JPH0795237B2 - Adaptive Multivariate Estimator - Google Patents

Description

TECHNICAL FIELD The present invention relates to classifying samples representing a real-time process into groups, each corresponding to a state of the real-time process, and more particularly, this classification is performed in real time using statistical techniques as each sample is generated.

BACKGROUND ART AND PROBLEM In many real-time processes, problems exist when attempting to estimate the current state of a changing environment from current and past samples of the process. One example of such a process is the production of speech by the human vocal tract. The sound produced by the vocal tract may have a fundamental frequency (voiced state) or may not have a fundamental frequency (unvoiced state). In addition, a third state may exist if no sound is produced (silence state). The problem of distinguishing between these three states is known as speech/silence discrimination. In low bit rate speech coders,
Inaccurate speech discrimination results in a degradation of speech quality. The difficulty in making these speech discriminations accurately lies in the fact that no single speech parameter, or classifier, can reliably distinguish between voiced and unvoiced speech. Combining multiple speech classifiers in a weighted sum fashion to make speech discrimination is well known to those skilled in the art. Such a method is described in D.P. Prezas et al., "Fast and Accurate Pitch Detection Using Pattern Recognition and Adaptive Time-Domain Analysis," Proceedings of the IEEE International Conference on Acoustics, Speech Synthesis and Signal Processing, Vol. 1, pp. 109-112, 1986.
April 2015 (“Fast and Accurate Pitch Detection Using
Pattern Recognition and Adaptive Time-Domain Analysis
sis", Proc. IEEE Int. Conf. Acoust.,Speech and Signal
Proc., Vol. 1, pp. 109-112, April 1986. As explained in that paper, if the weighted sum of the speech classifiers is greater than a certain threshold, the speech frame is declared voiced; if not, it is declared unvoiced. Mathematically, this relationship is expressed as a'X + b > 0, where "a" is a vector of weights, "X" is a vector of classifiers, and "b" is a scalar representing the threshold. The weights are selected to maximize performance on a training set of speech, where the voicing of each frame is known. These weights form a decision rule within the speech coder that provides significant speech quality improvements compared to those using a single parameter.

A problem with fixed weighted sum methods is that they do not perform well in changing speech environments, such as the result of telephone conversations conducted on a car or mobile phone, or perhaps due to different types of telephone microphones. The reason fixed weighted sum methods do not perform well in changing environments is that many speech classifiers are affected by background noise, nonlinear distortion, and filtering. If speechifications must be discriminated against speech with characteristics different from those of the training set, the weights will generally not provide satisfactory results.

One method for adapting the fixed weighted sum method to changing speech environments is proposed by C.P. Campbell
l) Another paper, "US Government LPC for Voiced/Unvoiced Classification of Speech"
-10E algorithm application," Proceedings of the IEEE International Conference on Acoustics, Sound Generation and Signal Processing, Tokyo, 1986, Vol. 9.11.4, pp. 473-
Page 476 (“Voiced/Unvoiced Classification of Spe.
ech with Application to the US Government LPC-10
E Algorithm", IEEE International Conference on Aco
ustics,Speech and Signal Processing,1986,Tokyo,Vo
1.9.11.4, pp.473-476). This paper proposes a method for improving the performance of a neural network by adding different levels of white noise to the training data for each set of weights and thresholds.
This paper discloses the use of different sets of weights and thresholds predefined from the same set of training data. For each frame, speech samples are processed with a set of weights and thresholds, after which one of these sets is selected based on the signal-to-noise ratio (SNR). The range of possible SNR values is divided into subranges, each of which is assigned to one of the sets. The SNR is calculated for each frame; the subrange is determined; and the frame is then classified as voiced or unvoiced. The problem with this method is that it is only effective when white noise is added to the training data and is not adaptable to a wide range of speech environments and speakers. Therefore, there is a need for a speech detector that can reliably determine whether speech is voiced or unvoiced for varying environments and different speakers.

