $\begin{array}{cc}{H}_{\mathrm{hpf1}}\left(z\right)=\frac{0.939819335-1.879638672z^{-1}+0.939819335z^{-2}}{1-1.933195469z^{-1}+0.935913085z^{-2}}& \left(1\right)\end{array}$

[0048] In terms of the VAD module 102, the preprocessed signal is analyzed to detect the presence of speech activity. This comprises the following operations: scaling the signal via an automatic gain control (AGC) mechanism to improve VAD performance for low level signals, windowing the Automatic Gain Control (AGC) scaled speech and computing a set of autocorrelation lags, performing a 10^th order autocorrelation LP analysis of the AGC scaled speech to determine a set of LP parameters which are used during pitch estimation, performing a preliminary pitch estimation based on the pitch candidates for the look-ahead part of the buffer, performing voice activity detection based on the autocorrelation lags and pitch estimate and the tone detection flag that is generated by examining the distance between adjacent line spectral frequencies (LSFs) which will be described in greater detail below with respect to conversion to line spectral frequencies.

[0049] This series of operations produces a VAD_FLAG and a VID_FLAG that have the following values depending on the detected voice activity: $\begin{array}{c}{\mathrm{VAD}}_{—}\mathrm{FLAG}=\{\begin{array}{cc}1& \mathrm{if}\mathrm{voice}\mathrm{activity}\mathrm{is}\mathrm{present},\\ 0& \mathrm{if}\mathrm{voice}\mathrm{activity}\mathrm{is}\mathrm{absent},\end{array}\\ {\mathrm{VID}}_{—}\mathrm{FLAG}=\{\begin{array}{cc}0& \mathrm{if}\mathrm{voice}\mathrm{activity}\mathrm{is}\mathrm{present},\\ 1& \mathrm{if}\mathrm{voice}\mathrm{activity}\mathrm{is}\mathrm{absent},\end{array}\end{array}$
It should be noted that the VAD_FLAG and the VID_FLAG represent the voice activity status of the look-ahead part of the buffer. A delayed VAD flag, VAD_FLAG_DL1 is also maintained to reflect the voice activity status of the current frame. In a presentation given during an IEEE speech and audio processing workshop in Finland during 1999, the entire contents of the documentation being incorporated by reference herein, the presenters F. Basbug, S. Nandkumar and K. Swamianthan described an AGC front-end for the VAD which itself is a variation of the voice activity detection algorithms used in cellular standards “TDMA cellular/PCS Radio Interface—Minimum Objective Standards for IS-136 B, DTX/CNG Voice Activity Detection”, which is also incorporated by reference in its entirety. A by-product of the AGC front-end is the global signal-to-noise ratio, which is used to control the degree of noise reduction.

[0050] The VAD flag is encoded explicitly only for unvoiced frames as indicated by the voicing measure flag. Voiced frames are assumed to be active speech. In the present embodiment of the invention, the VAD flag is not coded explicitly. The decoder sets the VAD flag to a one for all voiced frames. However, it will be appreciated by those skilled in the art that the VAD) flag can be coded explicitly without departing from the scope of the present invention.

[0051] Noise reduction module 104 provides noise reduction to the voice activity detected speech signal. Specifically, the preprocessed speech signal is processed by a noise reduction algorithm to produce a noise reduced speech signal. The following is a series of steps comprising the noise reduction algorithm: A trapezoidal windowing and the computing of the complex discrete Fourier transform (DFT) of the signal is performed. FIG. 2 depicts the part of the buffer that undergoes the DFT operation. A 256-point DFT (240 windowed samples+16 padded zeros) is used. The magnitude of the DFT is smoothed along the frequency axis across a variable window whose width is about 187.5 Hz in the first 1 KHz, about 250 Hz in the range of 1-2 KHz, and about 500 Hz in the range of 2-4 KHz regions. These values reflect a compromise between the conflicting objectives of preserving the format structure and having sufficient smoothness of the speech signal.

[0052] If the VVAD_FLAG, which is the VAD output prior to hangover, is a one which indicates voice activity, then the smoothed magnitude square of the DFT is taken to be the smoothed power spectrum of noisy speech S(k). However, if the VVAD_FLAG is a zero indicating voice inactivity, the smoothed DFT power spectrum is then used to update a recursive estimate of the average noise power spectrum N_av(k) as follows:
N_av(k)=0.9·N_av(k)+0.1·S(k) if VAD_FLAG=0   (2)
A spectral gain function is then computed based on the average noise power spectrum and the smoothed power spectrum of the noisy speech. The gain function G_nr(k) takes the following form: $\begin{array}{cc}{G}_{\mathrm{nr}}\left(k\right)=\frac{S\left(k\right)}{F_{\mathrm{nr}}{N}_{\mathrm{av}}\left(k\right)+S\left(k\right)}& \left(3\right)\end{array}$
Here, the factor F_nr is a factor that depends on the global signal-to-noise-ratio SNR_global that is generated by the AGC front-end for the VAD. The factor F_nr can be expressed as an empirically derived piecewise linear function of SNR_global that is monotonically non-decreasing. The gain function is close to unity when the smoothed power spectrum S(k) is much larger than the average noise power spectrum N_av(k). Conversely, the gain function becomes small when S(k) is comparable to or much smaller than N_av(k). The factor F_nr controls the degree of noise reduction by providing for a higher degree of noise reduction when the global signal-to-noise ratio is high (i.e., risk of spectral distortion is low since VAD and the average noise estimate are fairly accurate). Conversely, the factor restricts the amount of noise reduction when the global signal-to-noise ratio is low. For example, the risk of spectral distortion is high due to increased VAD inaccuracies and less accurate average noise power spectral estimate.

[0053] The spectral amplitude gain function is further clamped to a floor which is a monotonically non-increasing function of the global signal-to-noise ratio. This kind of clamping reduces the fluctuations in the residual background noise after noise reduction making the speech sound smoother. The clamping action is expressed as:
G_nr(k)=MAX(G_nr(k), T_global(SNR_global)   (4)
Thus, at high global signal-to-noise ratios, the spectral gain functions will be clamped to a lower floor since there is less risk of spectral distortion due to inaccuracies in the VAD or the average noise power spectral estimate N_av(k). But at lower global signal-to-noise ratio, the risks of spectral distortion outweigh the benefits of reduced noise and therefore a higher floor would be appropriate.

[0054] In order to reduce the frame-to-frame variation in the spectral amplitude gain function, a gain limiting device is used which limits the gain between a range that depends on the previous frame's gain for the same frequency. The limiting action can be expressed as follows:
G_nr^new(k)=MAX({S_nr^L.G_nr^old(k)}, MIN({S_nr^H.G_nr^old(k)}, G′_nr(k)))   (5)
The scale factors S_nr^L and S_nr^H are updated using a state machine whose actions depend on whether the frame is active, inactive or transient.

[0055] FIG. 3 depicts a flowchart 300 which performs scale factor updates in accordance with an embodiment of the present invention. The process 300 is occurs in noise reduction module 104 and is initiated at step 302 where input values VAD_FLAG and scale factors are received. The method 300 then proceeds to step 304 where a determination is made as to whether the VAD_FLAG is zero which indicates voice activity is absent. If the determination is affirmative the method 300 proceeds to step 306 where the scale factors are adjusted to be closer to unity. The method 300 then proceeds to step 308.

[0056] At step 308 a determination is made as to whether the VAD_FLAG was zero for the last two frames. If the determination is affirmative the method proceeds to step 310 where the scale factors are limited to be very close to unity. However, if the determination was negative, the method 300 then proceeds to step 312 where the scale factors are limited to be away from unity.

[0057] If the determination at step 304 was negative, the method 300 then proceeds to step 314 where the scale factors are adjusted to be away from unity. The method 300 then proceeds to step 316 where the scale factors are limited to be far away from unity.

[0058] The steps 310, 312 and 316 proceed to step 318 where the updated scale factors are outputted.

[0059] The final spectral gain function G_nr^new(k) is multiplied with the complex DFT of the preprocessed speech, attenuating the noise dominant frequencies and preserving signal dominant frequencies. An overlap-and-add inverse DFT is then performed on the spectral gain scaled DFT to compute a noise reduced speech signal over the interval of the noise reduction window

[0060] Since the noise reduction is carried out in the frequency domain, the availability of the complex DFT of the preprocessed speech is taken advantage of in order to carry out DTMF and Signaling tone detection. These detection schemes are based on examination of the strength of the power spectra at the tone frequencies, the out-of-band energy, the signal strength, and validity of the bit duration pattern. It should be noted that the incremental cost of having such detection schemes to facilitate transparent transmission of these signals is negligible since the power spectrum of the preprocessed speech is already available.

[0061] An embodiment of the invention will now be described in terms of LPC analysis filtering module 106. The noise reduced speech signal is subjected to a 10^th order autocorrelation method of LP analysis where {s_nr(n), 0≦n<400} denotes the noise reduced speech buffer, where {s_nr(n), 80≦n<240} is the current frame being encoded and {s_nr(n), 240≦n<320} is the look-ahead buffer 280 as shown in FIG. 2.

[0062] In the LP analysis of speech, the magnitude spectrum of short segments of speech is modeled by the magnitude frequency response of an all-pole minimum phase filter, whose transfer function is represented by $\begin{array}{cc}{H}_{\mathrm{lp}}\left(z\right)=\frac{1}{\sum _{m=0}^M{a}_m{z}^{-m}}& \left(6\right)\end{array}$
Here, {a_m, 0≦m≦M} are the LP parameters for the current frame and M=10 is the LP order. LP analysis is performed using the autocorrelation method with a modified Harming window of size 40 ms (320 samples) which includes the 20 ms current frame and the 20 ms lookahead frame as shown in FIG. 2.

