Title of Invention

A TECHNIQUE FOR GENERATING A MOTHER WAVELET FROM THE SPEECH SIGNAL FOR EVALUATION OF WAVELET TRANSFORMS TO CAPTURE THE TIME VARIATIONS IN THE SPEECH SIGNAL

Abstract An innovative method proposed for generation of orthogonal basis function of mother wavelet from the voiced speech segment for processing, analysis, synthesis, recognition of speech as well as speaker, consists of a tracking a pitch synchronous vector from the voiced speech segment, processing it suitably so that it becomes localized in time domain by passing it via a time domain window and processing it so that it becomes localized in frequency domain and then suitably down sampling it so that it represents a wavelet i.e. coefficients of low pass filter and then evaluating coefficients for high pass filter, verifies that it actually forms an orthogonal basis function and can concentrate the energy in small of wavelet coefficient and can capture the speech and speaker specific time varying parameter,
Full Text FORM - 2
THE PATENTS ACT, 1970
(39 OF 1970)
COMPLETE SPECIFICATION.
(Section 10; Rule 13)
Title: - A technique for generating a mother wavelet from the speech signal for evaluation of wavelet transforms to capture the time variations in the speech signal,
Dr.S.D.apte,
Having office at Plot No.9, Revenue Colony
Santh Shivaji Nager,
Sangli,- 416 416. (M.H), INDIA.

The following specification particularly describes the nature of this invention and the manner in which it is performed:-


ORIGINAL
123/MUM/2002

The present invention relates to an innovative technique for generating a mother wavelet from the speech signal for evaluation of wavelet transform to capture the time variation in the speech signal.
The existing mother wavelets are not suitable for extracting the speaker specific or speech or specific parameters from the speech signal. If the wavelet closely resembles the signal under consideration, then the gain resolution of decomposition improve. The select a suitable wavelet for speech signal, multiple mother wavelets will have to be used. This increases the computational complexity in the wavelet decomposition. The methods already existing for generation of mother wavelet are highly computationally involved. For calculate the coefficients the matrix is formed with mirror filters as Co, C1, C2, C3, & C3, -C2, Cl,-CO. The matrix multiplied by its transpose must be the identity matrix .The non-linear equations in 4 unknown are solved to generate mother wavelet coefficients. Here, if it is required to evaluate to 16coefficients in 16 unknowns are to be solved. The exact solution may not exist and the strict orthogonally and compact support is not maintained.
Objects;-
An object of the present invention is to provide a way to generate mother wavelet from the speech signal using a simple and innovative technique that does not involve any computational completely.
A further object of the present invention is to improve the gain and resolution of wavelet decomposition. Allow the user to improve the resolution by selecting required number of points in the wavelet. And still maintain simplicity of calculation.
A still further object of the invention is to provide a way to generate multiple frames of references for improving the verification accuracy of speech reorganization or speaker verification.
The present invention provides a method for generating an orthogonal basis function for the analysis of speech signals. The mother wavelet must satisfy certain conditions in order that it can be used as a basis function. It must be localized in time domain. It must also have finite Fourier transform. The voiced speech signal is first segmented into optimum processing interval by

dividing it into pitch synchronous vectors. A pitch synchronous vector is further used for the generation of mother wavelet. It is first localized in time domain by passing it via a time domain window. Its Fourier transform is then made finite. Extra zeros are further appended to make the segment size equal to 128.The segment is further suitably down sampled to generate the required number of filter coefficients. The quadrature mirror filter is designed form the coefficients generated after down sampling, the coefficients generated are energy normalized and amplitude normalized .Theses quadrature mirror filter pair is used as a mother wavelet i.e. orthogonal basis function.
It is experimentally verified that the generated mother wavelet using the stated method indeed generates an orthogonal basis function .It improves the verification accuracy for speech recognition and speaker verification. It gives better for speech synthesis and recovery of missing speech packets.
The invention will now be described in greater detail by way of an example, with reference to the accompanying figures, in which,
Fig.l shows a speech utterance of a speaker
Fig.2 shows the pitch synchronous speech vector for the utterance in Fig.l, Fig.3A & figure 3B depicts the generated mother wavelet coefficient in which figure 3A shows LPF coefficient, figure 3B shows HPF coefficients. Fig.4 represents the WT coefficient of the pitch synchronous vector respect to the generated mother wavelet.
Referring to Fig.l, it is observed that most of the energy is concentrated in the vowel part of the utterance; hence, vowel part is extracted form the speech utterance. The vowel part is further processed to track the optimum processing interval i.e. pitch synchronous vector for the utterance in Fig.l is shown in fig.2.
Referring Fig.3 it is seen that the pitch synchronous vector is processed further to design a mother wavelet. The design is such that the mother wavelet is localized in time domain. It is passed via a time domain window. Its Fourier transform is made zero and then extra zeros are appended so that the size of the vector is 128.Now it is suitably down sampled to generate the required low pass filter coefficient represented in Fig.3.