Solution The above problems are solved and a technical advance is achieved by an apparatus that determines statistical distributions for a plurality of process states in response to real-time samples from a physical process and establishes discriminant regions from these distributions. The latter regions are used to determine the current process state at the time each process sample is generated. When used to make speech discrimination, the apparatus adapts to changing speech environments by using the state of a speech classifier. Statistical techniques are based on the classifiers and used to modify the discriminant regions used for speech discrimination. The apparatus preferably estimates statistical distributions for both voiced and unvoiced frames and uses these statistical distributions to determine the discriminant regions. The latter regions are then used to discriminate whether the current speech frame is voiced or unvoiced.

The voiced speech detector preferably calculates the probability that the current speech frame is unvoiced, the probability that the current speech frame is voiced, and the overall probability that a frame will be unvoiced. Using these three probabilities, the detector then calculates a probability distribution for unvoiced frames and a probability distribution for voiced frames. Furthermore, the calculation to determine the probability that the current speech frame is voiced or unvoiced is performed using a maximum likelihood statistical method, which is also responsive to a weight vector and a threshold value in addition to the probability. In another embodiment, the weight vector and threshold value are adaptively calculated for each frame. This adaptive calculation of the weight vector and threshold value allows the detector to quickly adapt to changing speech environments.

The apparatus for determining the presence of a fundamental frequency within a speech frame preferably includes a circuit for calculating a set of statistical parameters in response to a set of classifiers representing speech attributes of the speech frame, a second circuit for calculating a set of weights each associated with one of the classifiers in response to a set of parameters defining a statistical distribution, and finally a third circuit for determining the presence of a fundamental frequency within the speech frame in response to the calculated set of weights, the classifiers, and the set of parameters, or, in conventional terms, making an unvoiced/voiced decision.

The second circuit also preferably calculates a threshold value and a new weight vector and communicates these values to the first circuit, which in response to these values and the new set of classifiers determines another set of statistical parameters which are then used to determine the presence of a fundamental frequency for the next speech frame.

The first circuit preferably responds to the next set of classifiers, the new weight vector, and the threshold value to calculate the probability that the next frame will be unvoiced, the probability that the next frame will be voiced, and the overall probability that a frame will be unvoiced. These probabilities are then used to determine another set of statistical parameters, along with a set of values giving the average of the classifiers for the past and current frames.

The method for determining a speech decision comprises the steps of: estimating a statistical distribution for voiced and unvoiced frames; determining a decision region representing voiced and unvoiced speech in response to the statistical distribution; and making a speech decision in response to the decision region and the current speech frame. The statistical distribution further comprises a probability that the current speech frame is unvoiced,
The probability that the current speech frame is voiced and the overall probability that a frame will be unvoiced are calculated. These three probabilities are calculated as substeps of the step of determining the statistical distribution.

BRIEF DESCRIPTION OF THE DRAWINGS The present invention will be readily understood by reading the following detailed description in conjunction with the drawings in which: FIG. 1 is a block diagram of an apparatus employing the present invention; FIG. 2 is a block diagram representation of the present invention; FIGS. 3 and 4 are block diagrams of the statistical voiced speech detector 103 of FIG. 2;
FIG. 5 is a diagram illustrating in more detail the functions performed by block 340 of FIG.

DETAILED DESCRIPTION Figure 1 shows an apparatus for performing unvoiced/voiced discrimination operations using the statistical voiced detector that is the subject of this invention as one of the voiced detectors. The apparatus of Figure 1 uses two types of detectors: a discriminative voiced detector and a statistical voiced detector. The statistical voiced detector 103 detects changes in the audio environment and modifies the weights used to process the classifiers coming from the classifier generator 101 to more accurately classify unvoiced/voiced sounds.
The discriminative voiced sound detector 102 is an adaptive detector that enables voiced sound discrimination. The discriminative voiced sound detector 102 is used during initial start-up, i.e., in rapidly changing speech environment conditions, when the statistical voiced sound detector 103 has not yet fully adapted to the initial or new speech environment.