[0063] The noise reduced speech signal over the LP analysis window {s_nr(n), 80≦n<400} is windowed using a modified Hanning window function {w_lp(n), 0≦n<320} defined as follows: $\begin{array}{cc}{w}_{\mathrm{lp}}\left(n\right)=\{\begin{array}{cc}0.5-0.5\cos \left(\frac{2\pi n}{319}\right),& 0\le n<240,\\ \frac{\left(0.5-0.5\cos \left(\frac{2\pi n}{319}\right)\right)}{{\cos }^2\left(\frac{2\pi \left(n-240\right)}{320}\right)},& 240\le n<320\end{array}& \left(7\right)\end{array}$
The windowed speech buffer is computed by multiplying the noise reduced speech buffer with the window function as follows:
s_w(n)=s_nr(80+n)w_lp(n) 0≦n<240.   (8)
Normalized autocorrelation lags are computed from the windowed speech by $\begin{array}{cc}{r}_{\mathrm{lp}}\left(m\right)=\frac{\sum _{n=0}^{319-m}{s}_w\left(n\right)s_w\left(n+m\right)}{\sum _{n=0}^{319}{s}_w^2\left(n\right)}0\le m\le 10,& \left(9\right)\end{array}$
The autocorrelation lags are windowed by a binomial window with a bandwidth expansion of 60 Hz. The binomial window is given by the following recursive rule: $\begin{array}{cc}{l}_w\left(m\right)=\{\begin{array}{cc}1& m=0\\ l_w\left(m-1\right)\frac{4995-m}{4994+m}& 1\le m\le 10.\end{array}& \left(10\right)\end{array}$
Lag windowing is performed by multiplying the autocorrelation tags by the binomial window:
r_lpw(m)=r_lp(m)l_w(m) 1≦m≦10.   (11)
The zeroth windowed lag r_lpw(0) is obtained by multiplying by a white noise correction factor of about 1.0001, which is equivalent to adding a noise floor at −40 dB:
r_lpw(0)=1.0001r_lp(0).   (12)

[0064] Lag windowing and white noise correction are techniques are used to address problems that arise in the case of periodic or nearly periodic signals. For such signals, the all-pole LP filter is marginally stable, with its poles very close to the unit circle. It is necessary to prevent such a condition to ensure that the LP quantization and signal synthesis at the decoder 100B can be performed satisfactorily.

[0065] The LP paramerters that define a minimum phase spectral model to the short term spectrum of the current frame are determined by applying Levinson-Durbin recursions to the windowed autocorrelation lags {r_lpw(m), 0≦m≦10}. The resulting 10^th order LP parameters for the current frame are {a′_m, 0≦m≦10}, with a′₀=1. Since the LP analysis window is centered around the sample index of about 240 in the buffer, the LP parameters represent the spectral characteristics of the signal in the vicinity of this point.

[0066] During highly periodic signals, the spectral fit provided by the LP model tends to be excessively peaky in the low formant regions, resulting in audible distortions. To overcome this problem, a bandwidth broadening scheme has been employed in this embodiment of the present invention, where the formant bandwidth of the model is broadened adaptively, depending on the degree of peakiness of the spectral model. The LP spectrum is given by $\begin{array}{cc}S\left(e^{\mathrm{jw}}\right)=\frac{1}{\sum _{m=0}^M{a}_m^{\prime }{e}^{-\mathrm{jwm}}}-\pi \le w\le \pi .& \left(13\right)\end{array}$
where ω_m denotes the pitch frequency estimate of the m^th subframe (1≦m≦8) of the current frame in radians/sample. Given this pitch frequency, the index of the highest frequency pitch harmonic that falls within the frequency band of the signal (0-4000 Hz or 0-π radians) for the m^th subframe is given by $\begin{array}{cc}{K}_m=\lfloor \frac{\pi }{{\omega }_m}\rfloor 1\le m\le 8,& \left(14\right)\end{array}$
where, └x┘ denotes the largest integer less than or equal to x. The magnitude of the LPC spectrum is evaluated at the pitch harmonics by $\begin{array}{cc}\left|S\left(k\right)\right|=|S\left(e^{j{\omega }_{g}k}\right)❘=\frac{1}{\left|\sum _{m=0}^M{a}_m^{\prime }{e}^{-1{\omega }_g\mathrm{km}}\right|}0\le k\le K_g.& \left(15\right)\end{array}$
It should be noted that ω₈ corresponds to the 8^th subframe has been used here since the LP parameters have been evaluated for a window centered around a sample of about 240 as shown in FIG. 2. A logarithmic peak-to-average ratio of the harmonic spectral magnitudes is computed as $\begin{array}{cc}\mathrm{PAR}=10{\log }_{10}\left\{\frac{\underset{1\le k\le K_g}{\mathrm{MAX}}S\left(k\right)}{\frac{1}{\left(K_g-1\right)}\left\{\sum _{k=1}^{K_g}S\left(k\right)-\underset{1\le k\le \mathrm{Kg}}{\mathrm{MAX}}S\left(k\right)\right\}}\right\}.& \left(16\right)\end{array}$
The peak-to-average ratio ranges from 0 dB (for flat spectra) to values exceeding 20 dB (for highly peaky spectra). The expansion in formant bandwidth (expressed in Hz) is then determined based on the log peak-to-average ratio according to a piecewise linear characteristic: $\begin{array}{cc}{\mathrm{dw}}_{l_p}=\{\begin{array}{cc}10+2\mathrm{PAR}& \mathrm{PAR}\le 5,\\ 20+12\left(\mathrm{PAR}-5\right),& \mathrm{PAR}\le 10,\\ 80+4\left(\mathrm{PAR}-10\right),& \mathrm{PAR}\le 20,\\ 120& \mathrm{PAR}>20.\end{array}& \left(17\right)\end{array}$
The expansion in bandwidth ranges from a minimum of about 10 Hz for flat spectra to a maximum of about 120 Hz for highly peaky spectra. Thus, the bandwidth expansion is adapted to the degree of peakiness of the spectra. The above piecewise linear characteristic have been experimentally optimized to provide the right degree of bandwidth expansion for a range of spectral characteristics. A bandwidth expansion factor a_bw to apply this bandwidth expansion to the LP spectrum is obtained by $\begin{array}{cc}{\alpha }_{\mathrm{bw}}=e^{-\frac{\pi {\mathrm{dw}}_{\mathrm{ip}}}{8000}}.& \left(18\right)\end{array}$
The LP parameters representing the bandwidth expanded LP spectrum are determined by
a_m=a′_mα_bw^m 0≦m≦10.   (19)

[0067] The bandwidth expanded LP filter coefficients are converted to line spectral frequencies (LSFs) for quantization and interpolation purposes which is described in “Line Spectral Representation of Linear Predictive Coefficients of Speech Signals” Journal of Acoustical Society of America, vol. 57, no. 1, 1975 by F. Itakura which is incorporated by reference in its entirety. An efficient approach to computing LSFs from LP parameters using Chebychev polynomials is described in “The Computation of Line Spectral Frequencies Using Chebyshev Polynomials,” IEEE Trans. On ASSP, vol. 34, no 6, pages 1419-1426, December 1986 by P. Kabal and R. P. Ramachandran which is herein incorporated by reference in its entirety. The resulting LSFs for the current frame are denoted by {λ(m), 0≦m≦10}.

[0068] The LSF domain also lends itself to detection of highly periodic or resonant inputs. For such signals, the LSFs located near the signal frequency have very small separations. If the minimum difference between adjacent LSF values falls below a threshold for a number of consecutive frames, it is highly probable that the input signal is a tone.

[0069] FIG. 4 describes a method 400 for tone detection in accordance with an embodiment of the present invention. The method 400 occurs in LPC analysis filtering module 106 and is initiated at step 402 where a tone counter is set illustratively for a maximum of 16. The method 400 then proceeds to step 404 where a determination is made as to whether the LSF value falls below a minimum threshold of for example 0.008. If the determination is answered negatively, the method 400 then proceeds to step 406 where the tone counter detects that the LSF value is above the threshold.

[0070] If the method 404 is answered affirmatively, the tone counter detects that the LSF value is below the threshold and increments the counter by one. The methods 406 and 412 proceed to step 408.

[0071] At step 408 a determination is made as to whether the tone counter is at its maximum value. If the method 408 is answered negatively, the method 400 proceeds to step 410 where a tone flag equals false indication is provided. If the method 408 is answered negatively, the method 400 then proceeds to step 414 where a tone flag equals true indication is provided.

[0072] The steps 410 and 44 proceed to step 416 where the method 400 continues checking for tones. Specifically, method 400 provides a tone flag indication which is a one if a tone has been detected and a zero otherwise. This flag is also used in voice activity detection.

[0073] The invention will now be described in reference to the pitch estimation and interpolation module 110. Pitch estimation is performed based on an autocorrelation analysis of a spectrally flattened low pass filtered speech signal. Spectral flattening is accomplished by filtering the AGC scaled speech signal using a pole-zero filter, constructed using the LP parameters of AGC scaled speech signal. If {a_m^agc, 0≦m≦10} are the LP parameters of AGC scaled speech signal, the pole-zero filter is given by $\begin{array}{cc}{H}_{\mathrm{sf}}\left(z\right)=\frac{\sum _{m=0}^M{a}_m^{\mathrm{age}}{z}^{-m}}{\sum _{m=0}^M{a_m^{\mathrm{age}}\left(0.8\right)}^m{z}^{-m}}.& \left(20\right)\end{array}$
The spectrally flattened signal is low-pass filtered by a 2^nd order IIR filter with a 3 dB cutoff frequency of 1000 Hz. The transfer function of this filter is $\begin{array}{cc}{H}_{\mathrm{lpf}1}\left(z\right)=\frac{0.06745527-0.134910548z^{-1}+0.06745527z^{-2}}{1-1.14298050z^{-1}+0.41280159z^{-2}}.& \left(21\right)\end{array}$

[0074] The resulting signal is subjected to an autocorrelation analysis in two stages. In the first stage, a set of four raw normalized autocorrelation functions (ACF) are computed over the current frame. The windows for the raw ACFs are staggered by 40 samples as shown in FIG. 2. The raw ACF for the i^th window is computed by $\begin{array}{cc}{r}_{\mathrm{raw}}\left(i,l\right)=\frac{\sum _{n=40\left(i-1\right)}^{40\left(i-1\right)+239-l}{S}_{\mathrm{sf}}\left(n\right)s_{\mathrm{sf}}\left(n+l\right)}{\sum _{n=40\left(i-1\right)}^{40\left(i-1\right)+239}{S}_{\mathrm{sf}}^2\left(n\right)}15\le l\le 125,2\le i\le 5.& \left(22\right)\end{array}$

[0075] In each frame, raw ACFs corresponding to windows 2, 3, 4 and 5 as shown in FIG. 2 are computed. In addition, a raw ACF for window 1 is preserved from the previous frame. For each raw ACF, the location of the peak within the lag range 20≦l≦120 is determined.

