AU741881B2

AU741881B2 - Method and apparatus for determining paremeters of a model of a power spectrum of a digitised waveform

Info

Publication number: AU741881B2
Application number: AU71457/00A
Authority: AU
Inventors: Jason Lukasiak; Burnett Ian Shaw
Original assignee: University of Wollongong; Motorola Australia Pty Ltd
Current assignee: University of Wollongong; Motorola Solutions Australia Pty Ltd
Priority date: 1999-11-12
Filing date: 2000-11-06
Publication date: 2001-12-13
Anticipated expiration: 2020-11-06
Also published as: AU7145700A

Description

P/00/011 Regulation 3.2

AUSTRALIA

Patents Act 1990 r fe.

ORIGINAL

COMPLETE SPECIFICATION STANDARD PATENT Invention Title: "METHOD AND APPARATUS FOR DETERMINING PARAMETERS OF A MODEL OF A POWER SPECTRUM OF A DIGITISED WAVEFORM" The following statement is a full description of this invention, including the best method of performing it known to me/us: CR1054AC METHOD AND APPARATUS FOR DETERMINING PARAMETERS OF A MODEL OF A POWER SPECTRUM OF A DIGITISED WAVEFORM This invention relates to a method and apparatus for determining parameters of a model of a power spectrum of a digitised waveform, especially, oooo though not exclusively, so as to improve the performance of a class of methods used for compressing digitised speech for storage or for transmission over digital 10 communication channels.

BACKGROUND OF THE INVENTION 9ooo o o 0o0 Digitised speech signals consist of sequences of numerical values 15 representing samples of a continuous speech waveform. In speech compression systems these samples are analysed to obtain parameters, which can subsequently be used as inputs to some process for generating an approximation to the original o: samples.

An important step in many techniques for performing such analysis is to determine a set of parameters of a filter, such that the squared magnitude of the filter's frequency response approximates the power spectrum of the input speech signal, according to some criterion. The filter parameters are, therefore, equivalently regarded as parameters of a model of the power spectrum of the speech signal. By applying an appropriate excitation signal to the input of the filter, an output signal that approximates the speech signal can be produced.

A common method of modelling the power spectrum of the speech signal is by means of an autoregressive filter. The way that such a filter determines each sample of its output involves multiplying each of several preceding output samples by a distinct coefficient, or parameter, and adding the resulting products and a corresponding sample of the filter input. A number of methods are known for determining the coefficients such that the squared magnitude of the frequency response of the resulting autoregressive filter approximates the power spectrum of the speech signal.

The filter parameters can be used to generate an approximation to the speech signal by applying an appropriate excitation signal to the input of the filter.

:°oooo The approximation is referred to here as the synthesised signal. The excitation signal can be obtained in a number of ways. In many systems, the excitation signal is taken over some fixed interval to be one of a finite number of i" ~predetermined sample sequences. One sequence is chosen by using each of the possible predetermined sample sequences to generate a possible synthesised signal •go• oo in the manner described above. The one which results in the synthesised signal ooo* that best approximates the input speech signal, according to some criterion is 15 chosen.

The synthesised signal generated using the excitation signal chosen as described above will not be identical to the input speech signal. The error depends largely on the number of possible predetermined excitation sequences, with more ooe° sequences allowing a smaller error to be attained. If the number of sequences is not sufficiently large, the synthesised signal will contain audible distortion, compared with the input signal.

It is highly desirable for a system to be designed in such a way that the errors in the synthesised signal contribute as little as possible to audible distortion.

The degree to which errors in the sample values of the synthesised signal give rise to audible distortion depends on the properties of the human auditory system.

Minimising the contribution of sample errors to audible distortion therefore requires that the properties of the human auditory system be exploited in some way. Several methods of exploiting the properties of the auditory system are 3 known. For example, these properties can be exploited in the choice of the criterion used to select a single excitation signal sequence from the possible predetermined sequences, or in the design of the sequences themselves. The present invention, in contrast, exploits the properties of the auditory system in the method of determining the filter parameters.