Consider now the overall operation of the apparatus shown in Figure 1. A classifier generator 101 is responsive to each speech frame to generate classifiers which are preferably the log of the speech energy, the log of the LPC (Linear Predictive Distribution) gain, the log-area ratio of the first reflection coefficients, and the squared correlation coefficient of two speech segments one frame long offset by one pitch period. The calculation of these classifiers is as follows:
This involves digitally sampling analog audio, forming frames of digital samples, and processing these frames, as is well known to those skilled in the art. Generator 101 transmits classifiers via path 106 to detectors 102 and 103.

Detectors 102 and 103 make unvoiced/voiced decisions in response to the classifiers received via path 106 and transmit these decisions via paths 107 and 110, respectively, to multiplexer 105.
These detectors also determine distance measures between the voiced and unvoiced frames and transmit these measures to comparator 104 via paths 108 and 109.
These distances are preferably Mahalanobis distances or other generalized distances. Comparator 104 controls multiplexer 105 in response to the distances received via paths 108 and 109, so that the latter multiplexer selects the detector output producing the maximum distance.

FIG. 2 shows the statistical voiced speech detector 103 in more detail.
For each speech frame, a set of classifiers, also called a vector of classifiers, is received from classifier generator 101 via path 106. Silence detector 201 is responsive to these classifiers to determine whether speech is present in the frame. If speech is present, detector 201 transmits a signal via path 210. If speech is not present in the frame (silence), then only do subtractor 207 and U/V (unvoiced/voiced) classifier 205 operate for that particular frame. As to whether speech is present, classifier 20
5, the unvoiced/voiced sound is discriminated for each frame.

Classifier averager 202, in response to signals from detector 201, maintains an average of the individual classifiers received via path 106 by averaging the classifier for the current frame with the classifier for the previous frame. If speech (not silence) is present in the frame, silence detector 201 sends a signal via path 210 to statistics calculator 203, generator 206, and averager 202.

Statistics calculator 203 calculates statistical distributions for voiced and unvoiced frames. In particular, calculator 203 calculates the overall probability that a frame is unvoiced and the probability that a frame is voiced in response to signals received via path 210. Statistical calculator 203 also calculates the statistic that each classifier would have if the frame were unvoiced and the statistic that each classifier would have if the frame were voiced. Calculator 203 also calculates the covariance matrix of the classifiers. The statistic is preferably a mean value.
The calculations made by 203 are based not only on the current frame but also on previous frames. Statistical calculator 203 bases these calculations not only on the classifiers for the current frame received via path 106 and the average of the classifiers received via path 211, but also on weights for each classifier and a threshold value received via path 213 from weight calculator 204 which determines whether the frame is unvoiced or voiced.

Weight calculator 204 recalculates the weight vector a for each classifier and the threshold value b for the current frame in response to the classifier probabilities, covariance matrices, and statistics for the current frame generated by calculator 203 and received over path 212. These new a and b values are then transmitted back to statistics calculator 203 over path 213.

Weight calculator 204 also transmits weights and statistics for classifiers in both the unvoiced and voiced domains to discriminator 205 via path 214 and to generator 206 via path 208.
6. The latter generator responds to this information by calculating a distance measure which is then transmitted to comparator 104 via path 109 as shown in FIG.

A U/V (unvoiced/voiced) discriminator 205 determines whether the frame is unvoiced or voiced in response to the information transmitted over paths 214 and 215 and transmits this discriminator over path 110 to multiplexer 105 of FIG. 1.