[0076] In the second stage, each raw ACF is reinforced by the preceding and the succeeding raw ACF, resulting in a composite ACF. For each lag l in the raw ACF in the range 20≦l≦120, peak values within a small range of lags [(l−w_c(l)), (l+w_c(l))] are determined in the preceding and the succeeding raw ACFs. These peak values reinforce the raw ACF at each lag l, via a weighted combination: $\begin{array}{cc}{r}_{\mathrm{comp}}\left(i,l\right)=\frac{w_c\left(l\right)+1-0.1m_{\mathrm{peak}}\left(l\right)}{\left(w_c\left(l\right)+1\right)}\left[\underset{l-w_c\left(l\right)\le m\le l+w_c\left(l\right)}{\mathrm{MAX}}{r}_{\mathrm{raw}}\left(i-1,m\right)\right]+r_{\mathrm{raw}}\left(i,l\right)+\frac{w_c\left(l\right)+1-0.1n_{\mathrm{peak}}\left(l\right)}{\left(w_c\left(l\right)+1\right)}\left[\underset{l-w_c\left(l\right)\le n\le l+w_c\left(l\right)}{\mathrm{MAX}}{r}_{\mathrm{raw}}\left(i+1,n\right)\right]20\le l\le 120,2\le i\le 5.& \left(23\right)\end{array}$
Here, w_c(l) determines the window length based on the lag index l: $\begin{array}{cc}{w}_c\left(l\right)=\{\begin{array}{cc}2& l<30\\ \lfloor 0.05l+0.5\rfloor & 30\le l\le 70\\ 4& l>70.\end{array}& \left(24\right)\end{array}$

[0077] Also, m_peak(l) and n_peak(l) are the locations of the peaks within the window. The weighting attached to the peak values from the adjacent ACFs ensures that the reinforcement diminishes with increasing difference between the peak location and the lag l. The reinforcement boosts a peak value if peaks also occur at nearby lags in the adjacent raw ACFs. This increases the probability that such a peak location is selected as the pitch period. ACF peaks locations due to an underlying periodicity do not change significantly across a frame. Consequently, such peaks are strengthened by the above process. On the other hand, spurious peaks are unlikely to have such a property and consequently arc diminished. This improves the accuracy of pitch estimation.

[0078] Within each composite ACF the locations of the two strongest peaks are obtained. These locations are the candidate pitch lags for the corresponding pitch window, and take values in the range 20-120 which is inclusive. In conjunction with the two peaks from the last composite ACF of the previous frame i.e., for window 5 in the previous frame, results in a set of 5 peak pairs, leading to 32 possible pitch tracks through the current frame. A pitch metric is used to maximize the continuity of the pitch track as well as the value of the ACF peaks along the pitch track to select one of these pitch tracks. The end point of the optimal pitch track determines the pitch period p₈ and a pitch gain β_pitch for the current frame. Note that due to the position of the pitch windows, the pitch period and pitch gain are aligned with the right edge of the current frame The pitch period is integer valued and takes on values in the range 20-120. It is mapped to a 7-bit pitch index l*_p in the range of about 0-101.

[0079] In respect to the prototype extraction module 108 and the pitch estimation and interpolation module 110, the pitch period is converted to the radian pitch frequency corresponding to the right edge of the frame by $\begin{array}{cc}{\omega }_8=\frac{2\pi }{p_8}.& \left(24\right)\end{array}$
A subframe pitch frequency contour is created by linearly interpolating between the pitch frequency of the left edge ω₀ and the pitch frequency of the right edge ω₈: $\begin{array}{cc}{\omega }_m=\frac{\left(8-m\right){\omega }_0+m{\omega }_8}{8},1\le m\le 8.& \left(25\right)\end{array}$
If there are abrupt discontinuities between the left edge and the right edge pitch frequencies, the above interpolation is modified to make a switch from the pitch frequency to its integer multiple or submultiple at one of the subframe boundaries. It should be noted that the left edge pitch frequency ω₀ is the right edge pitch frequency of the previous frame.
The index of the highest pitch harmonic within the 4000 Hz band is computed for each subframe by $\begin{array}{cc}{K}_m\lfloor \frac{\pi }{{\omega }_m}\rfloor ,1\le m\le 8.& \left(26\right)\end{array}$

[0080] The LSFs are quantized by a hybrid scalar-vector quantization scheme. The first 6 LSFs are scalar quantized using a combination of intraframe and interframe prediction using 4 bits/LSF. The last 4 LSFs are vector quantized using 7 bits. Thus, a total of 31 bits are used for the quantization of the 10-dimensional LSF vector.

[0081] The 16 level scalar quantizers for the first 6 LSFs in a preferred embodiment of the present invention is designed using a Linde-Buzo-Gray algorithm. An LSF estimate is obtained by adding each quantizer level to a weighted combination of the previous quantized LSF of the current frame and the adjacent quantized LSFs of the previous frame: $\begin{array}{cc}\tilde{\lambda }\left(l,m\right)=\left\{\begin{array}{cc}{S}_{L,m}\left(l\right)+0.375{\hat{\lambda }}_{\mathrm{prev}}\left(m+1\right),& m=0,\\ S_{L,m}\left(l\right)+0.375(\hspace{1em}{\hat{\lambda }}_{\mathrm{prev}}\left(m+1\right)-& \\ {\hat{\lambda }}_{\mathrm{prev}}\left(m-1\right)\hspace{1em})+\hat{\lambda }\left(m-1\right),& 1\le m\le 5,\end{array}\right\}0\le l\le 15.& \left(27\right)\end{array}$
Here, {{circumflex over (λ)}(m), 0≦m<6} are the first 6 quantized LSFs of the current frame and {{circumflex over (λ)}_prev(m), 0≦m≦10} are the quantized LSFs of the previous frame. {S_L,m(l), 0≦m<6, 0≦l≦15} are the 16 level scalar quantizer tables for the first 6 LSFs. The squared distortion between the LSF and its estimate is minimized to determine the optimal quantizer level: $\begin{array}{cc}{\underset{0\le l\le 15}{\mathrm{MIN}}\left(\lambda \left(m\right)-\tilde{\lambda }\left(l,m\right)\right)}^{2}0\le m\le 5.& \left(28\right)\end{array}$

[0082] If l*_L—S—m is the value of l that minimizes the above distortion, the quantized LSFs are given by: $\begin{array}{cc}\hat{\lambda }\left(m\right)=\{\begin{array}{cc}{S}_{L,m}\left(l_{\mathrm{L\_S}\mathrm{\_m}}^{*}\right)+0.375{\hat{\lambda }}_{\mathrm{prev}}\left(m+1\right),& m=0\\ S_{L,m}\left(l_{\mathrm{L\_S}\mathrm{\_m}}^{*}\right)+0.375(\hspace{1em}{\hat{\lambda }}_{\mathrm{prev}}\left(m+1\right)-& \\ {\hat{\lambda }}_{\mathrm{prev}}\left(m-1\right)\hspace{1em})+\hat{\lambda }\left(m-1\right),& 1\le m\le 5.\end{array}& \left(29\right)\end{array}$
The last 4 LSFs are vector quantized using a weighted mean squared error (WMSE) distortion measure. The weight vector {W_L(m), 6≦m≦9} is computed by the following procedure: $\begin{array}{cc}\mathrm{p1}\left(m\right)=\prod _{i=0,2,4,6,8}\left\{4+{\cos }^2\left(2\pi \lambda \left(m\right)\right)+{\cos }^2\left(2\pi \lambda \left(i\right)\right)-8\cos \left(2\pi \lambda \left(m\right)\right)\cos \left(2\pi \lambda \left(i\right)\right)\right\},6\le m\le 9.& \left(30\right)\\ \mathrm{p2}\left(m\right)=\prod _{i=1,3,5,7,9}\left\{4+{\cos }^2\left(2\pi \lambda \left(m\right)\right)+{\cos }^2\left(2\pi \lambda \left(i\right)\right)-8\cos \left(2\pi \lambda \left(m\right)\right)\cos \left(2\mathrm{\pi \lambda }\left(i\right)\right)\right\},6\le m\le 9.& \left(31\right)\\ W_L\left(m\right)={\left[\frac{1.09-0.6\cos \left(2\pi \lambda \left(m\right)\right)}{\begin{array}{c}(0.5+0.5\cos \left(2\pi \lambda \left(m\right)\right)\mathrm{p1}\left(m\right)+\\ (0.5-0.5\cos \left(2\pi \lambda \left(m\right)\right)\mathrm{p2}\left(m\right)\end{array}}\right]}^{0.25},6\le m\le 9.& \left(32\right)\end{array}$