I Claim,

l.An innovative method proposed for generation of orthogonal basis function of mother wavelet from the voiced speech segment for processing, analysis, synthesis, recognition of speech as well as speaker, consists of a tracking a pitch synchronous vector from the voiced speech segment, processing it suitably so that it becomes localized in time domain by passing it via a time domain window and processing it so that it becomes localized in frequency domain and then suitably down sampling it so that it represents a wavelet i.e. coefficients of low pass filter and then evaluating coefficients for high pass filter, verifies that it actually forms an orthogonal basis function and can concentrate the energy in small of wavelet coefficient and can capture the speech and speaker specific time varying parameter,
2. A method as claimed in claim 1, wherein a voiced speech signal is used for the generation of mother wavelet,
3. A method as claimed in claim 1&2, wherein a voiced speech signal is first segmented in to optimum processing interval by dividing it into pitch synchronous vectors,
4. A method as claimed in above claims wherein the quadrature mirror filter designed from the coefficients generated after down sampling is used as a mother wavelet,
5. An innovative method proposed for generation of orthogonal basis
function of mother wavelet from the voiced speech segment for processing,
such as here described with reference to accompanying drawings.

Dated this 11th Day of Feb.2002
(Joseph Varikasery)
of Varikasery&varikasery
5

Documents:

123-mum-2002-abstract(11-2-2002).pdf

123-mum-2002-cancelled pages(08-09-2005).pdf

123-mum-2002-claims(11-2-2002).pdf

123-mum-2002-claims(granted)(08-09-2005).doc

123-mum-2002-claims(granted)(08-09-2005).pdf

123-mum-2002-claims(granted)-(8-10-2007).pdf

123-mum-2002-correspondence(ipo)-(12-11-2007).pdf

123-mum-2002-correspopndence(07-07-2004).pdf

123-mum-2002-correspopndence(ipo)(25-08-2006).pdf

123-mum-2002-description(complete)-(11-2-2002).pdf

123-mum-2002-description(granted)-(8-10-2007).pdf

123-mum-2002-drawing(08-09-2005).pdf

123-mum-2002-drawing(11-2-2002).pdf

123-mum-2002-drawing(granted)-(8-10-2007).pdf

123-mum-2002-form 1(11-02-2002).pdf

123-mum-2002-form 19(07-07-2004).pdf

123-mum-2002-form 2(complete)-(11-2-2002).pdf

123-mum-2002-form 2(granted)(08-09-2005).doc

123-mum-2002-form 2(granted)(08-09-2005).pdf

123-mum-2002-form 2(granted)-(8-10-2007).pdf

123-mum-2002-form 2(title page)-(11-2-2002).pdf

123-mum-2002-form 2(title page)-(granted)-(8-10-2007).pdf

123-mum-2002-power of attorney(25-08-2003).pdf

123-mum-2002-specification(amended)-(8-9-2005).pdf

abstract1.jpg


Patent Number 210672
Indian Patent Application Number 123/MUM/2002
PG Journal Number 43/2007
Publication Date 26-Oct-2007
Grant Date 08-Oct-2007
Date of Filing 11-Feb-2002
Name of Patentee DR. S. D. APTE
Applicant Address PLOT NO.9, REVENUE COLONY, SANTH SHIVAJI NAGAR, SANGLI
Inventors:
# Inventor's Name Inventor's Address
1 DR. S.D. APTE PLOT NO.9, REVENUE COLONY, SANTH SHIVAJI NAGAR, SANGLI 416 416
PCT International Classification Number G10L 15/00
PCT International Application Number N/A
PCT International Filing date
PCT Conventions:
# PCT Application Number Date of Convention Priority Country
1 NA