2, where the operation of each block is given in more detail in vector and matrix mathematics. The meaner 202, statistics calculator 203, and weight calculator 204 are based on the method described in "Estimating the Components of a Gaussian Mixture" by N.E. Day.
We implement a modified EM algorithm similar to that described in the paper entitled "Decaying Means of a Mixture of Normal Distributions," Biometrika, Vol. 56, No. 3, pp. 463-474, 1969.
Using the concept of a classifier averager, classifier averager 202 calculates the average of the classifiers for the current frame and previous frames by calculating Equations 1, 2, and 3 below.

n = n + 1 if n < 2000 (1) z = = 1/n (2) X _n = (1 - z) X _n-1 + zx _n (3) x _n is a vector indicating the classifier for the current frame,
n is the number of frames to be processed up to 2000. z represents the average coefficient of the broken down data, and _Xn represents the average of all classifiers for the current frame and past frames. The statistical calculator 203 calculates z, _xn
In response to receiving the and X _n information, the covariance matrix T is calculated by first calculating a matrix of sums of squares and products Q _n as follows:

_Qn = (1 - z)Qn _-1 + _zxnx'n (4) Once _Qn is _calculated , T is calculated as follows:

T = _Qn - _{Xn X'n} . (5) The mean is subtracted from the classifier as follows:

_xn = _xn - _Xn (6) Next, the calculator 203 determines the probability that the frame represented by the current vector _Xn is unvoiced by solving the following equation (7): where the components of vector a correspond to the logarithm of the speech energy, 0.3918606, and the LPC
Preferably, the component corresponding to the logarithm of the gain is initialized equal to −0.0520902, the component corresponding to the logarithm of the area ratio of the first reflection coefficient is initialized equal to 0.5637082, and the component corresponding to the squared correlation coefficient is initialized equal to 1.361249, and b is initially initialized equal to −8.36454.

After solving equation (7), calculator 203 determines the probability that the classifier represents a voiced frame by solving the equation:

P(v| _xn )=1-P(u| _xn ) (8) Calculator 203 then determines the overall probability that a frame will be unvoiced by solving equation (9) for _pn .

_pn = (1 - z)pn _-1 + zP(u| _xn ). (9) After determining the probability that the frame will be unvoiced,
Next, calculator 203 determines two vectors u and v that give the mean values of each classifier for both unvoiced and voiced frames. Vectors u and v are the statistical means for unvoiced and voiced frames, respectively. Vector u, which is the statistical mean unvoiced vector,
contains the mean value of each classifier if the frame is unvoiced; and the vector v, the statistical mean voiced vector, gives the mean value for each classifier if the frame is voiced. As shown below, the vector u for the current frame is solved by computing equation (10), and the vector v for the current frame is determined by computing equation (11).

u _n = (1-z)u _n-1 +zx _n P(u|x _n )/p _n -zx _n (10) v _n = (1-z)v _n-1 +zx _n P(v|x _n )/(1-p _n )-zx _n
(11) where calculator 203 transmits vectors u and v, matrix T, and probability p to weight calculator 204 via path 212. Weight calculator 204 responds to this information by calculating new values for vector a and scalar b. These new values are then transmitted via path 213 to statistical calculator 203.
This allows the detector 103 to quickly adapt to changing environments.
Even if new values for a are not transmitted back to the statistical calculator 203, the detector 103 will continue to adapt to the changing environment because the vectors u and v are kept up to date. As will be apparent, the classifier 205 uses the vectors u and v as well as the vector a and scalar b to make a speech classification.
When n is preferably greater than 99, the vector a and scalar b are calculated as follows: Vector a is determined by solving the equation:

The scalar b is determined by solving the following equation:

After calculating equations (12) and (13), weight calculator 204
transmits the vectors a, u, and v to the U/V discriminator 205 via path 214. If the frame contains silence, only equation (6) is calculated.