[0083] A set of predetermined mean values {λ_dc(m), 6≦m<9} are used to remove the DC bias in the last 4 LSFs prior to quantization. These LSFs are estimated based on the mean removed quantized LSFs of the previous frame:
{tilde over (λ)}(l,m)=V_L(l,m−6)+λ_dc(m)+0.5({circumflex over (λ)}_prev(m)−λ_dc(m)), 0≦l≦127, 6≦m≦9.   (33)

[0084] Here {V_L(l,m), 0≦l≦127, 0≦m<3} is the 128 level, 4-dimensional codebook for the last 4 LSFs. The optimal code vector is determined by minimizing the WMSE between the estimated and the original LSF vectors: $\begin{array}{cc}\underset{0\le l\le 127}{\mathrm{MIN}}\sum _{m=6}^9{W}_L\left(m\right){\left(\lambda \left(m\right)-\hat{\lambda }\left(l,m\right)\right)}^2.& \left(34\right)\end{array}$

[0085] If l*_L—V is the value of l that minimizes the above distortion, the quantized LSF subvector is given by:
{circumflex over (λ)}(m)=V_L(l*_L—V,m−6)+λ_dc(m)+0.5({circumflex over (λ)}_prev(m)−λ_dc(m)), 6≦m≦9.   (35)

[0086] The stability of the quantized LSFs is checked by ensuring that the LSFs are monotonically increasing and are separated by a minimum value of about 0.008. If this criteria is not satisfied, stability is enforced by reordering the LSFs in a monotonically increasing order. If a minimum separation is not achieved, the most recent stable quantized LSF vector from a previous frame is substituted for the unstable LSF vector. The 6 4-bit SQ indices {l*_L—S—m, 0≦m≦5} and the 7-bit VQ index l*_L—V are transmitted to the decoder. Thus the LSFs are encoded using a total of 31 bits.

[0087] The inverse quantized LSFs are interpolated each subframe by preferably linear interpolation between the current LSFs {{circumflex over (λ)}(m), 0≦m≦10} and the previous LSFs {{circumflex over (λ)}_prev(m), 0≦m≦10}. The interpolated LSFs at each subframe are converted to LP parameters {â_m(l), 0≦m≦10, 1≦l≦8}.

[0088] The prediction residual signal for the current frame is computed using the noise reduced speech signal {s_nr(n)} and the interpolated LP parameters. Residual is computed from the midpoint of a subframe to the midpoint of the next subframe, using the interpolated LP parameters corresponding to the center of this interval. This ensures that the residual is computed using locally optimal LP parameters. The residual for the past data as shown in FIG. 2 is preserved from the previous frame and is also used for PW extraction.

[0089] Further, residual computation extends 93 samples into the look-ahead part of the buffer to facilitate PW extraction. LP parameters of the last subframe are used computing the look-ahead part of the residual. By denoting the interpolated LP parameters for the j^th subframe (0≦j≦8) of the current frame by {â_m(j), 0≦m≦10}, residual computation can be represented by: $e_{\mathrm{lp}}\left(n\right)=\{\begin{array}{cc}\sum _{m=0}^M{s}_{\mathrm{nr}}\left(n-m\right){\hat{a}}_m\left(0\right)& 80\le n<90,\\ \sum _{m=0}^M{s}_{\mathrm{nr}}\left(n-m\right){\hat{a}}_m\left(j\right)& 1\le j\le 720j+70\le n<20j+90,\\ \sum _{m=0}^M{s}_{\mathrm{nr}}\left(n-m\right){\hat{a}}_m\left(8\right)& 230\le n\le 332.\end{array}$
The residual for past data, {e_lp(n), 0≦n<80} is preserved from the previous frame.

[0090] The invention will now be discussed in reference to PW extraction. The prototype waveform in the time domain is essentially the waveform of a single pitch cycle, which contains information about the characteristics of the glottal excitation. A sequence of PWs contains information about the manner in which the excitation is changing across the frame. A time-domain PW is obtained for each subframe by extracting a pitch period long segment approximately centered at each subframe boundary. The segment is centered with an offset of up to ±10 samples relative to the subframe boundary, so that the segment edges occur at low energy regions of the pitch cycle. This minimizes discontinuities between adjacent PWs. For the m^th subframe, the following region of the residual waveform is considered to extract the PW: $\begin{array}{cc}\left\{e_{\mathrm{lp}}\left(80+20m+n\right),-\frac{p_m}{2}-12\le n\le \frac{p_m}{2}+12\right\}& \left(37\right)\end{array}$
where p_m is the interpolated pitch period (in samples) for the m^th subframe. The PW is selected from within the above region of the residual, so as to minimize the sum of the energies at the beginning and at the end of the PW. The energies are computed as sums of squares within a 5-point window centered at each end point of the PW, as the center of the PW ranges over the center offset of about ±10 samples: $\begin{array}{cc}{E}_{\mathrm{end}}\left(i\right)=\sum _{j=-2}^2{e}_{\mathrm{lp}}^2\left(80+20m-\frac{p_m}{2}+i+j\right)+\sum _{j=-2}^2{e}_{\mathrm{lp}}^2\left(80+20m+\frac{p_m}{2}+i+j\right)-10\le i\le 10.& \left(38\right)\end{array}$

[0091] The center offset resulting in the smallest energy sum determines the PW. If i_mm(m) is the center offset at which the segment end energy is minimized, i.e.,
E_end(i_min(m))≦E_end(i) −10≦i≦10,   (39)
the time-domain PW vector for the m^th subframe is $\left\{e_{\mathrm{lp}}\left(80+20m-\frac{p_m}{2}+i_{\min }\left(m\right)+n\right),0\le n<p_m\right\}.$
This is transformed by a p_m-point discrete Fourier transform (DFT) into a complex valued frequency-domain PW vector: $\begin{array}{cc}{P}_m^{\prime }\left(k\right)=\sum _{n=0}^{p_m-1}\begin{array}{cc}{e}_{\mathrm{lp}}\left(80+20m-\frac{p_m}{2}+i_{\min }\left(m\right)+n\right)e^{-j{\omega }_m\mathrm{kn}}& 0\le k\le K_m.\end{array}& \left(40\right)\end{array}$
Here ω_m is the radian pitch frequency and K_m is the highest in-band harmonic index for the m^th subframe (see equation 17). The frequency domain PW is used in all subsequent operations in the encoder. The above PW extraction process is carried out for each of the 8 subframes within the current frame, so that the residual signal in the current frame is characterized by the complex PW vector sequence {P′_m(k), 0≦k≦K_m, 1≦m≦8}. In addition, an approximate PW is computed for subframe 1 of the look ahead frame, to facilitate a 3-point smoothing of PW gain and magnitude. Since the pitch period is not available for the look-ahead part of the buffer, the pitch period at the end of the current frame, i.e., p₈, is used in extracting this PW. The region of the residual used to extract this extra PW is $\begin{array}{cc}\left\{e_{\mathrm{lp}}\left(260+n\right),-\frac{p_8}{2}-12\le n\le \frac{p_8}{2}+12\right\}.& \left(41\right)\end{array}$

[0092] By minimizing the end energy sum as before, the time-domain PW is obtained as $\left\{e_{\mathrm{lp}}\left(260-\frac{p_8}{2}+i_{\min }\left(9\right)+n\right),0\le n<p_8\right\}.$
The frequency-domain PW vector is designated by P₉ and is computed by the following DFT: $\begin{array}{cc}{P}_9^{\prime }\left(k\right)=\sum _{n=0}^{p_8-1}{e}_{\mathrm{lp}}\left(260-\frac{p_8}{2}+i_{\min }\left(9\right)+n\right)e^{-{\mathrm{j\omega }}_{8}k\pi }0\le k\le K_8.& \left(42\right)\end{array}$
It should be noted that the approximate PW is only used for smoothing operations and not as the PW for subframe 1 during the encoding of the next frame. Rather, it is replaced by the exact PW computed during the next frame.

[0093] Each complex PW vector can be further decomposed into a scalar gain component representing the level of the PW vector and a normalized complex PW vector representing the shape of the PW vector. Such a decomposition, permits vector quantization that is efficient in terms of computation and storage with minimal degradation in quantization performance. The PW gain is the root-mean square (RMS) value of the complex PW vector. It is obtained by $\begin{array}{cc}{g}_{\mathrm{pw}}^{\prime }\left(m\right)=\sqrt{\frac{1}{2K_m+2}\sum _{k=0}^{K_m}{P_m^{\prime }\left(k\right)}^2}1\le m\le 8.& \left(43\right)\end{array}$

[0094] PW gain is also computed for the extra PW by $\begin{array}{cc}{g}_{\mathrm{pw}}^{\prime }\left(9\right)=\sqrt{\frac{1}{2K_8+2}\sum _{k=0}^{K_8}{P_9^{\prime }\left(k\right)}^2}.& \left(44\right)\end{array}$

[0095] A normalized PW vector sequence is obtained by dividing the PW vectors by the corresponding gains: $\begin{array}{cc}{P}_m\left(k\right)=\frac{P_m^{\prime }\left(k\right)}{g_{\mathrm{pw}}^{\prime }\left(m\right)}0\le k\le K_m,1\le m\le 8.& \left(45\right)\end{array}$
And for the extra PW: $\begin{array}{cc}{P}_9\left(k\right)=\frac{P_9^{\prime }\left(k\right)}{g_{\mathrm{pw}}^{\prime }\left(9\right)}0\le k\le K_8.& \left(46\right)\end{array}$

[0096] For a majority of frames, especially during stationary intervals, gain values change slowly from one subframe to the next. This makes it possible to decimate the gain sequence by a factor of about 2, thereby reducing the number of values that need to be quantized. Prior to decimation, the gain sequence is smoothed by a 3-point window, to eliminate excessive variations across the frame. The smoothing operation is in the logarithmic gain domain and is represented by
g″_pw(m)=0.3 log₁₀ g′_pw(m−1)+0.4 log₁₀ g′_pw(m)+0.3 log₁₀ g′_pw(m+1) 1≦m≦8.   (47)

[0097] Conversion to logarithmic domain is advantageous since it corresponds to the scale of loudness of sound perceived by the human ear. The smoothed gain values are transformed by the following transformation: $\begin{array}{cc}{g}_{\mathrm{pw}}\left(m\right)=\{\begin{array}{ccc}0& g_{\mathrm{pw}}^{″}\left(m\right)>4.5,& \\ 90-20g_{\mathrm{pw}}^{″}\left(m\right)& 0\le g_{\mathrm{pw}}^{″}\left(m\right)\le 4.5,& 1\le m\le 8\\ 90& g_{\mathrm{pw}}^{″}\left(m\right)<0.& \end{array}& \left(48\right)\end{array}$

[0098] This transformation limits extreme (very low or very high) values of the gain and thereby improves quantizer performance, especially for low-level signals. The transformed gains are decimated by a factor of 2, requiring that only the even indexed values, i e., {g_pw(2), g_pw(4), g_pw(6), g_pw(8)}, are quantized.