In this specification, including the claims, the terms "comprises", "comprising" or similar terms are intended to mean a non-exclusive inclusion, :°oooo such that a method or apparatus that comprises a list of elements does not include those elements solely, but may well include other elements not listed.

i "'"BRIEF SUMMARY OF THE INVENTION It is an object of the present invention to provide a method and apparatus oooo for determining parameters of a model of the power spectrum of a digitised waveform.

Accordingly, in one aspect, the invention provides a method of determining parameters of a model of the spectrum of a digitised waveform, the method comprising the steps of: receiving a first input signal; determining a first power spectrum estimate of the first input signal, wherein the power spectrum estimate is a function of frequency; computing a masking function, wherein the masking function is a function of frequency that depends on the first power spectrum estimate; modifying the first power spectrum estimate to produce a second power spectrum estimate, wherein a value of the second power spectrum estimate is less than a value of the first power spectrum estimate at frequencies at which a value of the masking function is greater than the value of the first power spectrum estimate; and I 4 determining a set of filter parameters, wherein a squared magnitude of a frequency response associated with a filter controlled by the set of filter parameters is an approximation to the second power spectrum estimate.

Suitably, the step of determining the first power spectrum estimate may include calculating squared magnitudes of Fast Fourier Trnasforms of the first input signal.

oooo Preferably, the step of determining a set of filter parameters may include ocalculating an autocorrelation function from the second power spectrum estimate.

ooo• l Preferably, the filter parameters may be for an autoregressive filter.

~According to a second aspect, the invention provides an apparatus for filtering a first signal based on the power spectrum of a second signal, the apparatus comprising: an input terminal for receiving an input signal; a first analyser having an input coupled to the input terminal and an output, for determining a first power spectrum estimate of the first input signal, wherein the power spectrum estimate is a function of frequency; a second analyser having an input coupled to the output of the first analyser and an output, for computing a masking function, wherein the masking function is a function of frequency that depends on the first power spectrum estimate; a transformer having a first input coupled to the output of the first analyser, a second input coupled to the output of the second analyser and an output, for modifying the first power spectrum estimate to produce a second power spectrum estimate, wherein the value of the second power spectrum estimate is less than the value of the first power spectrum estimate at frequencies at which the value of the masking function is greater than the value of the first power spectrum estimate; and a third analyser having an input coupled to the output of the transformer and an output, for determining a set of filter parameters, wherein a squared magnitude of a frequency response associated with a filter controlled by the set of filter parameters is an approximation to the second power spectrum estimate.

Preferably, in use the first analyzer may determine the first power spectrum estimate by calculating squared magnitudes of Fast Fourier Trnasforms of the o first input signal.

Suitably, in use the third analyzer may determine the set of filter 10 parameters by calculating an autocorrelation function from the second power spectrum estimate.

Preferably, the filter can be an autoregressive filter.

BRIEF DESCRIPTION OF THE DRAWINGS An embodiment of the invention will now be more fully described, by way of example, with reference to the drawings, of which: ~FIG. I shows a block diagram of an apparatus used to determine parameters of a model of the power spectrum of a digitised signal; FIG. 2 shows a flowchart of a method for determining parameters of a model of the power spectrum of a digitised signal utilising the apparatus illustrated in FIG. 1; FIG. 3 shows a block diagram of the second analyser used for determining a masking threshold function shown in FIG. 1; and FIG. 4 shows a spreading function used by the apparatus of FIG. 3 in determining the masking threshold function.

1 1 6 DETAILED DESCRIPTION OF THE DRAWINGS Thus, the overall scheme of the invention according to one embodiment is shown in FIG. 1, FIG. 2, FIG. 3 and FIG.4.

As shown in FIG. 1, an input signal is delivered from an input terminal 100 to a first analyser 101. The first analyser 101 computes a Fast Fourier Transform (FFT) of the input signal and then calculates the squared magnitudes of the FFT o coefficients. The resultant squared magnitudes represent the first power spectrum of the input signal at discrete frequency values, and appear at an output of the first 10 analyser 101. This procedure is shown at steps 201 and 202 in FIG. 2.

i "'"The first power spectrum of the input signal is fed from the output of the first analyser 101 to an input of a second analyser 102, which calculates a masking 0° OO° threshold function for the input power spectrum, for example using the method described in, "Transform coding of audio signals using perceptual noise criteria", by J.D. Johnston, IEEE Journal on Selected Areas in Communications, vol.6, pp314-323, February 1988.