In response to this transmitted information, the discriminator 205 determines whether the current frame is voiced or unvoiced. If the component of the vector ( _vn -u _n ) corresponding to the output is positive, then the frame is declared to be voiced if the following expression is true:

a'xn _- a'(u _n +v _n )/2>0; (14) or if the component of the vector (v _n -u _n ) corresponding to the output is negative, then the frame is declared to be voiced if the following is true:

a'xn _- a'(u _n +v _n )/2 < 0. (15) Equation (14) can also be rewritten as follows:

_a'xn +b-log[(1- _pn )/ _pn ]>0 Equation (15) can also be rewritten as follows:

_a'xn + b-log[(1- _pn )/ _pn ]>0 If the above condition is not met, the classifier 205 declares the frame to be unvoiced.
5) represents the decision region for voiced discrimination. (14)
and the log term in the rewritten form of (15) can be omitted with minor performance changes. In this embodiment, the component corresponding to the output is preferably the log of the speech energy.

Generator 206 responds to information received from calculator 204 via path 214 to calculate a distance measure A as follows: First, a discrimination variable d is calculated according to equation (16) as follows: d = ∑ i ⁢ ...

d= _a'xn +b-log[(1- _pn )/ _pn ]. (16) Various types of speech detectors may be used, as would be apparent to one skilled in the art, to generate values similar to d for use in the following equations. One such detector would be an autocorrelation detector. If the frame is voiced, then equations (17) through (20) are solved as follows:

m ₁ = (1 − z) m ₁ + zd, (17) s ₁ = (1 − z) s ₁ + zd ² , (18) k ₁ = s ₁ − ▲m ² ₁ ▼ (19) where m ₁ is the mean for the voiced frames and k ₁ is the variance for the voiced frames.

The probability P _d that the classifier 205 will declare a frame to be unvoiced is calculated as follows:

_Pd = (1 - z) _Pd (20) Preferably, _Pd is initially set to 0.5.

If the frame is unvoiced, then equations (21) through (24) are solved as follows:

_m0 = (1 - z) _m0 + zd, (21) _s0 = (1 - z) _s0 + ^zd2 , (22) _k0 = _s0 - ^m20 , (23) The probability _Pd that the classifier 205 will declare a frame _to be unvoiced is calculated as follows:

_Pd = (1 - z) _Pd + z. (24) After calculating equations (16) through (22), a distance measure or merit value is calculated as follows: Pd = (1 - z) Pd + z.

Equation (25) is different from Equation (25), which calculates the distance measure using Hotelling's two-sample ^T2 statistic, in that the larger the merit value, the greater the separation. However, other merit values exist where the smaller the merit value, the greater the separation. Preferably, the distance measure may be the Mahalanobis distance, given by

Preferably, the third method is given by:

Preferably, a fourth method for calculating the distance measure is given by:

A ² =a'(v _n -u _n ) (28) The decision detector 102 performs the unvoiced/voiced decision by transmitting information indicating a voiced frame to the multiplexer 105 via path 107 if a'X+b>0. If this condition is not true, then the detector 10
2 indicates an unvoiced frame. The values for the vector a and scalar b used by detector 102 are preferably the same as the initial values of a and b for statistical voiced detector 103.

Detector 102 determines the distance measure in a manner similar to generator 206 by performing calculations similar to those given in equations (16) through (28).

3 and 4 show the statistical voiced sound detector 103 of FIG.
The operations performed by the method are shown in more detail in flow chart form. Blocks 02 and 300 correspond to blocks 200 and 210 of FIG.
2. Blocks 302 and 318 implement statistical calculator 203. Blocks 320 and 322 implement weight calculator 204, and blocks 326 through 338 implement block 205 of FIG. 2. Generator 206 of FIG. 2 implements block
340. The subtractor 207 is implemented by block 308 or block 324.

Block 302 calculates a vector representing the average of the classifiers for the current frame and all previous frames. Block 300 determines whether speech or silence is present in the current frame, and if silence is present in the current frame, block 324 subtracts the average for each classifier from each classifier before transferring control to decision block 326. However, if speech is present in the current frame, then blocks 304 through 322
The statistical and weighting calculations are performed by: First, a mean vector is determined in block 302. Second, a matrix of sums of squares and products is calculated in block 304. Then, in block 306, the latter matrix is used together with a vector X indicating the means of the classifiers for the current and previous frames to calculate the covariance matrix T. Next, in block
At 308, the mean X is subtracted from the classifier vector x _n .