[0099] At the decoder 100B, the odd indexed values are obtained by linearly interpolating between the inverse quantized even indexed values.

[0100] A 256 level, 4-dimensional vector quantizer is used to quantize the above gain vector. The design of the vector quantizer is one of the novel aspects of this algorithm. The PW gain sequence can exhibit two distinct modes of behavior. During stationary signals, such as voiced intervals, variations of the gain sequence across a frame are small.

[0101] On the other hand, during non-stationary signals such as voicing onsets, the gain sequence can exhibit large variations across a frame. The vector quantizer used must be able to represent both types of behavior. On the average, stationary frames far outnumber the non-stationary frames.

[0102] If a vector quantizer is trained using a database, which does not distinguish between the two types, the training is dominated by stationary frames leading to poor performance for non-stationary frames. To overcome this problem, the vector quantizer design was modified by classifying the PW gain vectors classified into a stationary class and a non-stationary class.

[0103] For the 256 level codebook, 192 levels were allocated to represent stationary frames and the remaining 64 were allocated for non-stationary frames. The 192 level codebook is trained using the stationary frames, and the 64 level codebook is trained using the non-stationary frames. The training algorithm with a binary split and random perturbation is based on the generalized Lloyd algorithm disclosed in “An algorithm for Vector Quantization Design”, by Y. Linde, A. Buzo and R. Gray, pages 84-95 of IEEE Transactions on Communications, VOL. COM-28, No. 1, January 1980 which is incorporated by reference in its entirety. In the case of the stationary codebook, a ternary split is used to derive the 192 level codebook from a 64 level codebook in the final stage of the training process. The 192 level codebook and the 64 level codebook are concatenated to obtain the 256-level gain codebook. The stationary/non-stationary classification is used only during the training phase. During quantization, stationary/non-stationary classification is not performed. Instead, the entire 256-level codebook is searched to locate the optimal quantized gain vector. The quantizer uses a mean squared error (MSE) distortion metric: $\begin{array}{cc}{D}_g\left(l\right)=\sum _{m=1}^4{\left[g_{\mathrm{pw}}\left(2m\right)-V_g\left(l,m\right)\right]}^{2}0\le l\le 255,& \left(49\right)\end{array}$
where, {V_g(l,m), 0≦l≦255, 1≦m≦4} is the 256 level, 4-dimensional gain codebook and D_g(l) is the MSE distortion for the l^th codevector. In another embodiment of the present invention the optimal codevector {V_g(l*_g,m), 1≦m≦4} is the one which minimizes the distortion measure over the entire codebook, i.e.,
D_g(l*_g)≦D_g(l) 0≦l≦255.   (50)
The 8-bit index of the optimal code-vector l*_g is transmitted to the decoder as the gain index.

[0104] FIG. 5 is a block diagram showing the separation of stationary and nonstationary components of a PW in accordance with an embodiment of the present invention and occurs in compute subband nonstationary measure module 116. In the FDI algorithm, only the PW magnitude information is explicitly encoded. PW Phase is not encoded explicitly since the replication of phase spectrum is not necessary for achieving a natural quality in reconstructed speech. However, this does not imply that an arbitrary phase spectrum can be employed at the decoder. One important requirement on the phase spectrum used at the decoder 100B is that it produces the correct degree of periodicity i.e., pitch cycle stationarity across the frequency band. Achieving the correct degree of periodicity is extremely important to reproduce natural sounding speech.

[0105] The generation of the phase spectrum at the decoder 100B is facilitated by measuring pitch cycle stationarity at the encoder as a ratio of the energy of the non-stationary component to that of the stationary component in the PW sequence. Further, this energy ratio is measured over 5 subbands spanning the frequency band of interest, resulting in a 5-dimensional vector nonstationarity measure in each frame. This vector is quantized and transmitted to the decoder, where it is used to generate phase spectra that lead to the correct degree of periodicity across the band. The first step in measuring the stationarity of PW is to align the PW sequence.

[0106] In order to measure the degree of stationarity of the PW sequence, it is necessary to align each PW to the preceding PW. The alignment process applies a circular shift to the pitch cycle to remove apparent differences in adjacent PWs that are due to temporal shifts or variations in pitch frequency. Let {tilde over (P)}_m−1 denote the aligned PW corresponding to subframe m−1 and let {tilde over (θ)}_m−1 be the phase shift that was applied to P_m−1 to derive {tilde over (P)}_m−1. In other words,
{tilde over (P)}_m−1(k)=P_m−1(k)e^{j{tilde over (θ)}m−1k} 0≦k≦K_m−1.   (51)

[0107] For the alignment of P_m to {tilde over (P)}_m−1, if the residual signal is perfectly periodic with the pitch period being an integer number of samples, P_m and P_m−1 are identical except for a circular shift. In this case, the pitch cycle for the m^th subframe is identical to the pitch cycle for the m−1^th subframe, except that the starting point for the former is at a later point in the pitch cycle compared to the latter. The difference in starting point arises due to the advance by a subframe interval and differences in center offsets at subframes m and m−1. With the subframe interval of 20 samples and with center offsets of i_mm(m) and i_mm(m−1), it can be seen that the m^th pitch cycle is ahead of the m−1^th pitch cycle by 20+i_min(m)−i_min(m−1) samples. If the pitch frequency is ω_m, a phase shift of −ω_m(20+i_min(m)−i_min(m−1)) is necessary to correct for this phase difference and align P_m with P_m−1. In addition, since P_m−1 has been circularly shifted by {tilde over (θ)}_m−1 to derive {tilde over (P)}_m−1, it follows that the phase shift needed to align P_m with {tilde over (P)}_m−1 is a sum of these two phase shifts and is given by
{tilde over (θ)}_m−1−ω_m(20+i_min(m)−i_min(m−1)).   (52)

[0108] In practice, the residual signal is not perfectly periodic and the pitch period can be non-integer valued. In such a case, the above cannot be used as the phase shift for optimal alignment. However, for quasi-periodic signals, the above phase angle can be used as a nominal shift and a small range of angles around this nominal shift angle are evaluated to find a locally optimal shift angle. Satisfactory results have been obtained with about an angle range of ±0.2π centered around the nominal shift angle, searched in steps of about 0.04π. For each shift within this range, the shifted version of P_m is correlated against {tilde over (P)}_m−1. The shift angle that results in the maximum correlation is selected as the locally optimal shift. This correlation maximization can be represented by $\underset{-5\le i\le 5}{\mathrm{MAX}}\sum _{k=0}^{K_m}\mathrm{Re}\left[{\tilde{P}}_{m-1}\left(k\right)P_m^{\prime }\left(k\right)e^{-j\left({\tilde{\theta }}_{m-1}-{\omega }_m\left(20+i_{\min }\left(m\right)-i_{\min }\left(m-1\right)\right)+004\mathrm{\pi i}\right)k}\right]$
where * represents complex conjugation and Re[ ] is the real part of a complex vector. If i=i_max maximizes the above correlation, then the locally optimal shift angle is
{tilde over (θ)}_m={tilde over (θ)}_m−1−ω_m(20+i_min(m)−i_min(m−1))+0.04πi_max   (54)
and the aligned PW for the m^th subframe is obtained from
 {tilde over (P)}_m(k)=P_m(k)e^{j{tilde over (θ)}mk} 0≦k≦K_m.   (55)

[0109] The process of alignment results in a sequence of aligned PWs from which any apparent dissimilarities due to shifts in the PW extraction window, pitch period etc. have been removed. Only dissimilarities due to the shape of the pitch cycle or equivalently the residual spectral characteristics are preserved. Thus, the sequence of aligned PWs provides a means of measuring the degree of change taking place in the residual spectral characteristics i.e., the degree of stationarity of the residual spectral characteristics. The basic premise of the FDI algorithm is that it is important to encode and reproduce the degree of stationarity of the residual in order to produce natural sounding speech at the decoder. Consider the temporal sequence of aligned PWs along the k^th harmonic track, i.e.,
{{tilde over (P)}_m(k), 1≦m≦8}.   (56)

[0110] If the signal is perfectly periodic, the k^th harmonic is identical for all subframes, and the above sequence is a constant as a function of m. If the signal is quasi-periodic, the sequence exhibits slow variations across the frame, but is still a predominantly low frequency waveform. It should be noted that here frequency refers to evolutionary frequency, related to the rate at which PW changes across a frame. This is in contrast to harmonic frequency, which is the frequency of the pitch harmonic. Thus, a high frequency harmonic component changing slowly across the frame is said to have low evolutionary frequency content. Or a low frequency harmonic component changing rapidly across the frame is said to have high evolutionary frequency content.

[0111] As the signal periodicity decreases, variations in the above PW sequence increase, with decreasing energy at lower frequencies and increasing energy at higher frequencies. At the other extreme, if the signal is aperiodic, the PW sequence exhibits large variations across the frame, with a near uniform energy distribution across frequency. Thus, by determining the spectral energy distribution of aligned PW sequences along a harmonic track, it is possible to obtain a measure of the periodicity of the signal at that harmonic frequency. By repeating this analysis at all the harmonics within the band of interest, a frequency dependent measure of periodicity can be constructed.