The masking threshold function from an output of the second analyser 102 r-r ~together with the power spectrum from the output of the first analyser 101 are delivered to a transformer 103. The transformer 103 comprises a masked frequency calculator 104, and a modified spectrum generator 105. The masking threshold function determined by the second analyser 102 and the power spectrum determined by the first analyser 101 are provided to the masked frequency calculator 104, which calculates frequencies of the input spectrum that are deemed inaudible. This is achieved by comparing the power spectrum value for each discrete frequency to the masking threshold function for the same frequency. If the power spectrum value is less than the threshold value the frequency is deemed to be masked. This is repeated for all frequencies in the input signal. A representation of the values of the masked frequencies appear at an output of the masked frequency calculator 104. This procedure is shown at step 203 in FIG. 2.

The power spectrum from the first analyser 101 and masked frequencies from the masked frequency calculator 104 are provided to the modified spectrum generator 105, which generates a modified spectrum such that the value of the modified power spectrum is less than the value of the power spectrum from the first analyser 101 at frequencies provided by the masked frequency calculator 104.

":The modified power spectrum (second power spectrum estimate) appears at an :o output of the modified spectrum generator 105. This procedure is shown at step 10 204 in FIG. 2.

~The modified power spectrum (second power spectrum estimate) from the output of the modified spectrum generator 105 is provided at an input to a filter #see parameter generator 106, which calculates a set of filter parameters (filter coefficients). In the preferred embodiment, the filter parameters are parameters of an autoregressive filter. They are calculated so that the squared magnitude of the S:°o frequency response of the filter is an approximation to the modified power spectrum provided by the modified spectrum generator 105. The filter parameters are provided at an output of the filter parameter generator 106. This procedure is shown at step 205 in FIG. 2.

The way that the filter parameters are determined from the modified power spectrum involves two steps. Firstly, an autocorrelation function is calculated from the modified power spectrum via an Inverse Fast Fourier Transform operation using any known technique, for example using the method described in Digital Processing of Speech Signals, by L.B. Rabiner and R.W. Schafer, Prentice Hall, New Jersey, 1978, page 164.

Secondly, parameters of an autoregressive filter are computed from the autocorrelation function using any known technique, for example using the Levinson-Durbin recursion as described in "Linear Prediction: A Tutorial Review", by J. Makhoul, Proceedings of IEEE, Vol. 63, pp.561-580, 1975. The filter parameters appear at the output of the filter parameter generator 106.

FIG. 3 shows the detailed operation of the second analyser 102. The input power spectrum from the first analyser 101 is provided to a sub-band analyser 301, which groups the input power spectrum values into non overlapped bands. The bands are of variable bandwidth and are designed to simulate the frequency response of the human ear, for example as described in, "Critical bands", by B.

Scharf, in Foundations of Modern Auditory Theory, edited by J. Tobias, New York: Academic Press, pp. 159-202, 1970. The energies of all frequency 0 10 components contained within each band are then summed together to give a total energy value for each band. A band energy waveform is generated. This is a function of discrete frequency whose value within each band described earlier is constant and equal to the total energy for the band. The band energy waveform appears at an output of the sub-band analyser 301.

15 The output of the sub-band analyser is provided as the input to an interband masking calculator 302. The inter-band masking calculator 302 convolves o. the band energy waveform with a spreading function to produce a spread band energy waveform. The spreading function is shown in FIG. 4. The spread band energy waveform appears at an output of the interband masking calculator 302.