Next, block 310 calculates the current weight vector a, the current threshold b, and the classifier vector for the current frame.
x _n is used to calculate the probability that the current frame is unvoiced. After calculating the probability that the current frame is unvoiced, block 312 calculates the probability that the current frame is voiced. Block 314 then calculates the overall probability p _n that a frame will be unvoiced.

Blocks 316 and 318 compute two vectors u and v. The values contained in vector u represent the statistical mean value that each classifier would have if the frame were unvoiced, while vector v contains values representing the statistical mean value that each classifier would have if the frame were voiced. The actual values of the classifiers for the current and previous frames are clustered around vector u or vector v.
If these frames are found to be unvoiced,
The vectors representing the classifiers for the previous and current frames are clustered around vector u;
Otherwise, the previous classifier vector is the vector v
are clustered around

After executing blocks 316 and 318, control is passed to decision block 320. If N is greater than 99, then
Control is passed to decision block 322; otherwise, control is passed to block 326. Upon receiving control, block 322 then calculates a new weight vector a and a new threshold value b. Vector a and value b are used by the blocks preceding it in FIG. 3 in the next succeeding frame. Preferably, if N is required to be greater than infinity, vector a and scalar b are never changed and detector 103 continues from block 326 to
It will adapt in response only to the vectors v and u as shown in 328.

Blocks 326 through 338 implement the u/v classifier 205 of FIG. 2. Block 326 determines whether the power term of vector v of the current frame is greater than or equal to the power term of vector u. If this condition is true, then decision block 328 is executed. The latter decision block tests for voiced or unvoiced speech. If the decision of block 328 is true, the frame is marked as voiced by block 330; otherwise, the frame is marked as unvoiced by block 332. If vector v
If it is less than the power term of
The function of is executed and functions similarly. Finally, block 340
calculates the distance measure.

Figure 5 details, in flow diagram form, the operations performed by block 340 of Figure 4. Decision block 501 determines whether a frame has been designated unvoiced or voiced by examining the results of blocks 330, 332, 336, or 338. If the frame is designated voiced, path 507 is selected. Block 510 calculates probability _Pd , block 502 recalculates mean _m1 for voiced frames, and block 503 recalculates variance _k1 for voiced frames. If the frame is classified as unvoiced, decision block 501 selects path 508. Block 509 recalculates probability _Pd , block 504 recalculates mean _m0 for unvoiced frames, and block 505 recalculates variance _k0 for unvoiced frames. Finally, block 506 calculates the distance measure by performing the indicated calculations.

──────────────────────────────────────────────────── Continued from the front page (56) References: JP 61-48898 (JP, A) JP 60-200300 (JP, A) JP 60-114900 (JP, A) IEEE Transactions on Acoustics, Speech, and Signal Processes sing, Vol. ASSP-24, No. 3, June 1976, pp. 201-212

Claims (20)