[0112] The relative distribution of spectral energy of variations of PW between low and high frequencies can be determined by passing the aligned PW sequence along each harmonic track through a low pass filter and a high pass filter. In an embodiment of the present invention, the low pass filter used is a 3^rd order chebyshev filter with a 3 dB cutoff at 35 Hz (for the PW sampling frequency of 400 Hz), with the following transfer function: $\begin{array}{cc}{H}_{\mathrm{lpf2}}\left(z\right)=\frac{0.063536-0.039167z^{-1}-0.039167z^{-2}+0.063536z^{-3}}{1-2.2255z^{-1}+1.7265z^{-2}+0.45231z^{-3}}.& \left(57\right)\end{array}$
The high pass filter used is also a 3^rd order chebyshev filter with a 3 dB cutoff at 18 Hz with the following transfer function: $\begin{array}{cc}{H}_{\mathrm{hpf2}}\left(z\right)=\frac{0.71923-2.1146z^{-1}+2.1146z^{-2}-0.71923z^{-3}}{1-2.2963z^{-1}+1.8542z^{-2}-5.1726z^{-3}}.& \left(58\right)\end{array}$

[0113] The output of the low pass filter is the stationary component of the PW that gives rise to pitch cycle periodicity and is denoted by {S_m(k), 0≦k≦K_m, 1≦m≦8}. The output of the high pass filter is the nonstationary component of PW that gives rise to pitch cycle aperiodicity and is denoted by {R_m(k), 0≦k≦K_m, 1≦m≦8}. The energies of these components are computed in subbands and then averaged across the frame.

[0114] The harmonics of the stationary and nonstationary components are grouped into 5 subbands spanning the frequency band of interest where the band-edges in Hz is defined by the array
B_rs=[1 400 800 1600 2400 3400].   (59)
The subband edges in Hz can be translated to subband edges in terms of harmonic indices such that the i^th subband contains harmonics with indices {η_m(i−1)≦k<η_m(i), 1≦i≦5} as follows: ${\eta }_m\left(i\right)=\left\{\begin{array}{cc}2+\lfloor \frac{B_{\mathrm{rs}}\left(i\right)K_m}{4000}\rfloor & \left\{1+\lfloor \frac{B_{\mathrm{rs}}\left(i\right)K_m}{4000}\rfloor \right\}<\frac{B_{\mathrm{rs}}\left(i\right)\pi }{4000{\omega }_m},\\ \lfloor \frac{B_{\mathrm{rs}}\left(i\right)K_m}{4000}\rfloor & \lfloor \frac{B_{\mathrm{rs}}\left(i\right)K_m}{4000}\rfloor >\frac{B_{\mathrm{rs}}\left(i\right)\pi }{4000{\omega }_m},\\ 1+\lfloor \frac{B_{\mathrm{rs}}\left(i\right)K_m}{4000}\rfloor & \mathrm{otherwise}.\end{array}\right\},0\le i\le 5,1\le m\le$
The energy in each subband is computed by averaging the squared magnitude of each harmonic within the subband. For the stationary component, the subband energy distribution for the m^th subframe is computed by $\begin{array}{cc}{\mathrm{ES}}_m\left(l\right)=\frac{1}{2\left({\eta }_m\left(l\right)-{\eta }_m\left(l-1\right)\right)}\sum _{k={\eta }_m\left(l-1\right)}^{{\eta }_m\left(l\right)-1}{S_m\left(k\right)}^{2}1\le l\le 5.& \left(61\right)\end{array}$
For the nonstationary component, the subband energy distribution for the m^th subframe is computed by $\begin{array}{cc}{\mathrm{ER}}_m\left(l\right)=\frac{1}{2\left({\eta }_m\left(l\right)-{\eta }_m\left(l-1\right)\right)}\sum _{k={\eta }_m\left(l-1\right)}^{{\eta }_m\left(l\right)-1}{R_m\left(k\right)}^{2}1\le l\le 5.& \left(62\right)\end{array}$
Next, these subframe energies are averaged across the frame: $\begin{array}{cc}{\mathrm{ES}}_{\mathrm{avg}}\left(l\right)=\frac{1}{8}\sum _{m=1}^8{\mathrm{ES}}_m\left(l\right),1\le l\le 5.& \left(63\right)\\ {\mathrm{ER}}_{\mathrm{avg}}\left(l\right)=\frac{1}{8}\sum _{m=1}^8{\mathrm{ER}}_m\left(l\right),1\le l\le 5.& \left(64\right)\end{array}$
The subband nonstationarity measure is computed as the ratio of the energy of the nonstationary component to that of the stationary component in each subband: $\begin{array}{cc}\left(l\right)=\frac{{\mathrm{ER}}_{\mathrm{avg}}\left(l\right)}{{\mathrm{ES}}_{\mathrm{avg}}\left(l\right)},1\le l\le 5.& \left(65\right)\end{array}$

[0115] If this ratio is very low, it indicates that the PW sequence has much higher energy at low evolutionary frequencies than at high evolutionary frequencies, corresponding to a predominantly periodic signal or stationary PW sequence. On the other hand, if this ratio is very high, it indicates that the PW sequence has much higher energy at high evolutionary frequencies than at low evolutionary frequencies, corresponding to a predominantly aperiodic signal or non stationary PW sequence. Intermediate values of the ratio indicate different mixtures of periodic and aperiodic components in the signal or different degrees of stationarity of the PW sequence. This information can be used at the decoder to create the correct degree of variation from one PW to the next, as a function of frequency and thereby realize the correct degree of periodicity in the signal.

[0116] In case of nonstationary voiced signals, where the pitch cycle is changing rapidly across the frame, the nonstationarity measure may have high values even in low frequency bands. This is usually a characteristic of unvoiced signals and usually translates to a noise-like excitation at the decoder. However, it is important that non-stationary voiced frames are reconstructed at the decoder with glottal pulse-like excitation rather than with noise-like excitation. This information is conveyed by a scalar parameter called a voicing measure, which is a measure of the degree of voicing of the frame. During stationary voiced and unvoiced frames, there is some correlation between the nonstationarity measure and the voicing measure. However, while the voicing measure indicates if the excitation pulse should be a glottal pulse or a noiselike waveform, the nonstationarity measure indicates how much this excitation pulse should change from subframe to subframe. The correlation between the voicing measure and the nonstationarity measure is exploited by vector quantizing these jointly.

[0117] The voicing measure is estimated for each frame based on certain characteristics correlated with the voiced/unvoiced nature of the frame. It is a heuristic measure that assigns a degree of voicing to each frame in the range 0-1, with a zero indicating a perfectly voiced frame and a one indicating a completely unvoiced frame.

[0118] The voicing measure is determined based on six measured characteristics of the current frame which are, the average of the nonstationarity measure in the 3 low frequency subbands, a relative signal power which is computed as the difference between the signal power of the current frame and a long term average signal power, the pitch gain, the average correlation between adjacent aligned PWs, the 1^st reflection coefficient obtained during LP Analysis, and the variance of the candidate pitch lags computed during pitch estimation.

[0119] The (squared) normalized correlation between the aligned PW of the m^th and m−1^th frames is obtained by $\begin{array}{cc}{\gamma }_m=\frac{{\left[\sum _{k=1}^6{\tilde{P}}_m\left(k\right){\tilde{P}}_{m-1}\left(k\right)\right]}^2}{\sum _{k=1}^6{{\tilde{P}}_m\left(k\right)}^2\sum _{k=1}^6{{\tilde{P}}_{m-1}\left(k\right)}^2}.& \left(66\right)\end{array}$

[0120] It should be noted that the upper limit of the summations are limited to 6 rather than K_m to reduce computational complexity. This subframe correlation is averaged across the frame to obtain an average PW correlation: $\begin{array}{cc}{\gamma }_{\mathrm{avg}}=\frac{1}{8}\sum _{m=1}^8{\gamma }_m.& \left(67\right)\end{array}$
The average PW correlation is a measure of pitch cycle to pitch cycle correlation after variations due to signal level, pitch period and PW extraction offset have been removed. It exhibits a strong correlation to the nature of glottal excitation. As mentioned earlier, the nonstationarity measure, especially in the low frequency subbands, has a strong correlation to the voicing of the frame. An average of the nonstationarity measure for the 3 lowest subbands provides a useful parameter in inferring the nature of the glottal excitation. This average is computed as $\begin{array}{cc}{\mathrm{avg}}_{}=\frac{1}{3}\sum _{l=1}^3{l}_{}.& \left(68\right)\end{array}$
It will be appreciated by those skilled in the art that subbands other than the three lowest subbands can be used without departing from the scope of the present invention.

[0121] The pitch gain is a parameter that is computed as part of the pitch analysis function. It is essentially the value of the peak of the autocorrelation function (ACF) of the residual signal at the pitch lag. To avoid spurious peaks, the ACF used in the embodiment of this invention is a composite autocorrelation function, computed as a weighted average of adjacent residual raw autocorrelation functions.

[0122] The pitch gain, denoted by β_pitch, is the value of the peak of a composite autocorrelation function. The composite ACF are evaluated once every 40 samples within each frame at 80, 120, 160, 200 and 240 samples as shown in FIG. 2. For each of the 5 ACF, the location of the peak ACF is selected as a candidate pitch period. The variation among these 5 candidate pitch lags is also a measure of the voicing of the frame. For unvoiced frames, these vales exhibit a higher variance than for voiced frames. The mean is computed as $\begin{array}{cc}{\mathrm{p\_cand}}_{\mathrm{avg}}=\frac{1}{5}\sum _{l=0}^4{\mathrm{p\_cand}}_l.& \left(69\right)\end{array}$
The variation is computed by the average of the absolute deviations from this mean: $\begin{array}{cc}{p}_{\mathrm{var}}=\frac{1}{5}\sum _{l=0}^4{\mathrm{p\_cand}}_{\mathrm{avg}}-{\mathrm{p\_cand}}_l.& \left(70\right)\end{array}$
This parameter exhibits a moderate degree of correlation to the voicing of the signal.