The spread band energy waveform from the inter-band masking calculator 302 is supplied to the input of a masking threshold calculator 303, which determines an Initial Masking Threshold Function (IMTF). The IMTF is calculated for each band from the following formula: IMTF(i) Energy(i) 0(i) where Energy(i) represents the total energy of the ith band of the spread band energy waveform measured in decibels; 0(i) is given by O(i) a(14.5 i) a)7 where

SFM

a min( ,1) SFMmax and

G

SFM 10log G m A,n SFMmax is an empirically determined value; Gm is the geometric mean of the S" 5 power spectrum estimate; and A, is the arithmetic mean of the first power spectrum.

The IMTF appears at the output of the masking threshold calculator 303.

The IMTF is supplied to a threshold nomaliser 304. The threshold normaliser 304 adjusts the IMTF to account for misestimation of the Energy(i) values resulting from the shape of the spreading function. One example of a procedure for this adjustment is described in "Transform coding of audio signals using perceptual noise criteria", by J.D. Johnston, IEEE Journal on Selected Areas in Communications, vol.6, pp 3 14-323. The output of the threshold normaliser 304 is the masking threshold function.

15 It will be appreciated that although one particular embodiment has been described here in detail, various modifications and improvements can be made by a person skilled in the art without departing from the scope of the present invention.

Claims

1. A method of determining parameters of a model of the spectrum of a digitised waveform, the method comprising the steps of receiving a first input signal; determining a first power spectrum estimate of the first input signal, wherein the power spectrum estimate is a function of frequency; computing a masking function, wherein the masking function is a function of frequency that depends on the first power spectrum estimate; modifying the first power spectrum estimate to produce a second power spectrum estimate, wherein a value of the second power spectrum estimate is less than a value of the first power spectrum estimate at frequencies at which a value of the masking function is greater than the value of the first power spectrum estimate; and 15 determining a set of filter parameters, wherein a squared magnitude of a frequency response associated with a filter controlled by the set of filter parameters is an approximation to the second power spectrum estimate.

2. A method as claimed in claim 1, wherein in the step of determining the first power spectrum estimate includes calculating squared magnitudes of Fast Fourier Transforms of the first input signal.

3. A method as claimed in claim 1, wherein in the step of determining a set of filter parameters includes calculating an autocorrelation function from the second power spectrum estimate.

4. A method as claimed in claim 3, wherein in the filter parameters are for an autoregressive filter. 11 An apparatus for filtering a first signal based on the power spectrum of a second signal, the apparatus comprising: an input terminal for receiving an input signal; a first analyser having an input coupled to the input terminal and an output, for determining a first power spectrum estimate of the first input signal, wherein the power spectrum estimate is a function of frequency; a second analyser having an input coupled to the output of the first analyser •10 and an output, for computing a masking function, wherein the masking function is a function of frequency that depends on the first power spectrum estimate; a transformer having a first input coupled to the output of the first analyser, oooo a second input coupled to the output of the second analyser and an output, for modifying the first power spectrum estimate to produce a second power spectrum estimate, wherein the value of the second power spectrum estimate is less than the .ooooi S•value of the first power spectrum estimate at frequencies at which the value of the masking function is greater than the value of the first power spectrum estimate; and a third analyser having an input coupled to the output of the transformer and an output, for determining a set of filter parameters, wherein a squared magnitude of a frequency response associated with a filter controlled by the set of filter parameters is an approximation to the second power spectrum estimate.

6. The apparatus of claim 5, wherein in use the first analyzer determines the first power spectrum estimate by calculating squared magnitudes of Fast Fourier Trnasforms of the first input signal.

7. A method as claimed in claim 5, wherein in use the third analyzer determines the set of filter parameters by calculating an autocorrelation function from the second power spectrum estimate.

8. A method as claimed in claim 7, wherein in the filter is an autoregressive filter.

9. A method of determining parameters of a model of the spectrum of a digitized wave form substantially as hereinbefore described with reference to the accompanying drawings.

10. An apparatus for filtering a first signal based on the power spectrum of a S second signal substantially as hereinbefore described with reference to the accompanying drawings. 15 DATED this Sixth day of November 2000 MOTOROLA AUSTRALIA PTY LTD and UNIVERSITY OF WOLLONGONG By Their Patent Attorneys FISHER ADAMS KELLY Registered Patent Attorney