[Claims] [Claim 1] An apparatus for determining the presence of a fundamental frequency in a non-training speech signal, comprising: means for generating a digital speech signal by sampling the speech signal in response to the non-training speech signal, forming frames of the digital speech signal, and processing each frame to generate a set of classifiers defining speech attributes; first means for calculating a first set of statistical distributions in response to the set of classifiers defining speech attributes of a first frame of the frames; second means for calculating a set of weights corresponding to one of the classifiers in response to the calculated set of first statistical distributions; and third means for determining the presence of a fundamental frequency in the first frame in response to the calculated weights, the set of classifiers, and the set of first statistical distributions. [Claim 2] The apparatus of claim 1, wherein said second means comprises: means for calculating a threshold value in response to said set of statistical distributions; and means for informing said first means of said set of weights and said threshold value to be used in calculating a second set of statistical distributions for a second frame different from said first frame. 3. The apparatus of claim 2, wherein said first means calculates a second set of statistical distributions further in response to the communicated set of weights and a second set of classifiers defining speech attributes of said second frame. 4. The method of claim 1, wherein the first means comprises: means for calculating an average of the classifiers for a previous frame; the average of the classifiers; and the set of weights notified.
and means for determining said second set of statistical distributions in response to said second set of classifiers. 5. The apparatus of claim 4, wherein said first means further comprises: means for detecting the presence of speech in each frame; and means for ceasing calculation of said second set of statistical distributions when no speech is detected in said second frame. 6. The apparatus of claim 5, wherein said first means further comprises: means for calculating a probability that said second set of classifiers represents an unvoiced frame and a probability that said second set of classifiers represents a voiced frame; and means for calculating an overall probability that a frame is unvoiced. 7. The apparatus of claim 6, wherein said first means further comprises means for calculating a set of average classifiers representing unvoiced frames and a set of average classifiers representing voiced frames. 8. The apparatus of claim 7, wherein said first means further comprises means for calculating a covariance matrix between said set of average classifiers representing unvoiced frames for said second frame and said set of classifiers representing unvoiced frames for said second frame. 9. The second means includes a covariance matrix;
9. The apparatus of claim 8, wherein said second set of statistical distributions is determined in response to said set of average classifiers for both voiced and unvoiced frames and said overall probability that a frame is unvoiced. 10. The apparatus of claim 9, wherein said third means determines the presence of said fundamental frequency in said second frame in response to said second set of statistical distributions and said set of average classifiers for voiced and unvoiced frames. [Claim 11] A method for determining the presence of a fundamental frequency in a non-training speech signal, comprising the steps of: generating a digital speech signal by sampling a non-training speech signal, forming frames of the digital speech signal, and processing each frame to generate a set of classifiers defining speech attributes; a first calculation step of calculating a first set of statistical distributions in response to the set of classifiers defining speech attributes of a first frame of one of the frames; a second calculation step of calculating a set of weights corresponding to one of the classifiers in response to the calculated set of first statistical distributions; and determining the presence of a fundamental frequency in the first frame in response to the calculated weights and set of classifiers and the first set of statistical distributions. 12. The method of claim 11, wherein said second calculating step comprises: calculating a threshold value in response to said set of statistical distributions; and reporting said set of weights and said threshold value for use in calculating a second set of statistical distributions for a second frame different from said first frame. 13. The method of claim 12, wherein said first calculating step further comprises calculating a second set of statistical distributions in response to the communicated set of weights and a second set of classifiers defining speech attributes of said second frame. 14. The method of claim 1, wherein said first calculation step comprises: calculating an average of said classifiers for a previous frame; and calculating said average of said classifiers and said set of signaled weights.
and determining said second set of statistical distributions in response to said second set of classifiers. 15. The method of claim 14, wherein the first calculating step further comprises the steps of: detecting the presence of speech in each frame; and ceasing calculation of the second set of statistical distributions when no speech is detected in the second frame. 16. The method of claim 15, wherein the first calculating step further comprises: calculating the probability that the second set of classifiers represents an unvoiced frame and the probability that the second set of classifiers represents a voiced frame; and calculating the overall probability that the frame is unvoiced. 17. The method of claim 16, wherein said first calculating step further comprises the step of calculating a set of average classifiers representing unvoiced frames and a set of average classifiers representing voiced frames. 18. The method of claim 17, wherein the first calculating step further comprises calculating a covariance matrix between the set of average classifiers representing unvoiced frames relative to the second frame and the set of classifiers representing unvoiced frames relative to the second frame.
17 ways. 19. The method of claim 18, wherein the second calculating step determines the set of second statistical distributions in response to a covariance matrix, the set of average classifiers for both voiced and unvoiced frames, and the overall probability that a frame is unvoiced. 20. The determining step comprises: determining the second set of statistical distributions in response to the second set of statistical distributions and the set of average classifiers for voiced and unvoiced frames.
20. The method of claim 19, further comprising determining the presence of the fundamental frequency in a frame.