[0123] The signal power also exhibits a moderate degree of correlation to the voicing of the signal. However, it is important to use a relative signal power rather than an absolute signal power, to achieve robustness to input signal level deviations from nominal values. The signal power in dB is defined as $\begin{array}{cc}{E}_{\mathrm{sig}}=10{\log }_{10}\left[\frac{1}{160}\sum _{n=80}^{239}{s}^2\left(n\right)\right].& \left(71\right)\end{array}$

[0124] An average signal power can be obtained by exponentially averaging the signal power during active frames. Such an average can be computed recursively using the following equation:
E_sigavg=0.95E_sigavg+0.05E_sig.   (72)

[0125] A relative signal power can be obtained as the difference between the signal power and the average signal power:
E_sigrel=E_sig−E_sigavg.   (73)

[0126] The relative signal power measures the signal power of the frame relative a long term average. Voiced frames exhibit moderate to high values of relative signal power, whereas unvoiced frames exhibit low values.

[0127] The 1^st reflection coeffient ρ₁ is obtained as a byproduct of LP analysis during Levinson-Durbin recursion. Conceptually it is equalivalent to the 1^st order normalized autocorrelation coefficient of the noise reduced speech. During voiced speech segments, the speech spectrum tends to have a low pass characteristic, which results in a ρ₁ close to 1. During unvoiced frames, the speech spectrum tends to have a flatter or high pass characteristic, resulting in smaller or even negative values for ρ₁.

[0128] To derive the voicing measure, each of these six parameters are nonlinearly transformed using sigmoidal functions such that they map to the range 0-1, close to 0 for voiced frames and close to 1 for unvoiced frames. The parameters for the sigmoidal transformation have been selected based on an analysis of the distribution of these parameters. The following are the transformations for each of these parameters: $\begin{array}{cc}{n}_{\mathrm{pg}}=1-\frac{1}{\left(1+e^{-12\left({\beta }_{\mathrm{such}}-0.48\right)}\right)}& \left(74\right)\\ n_{\mathrm{pw}}=\{\begin{array}{c}1-\frac{1}{\left(1+e^{-10\left({\gamma }_{\mathrm{avg}}-0.72\right)}\right)}{\gamma }_{\mathrm{avg}}\le 0.72\\ 1-\frac{1}{\left(1+e^{-13\left({\gamma }_{\mathrm{avg}}-0.72\right)}\right)}{\gamma }_{\mathrm{avg}}>0.72\end{array}& \left(75\right)\\ n_{ℛ}=\{\begin{array}{c}\frac{1}{\left(1+e^{-7\left({ℛ}_{\mathrm{avg}}-0.85\right)}\right)}{\mathrm{avg}}_{}\le 0.85\\ \frac{1}{\left(1+e^{-3\left({ℛ}_{\mathrm{avg}}-0.72\right)}\right)}{\mathrm{avg}}_{}>0.85\end{array}& \left(76\right)\\ n_E=1-\frac{1}{\left(1+e^{-125\left(E_{\mathrm{signal}}-2\right)}\right)}& \left(77\right)\\ n_{\mathrm{pv}}=\{\begin{array}{cc}0.5-12.5\left(p_{\mathrm{var}}-0.02\right)& p_{\mathrm{var}}<0.02\\ 10\left(0.07-p_{\mathrm{var}}\right)& p_{\mathrm{var}}<0.07\\ 1& p_{\mathrm{var}}\ge 0.07\end{array}{n}_{\rho }=\{\begin{array}{c}1-\frac{1}{\left(1+e^{-5\left({\rho }_1-0.85\right)}\right)}{\rho }_1\le 0.85\\ 1-\frac{1}{\left(1+e^{-13\left({\rho }_1-0.85\right)}\right)}{\rho }_1>0.85\end{array}& \left(78\right)\end{array}$
The voicing measure of the previous frame ν_prev determines the weighted sum of the transformed parameters which results in the voicing measure: $\begin{array}{cc}v=\{\begin{array}{c}0.35n_{\mathrm{pg}}+0.225n_{\mathrm{pw}}+0.15n_{ℛ}+0.085n_E+\\ 0.07n_{\mathrm{pv}}+0.12n_{\rho }{v}_{\mathrm{prev}}<0.3\\ 0.35n_{\mathrm{pg}}+0.2n_{\mathrm{pw}}+0.1n_{ℛ}+0.1n_E+\\ 0.05n_{\mathrm{pv}}+0.2n_{\rho }{v}_{\mathrm{prev}}\ge 0.3.\end{array}& \left(79\right)\end{array}$

[0129] The weights used in the above sum are in accordence with the degree of correlation of the parameter to the voicing of the signal. Thus, the pitch gain receives the highest weight since it is most strongly correlated, followed by the PW correlation. The 1^st reflection coefficient and low-band nonstationarity measure receive moderate weights. The weights also depend on whether the previous frame was strongly voiced, in which case more weight is given to the low-band nonstationarity measure. The pitch variation and relative signal power receive smaller weights since they are only moderately correlated to voicing.

[0130] If the resulting voicing measure ν is clearly in the voiced region (ν<0.45) or clearly in the unvoiced region (ν>0.6), it is not modified further. However, if it lies outside the clearly voiced or unvoiced regions, the parameters are examined to determined if there is a moderate bias towards a voiced frame. In such a case, the voicing measure is modified so that its value lies in the voiced region.

[0131] The resulting voicing measure ν takes on values in the range 0-1, with lower values for more voiced signals. In addition, a binary voicing measure flag is derived from the voicing measure as follows: $\begin{array}{cc}{v}_{\mathrm{flag}}=\{\begin{array}{cc}0& v\le 0.45,\\ 1& v>0.45.\end{array}& \left(80\right)\end{array}$

[0132] Thus, ν_flag is 0 for voiced signals and 1 for unvoiced signals. This flag is used in selecting the quantization mode for PW magnitude and the subband nonstationarity vector. The voicing measure ν is concatenated to the subband nonstationarity measure vector and the resulting 6-dimensional vector is vector quantized.

[0133] The subband nonstationarity measure can have occasional spurious large values, mainly due to the approximations and the averaging used during its computation. If this occurs during voiced frames, the signal is reproduced with excessive roughness and the voice quality is degraded. To prevent this, large values of the nonstationarity measure are attenuated. The attenuation charactersitic has been determined experimentally and is specified as follows for each of the five subbands: $\begin{array}{cc}\left(1\right)\Leftarrow \{\begin{array}{cc}\left(1\right)& v>0.6\mathrm{or}\left(1\right)\le 0.3+\\ & 0.1667v\\ 0.05+0.1667v+& v\le 0.6\mathrm{and}\left(1\right)>\\ \frac{0.5}{\left(1+e^{-5\left(ℛ\left(1\right)-0.3-0.1667v\right)}\right)}& 0.3+0.1667v\end{array}& \left(81\right)\\ \left(2\right)\Leftarrow \{\begin{array}{cc}\left(2\right)& v>0.6\mathrm{or}\left(2\right)\le 0.45+\\ & 0.1667v\\ 0.2+0.0833v+& v\le 0.6\mathrm{and}\left(2\right)>\\ \frac{0.5+0.1667v}{\left(1+e^{-5\left(ℛ\left(2\right)-0.45-0.1667v\right)}\right)}& 0.45+0.1667v\end{array}& \left(82\right)\\ \left(3\right)\Leftarrow \{\begin{array}{cc}\left(3\right)& v>0.6\mathrm{or}\left(3\right)\le 0.5+\\ & 0.5v\\ 0.1+0.5v+& v\le 0.6\mathrm{and}\left(3\right)>\\ \frac{0.8}{\left(1+e^{-5\left(ℛ\left(3\right)-0.5-0.5v\right)}\right)}& 0.5+0.5v\end{array}& \left(83\right)\\ \left(4\right)\Leftarrow \{\begin{array}{cc}\left(4\right)& v>0.6\mathrm{or}\left(4\right)\le 0.65+\\ & 0.5833v\\ 0.3+0.333v+& v\le 0.6\mathrm{and}\left(4\right)>\\ \frac{0.7+0.5v}{\left(1+e^{-5\left(ℛ\left(4\right)-0.3-0.3333v\right)}\right)}& 0.65+0.5833v\end{array}& \left(84\right)\\ \left(5\right)\Leftarrow \{\begin{array}{cc}\left(5\right)& v>0.6\mathrm{or}\left(5\right)\le 0.65+\\ & 0.5833v\\ 0.3+0.333v+& v\le 0.6\mathrm{and}\left(5\right)>\\ \frac{0.7+0.5v}{\left(1+e^{-5\left(ℛ\left(5\right)-0.3-0.3333v\right)}\right)}& 0.65+0.5833v\end{array}& \left(85\right)\end{array}$

[0134] Additionaly, for voiced frames, it is necessary to ensure that the values of the nonstationarity measure in the low frequency subbands are in a monotonically nondecreasing order. This condition is enforced for the 3 lower subbands according to the flow chart in FIG. 6.

[0135] FIG. 6 is a flow chart depicting a method 600 for enforcing monotonic measures in accordance with an embodiment of the present invention. The method 600 occurs in compute subband nonstationary measure module 116 and is initiated at step 602 where the adjustment for the R vector is begun. The method 600 then proceeds to step 604.

[0136] At step 604 a determination is made as to whether the voicing measure is less than 0.6. If the determination is answered negatively, the method proceeds to step 622. If the determination is answered affirmatively the method proceeds to step 606.

[0137] At step 606 a determination is made as to whether R1 is greater than R2. If the determination is answered negatively, the method proceeds to step 614. If the determination is answered affirmatively, the method proceeds to step 608.

[0138] At step 614 a determination is made as to whether R2 is greater than R3. If the determination is answered negatively the method proceeds to step 622. If the determination is answered affirmatively, the method proceeds to step 616.

[0139] At step 608 a determination is made as to whether 0.5(R1+R2) is less than or equal to R3. If the determination is answered affirmatively the method proceeds to step 610 where a formula is used to calculate R1 and R2. The method then proceeds to step 614.

[0140] If the determination at step 608 is answered negatively, the method proceeds to step 612 where a series of calculations is used to calculate R1, R2 and R3. The method then proceeds to step 614.

[0141] At step 616 a determination is made as to whether 0.5(R2+R3) is greater than or equal to R1. If the determination is answered affirmatively, the method proceeds to step 618 where a series of calculations is used to calculate R2 and R3. If the method is answered negatively, the method proceeds to step 620 where a series of calculations is used to calculate R1, R2 and R3.

[0142] The steps 614, 618 and 620 proceed to step 622 where the adjustment of the R vector ends.

[0143] The nonstationarity measure vector is vector quantized using a spectrally weighted quantization. The spectral weights are derived from the LPC parameters. First, the LPC spectral estimate corresponding to the end point of the current frame is estimated at the pitch harmonic frequencies. This estimate employs tilt correction and a slight degree of bandwidth broadening. These measures are needed to ensure that the quantization of formant valleys or high frequencies are not compromised by attaching excessive weight to formant regions or low frequencies. $\begin{array}{cc}{W}_8\left(k\right)=\frac{{\sum _{m=0}^{10}{a}_8^{\prime }\left(i\right){0.4}^m{e}^{j{w}_8\mathrm{k8}}}^2}{{\sum _{m=0}^{10}{a}_8^{\prime }\left(i\right){0.98}^m{e}^{-j{w}_8\mathrm{k8}}}^2}0\le k\le K_8.& \left(86\right)\end{array}$

[0144] This harmonic spectrum is converted to a subband spectrum by averaging across the 5 subbands used for the computation of the nonstationarity measure. $\begin{array}{cc}{\stackrel{\_}{W}}_8\left(l\right)=\frac{1}{\left({\eta }_8\left(l\right)-{\eta }_8\left(l-1\right)\right)}\sum _{k={\eta }_8\left(l-1\right)}^{{\eta }_8\left(l\right)-1}{W}_8\left(k\right)1\le l\le 5.& \left(87\right)\end{array}$

[0145] This is averaged with the subband spectrum at the end of the previous frame to derive a subband spectrum that corresponding to the center of the current frame. This average serves as the spectral weight vector for the quantization of the nonstationarity vector.
{overscore (W)}₄(l)=0.5({overscore (W)}₀(l)+{overscore (W)}₈(l)) 1≦l≦5.   (88)

[0146] The voicing measure is concatenated to the end of the nonstationarity measure vector, resulting in a 6-dimensional composite vector. This permits the exploitation of the considerable correlation that exists between these quantities. The composite vector is denoted by
_c={(1) (2) (3) (4) (5) ν}  (89)

[0147] The spectral weight for the voicing measure is derived from the spectral weight for the nonstationarity measure depending on the voicing measure flag. If the frame is voiced (ν_flag=0), the weight is computed as $\begin{array}{cc}{\stackrel{\_}{W}}_4\left(6\right)=\frac{0.33}{5}\sum _{l=1}^5{W}_4\left(l\right)v_{\mathrm{flag}}=0.& \left(90\right)\end{array}$

[0148] In other words, it is lower than the average weight for the nonstationary component. This ensures that that the nonstationary component is quantized more accurately than the voicing measure. This is desirable since for voiced frames, it is important to preserve the nonstationarity in the various bands to achieve the right degree of periodicty. On the other hand, for unvoiced frames, voicing measure is more important. In this case, its weight is larger than the maximum weight for the nonstationary component. $\begin{array}{cc}{\stackrel{\_}{W}}_4\left(6\right)=1.5\underset{1\le l\le 5}{\mathrm{MAX}}{W}_4\left(l\right)v_{\mathrm{flag}}=1.& \left(91\right)\end{array}$

[0149] A 64 level, 6-dimensional vector quantizer is used to quantize the composite nonstationarity measure-voicing measure vector. The first 8 codevectors (indices 0-7) assigned to represent unvoiced frames and the remaining 56 codevectors (indices 8-63) are assigned to respresent voiced frames. The voiced/unvoiced decision is made based on the voicing measure flag. The following weighted MSE distortion measure is used: $\begin{array}{cc}{D}_R\left(l\right)=\sum _{m=1}^6{{\stackrel{\_}{W}}_4\left(m\right)\left[c_{}\left(m\right)-V_R\left(l,m\right)\right]}^{2}0\le l\le 63,& \left(92\right)\end{array}$

[0150] Here, {V_R(l,m), 0≦l≦63, 1≦m≦6} is the 64 level, 6-dimensional composite nonstationarity measure-voicing measure codebook and D_R(l) is the weighted MSE distortion for the l^th codevector. If the frame is unvoiced (ν_flag=1), this distortion is minimized over the indices 0-7. If the frame is voiced (ν_flag=0), the distortion is minimized over the indices 8-63. Thus, $\begin{array}{cc}{D}_R^{mm}=\{\begin{array}{cc}\underset{0\le l\le 7}{\mathrm{MIN}}{D}_R\left(l\right)& v_{\mathrm{flag}}=1\\ \underset{8\le l\le 63}{\mathrm{MIN}}{D}_R\left(l\right)& v_{\mathrm{flag}}=0\end{array}& \left(93\right)\end{array}$

[0151] This partitioning of the codebook reflects the higher importance given to the representation of the nonstationarity measure during voiced frames. The 6-bit index of the optimal codevector l*_R is transmitted to the decoder as the nonstationarity measure index. It should be noted that the voicing measure flag, which is used in the decoder 100B for the inverse quantization of the PW magnitude vector, can be detected by examining the value of this index.

[0152] Up to this point, the PW vectors are processed in Cartesian (i.e., real-imaginary) form. The FDI codec 100 at 4.0 kbit/s encodes only the PW magnitude information to make the most efficient use of available bits. PW phase spectra are not encoded explicitly. Further, in order to avoid the computation intensive square-root operation in computing the magnitude of a complex number, the PW magnitude-squared vector is used during the quantization process.

[0153] The PW magnitude vector is quantized using a hierarchical approach, which allows the use of fixed dimension VQ with a moderate number of levels and precise quantization of perceptually important components of the magnitude spectrum. In this approach, the PW magnitude is viewed as the sum of two components: a PW mean component, which is obtained by averaging the PW magnitude across frequencies within a 7 band sub-band structure, and a PW deviation component, which is the difference between the PW magnitude and the PW mean. The PW mean component captures the average level of the PW magnitude across frequency, which is important to preserve during encoding. The PW deviation contains the finer structure of the PW magnitude spectrum and is not important at all frequencies. It is only necessary to preserve the PW deviation at a small set of perceptually important frequencies. The remaining elements of PW deviation can be discarded, leading to a small, fixed dimensionality of the PW deviation component.

[0154] The PW magnitude vector is quantized differently for voiced and unvoiced frames as determined by the voicing measure flag. Since the quantization index of the nonstationarity measure is determined by the voicing measure flag, the PW magnitude quantization mode information is conveyed without any additional overhead.

[0155] During voiced frames, the spectral characteristics of the residual are relatively stationary. Since the PW mean component is almost constant across the frame, it is adequate to transmit it once per frame. The PW deviation is transmitted twice per frame, at the 4^th and 8^th subframes. Further, interframe predictive quantization can be used in the voiced mode. On the other hand, unvoiced frames tend to be nonstationary. To track the variations in PW spectra, both mean and deviation components are transmitted twice per frame, at the 4^th and 8^th subframes. Prediction is not employed in the unvoiced mode.

[0156] The PW magnitude vectors at subframes 4 and 8 are smoothed by a 3-point window. This smoothing can be viewed as an approximate form of decimation filtering to down sample the PW vector from 8 vectors/frame to 2 vectors/frame.
{overscore (P)}_m(k)=0.3P_m−1(k)+0.4P_m(k)+0.3P_m−1(k), 0≦k≦K_m, m=4, 8.   (94)

[0157] The subband mean vector is computed by averaging the PW magnitude vector across 7 subbands. The subband edges in Hz are
B_pw=[1 400 800 1200 1600 2000 2600 3400]  (95)

[0158] To average the PW vector across frequencies, it is necessary to translate the subband edges in Hz to subband edges in terms of harmonic indices. The band-edges in terms of harmonic indices for subframes 4 and 8 can be computed by $\begin{array}{cc}{\kappa }_m\left(i\right)=\left\{\begin{array}{cc}2+\lfloor \frac{B_{\mathrm{pw}}\left(i\right)K_m}{4000}\rfloor & \left\{1+\lfloor \frac{B_{\mathrm{pw}}\left(i\right)K_m}{4000}\rfloor \right\}<\frac{B_{\mathrm{pw}}\left(i\right)\pi }{4000{\omega }_m},\\ \lfloor \frac{B_{\mathrm{pw}}\left(i\right)K_m}{4000}\rfloor & \lfloor \frac{B_{\mathrm{pw}}\left(i\right)K_m}{4000}\rfloor >\frac{B_{\mathrm{pw}}\left(i\right)\pi }{4000{\omega }_m},\\ 1+\lfloor \frac{B_{\mathrm{pw}}\left(i\right)K_m}{4000}\rfloor & \mathrm{otherwise}.\end{array}\right\},0\le i\le 7,m=4,8.& \left(96\right)\end{array}$