Bitstream-based feature extraction method for a front-end speech recognizer

Abstract

A feature extraction process for use in a wireless communication system provides automatic speech recognition based on both spectral envelope and voicing information. The shape of the spectral envelope is used to determine the LSPs of the incoming bitstream and the adaptive gain coefficients and fixed gain coefficients are used to generate the “voiced” and “unvoiced” feature parameter information.

1 . An automatic speech recognition system for receiving as an input a coded bitstream speech signal and generating therefrom feature recognition parameters, said recognition system including 
 a spectral envelope detection arrangement for generating a first set of feature recognition parameters derived from line spectrum pair (LSP) information from the encoded bitstream; and    an excitation signal extraction arrangement for generating a second set of feature recognition parameters derived from adaptive codebook gain and fixed codebook gain information related to voiced and unvoiced information, respectively.    
   
   
       2 . The automatic speech recognition system of  claim 1  wherein the bitstream is defined on a frame-by-frame basis, each frame include including a first set of bits associated with spectral envelope detection and the LSP coefficients for the generation of the first set of feature recognition parameters and a second set of bits associated with the excitation signal extraction and the voiced/unvoiced information for the generation of the second set of feature recognition parameters.  
   
   
       3 . The automatic speech recognition system of  claim 2  wherein the spectral envelope detection arrangement comprises 
 an inverse quantizer responsive to the first set of bits in each frame within the bitstream;    an interpolator coupled to the output of the inverse quantizer for providing an output at a predetermined rate;    an LSP-to-CEP conversion unit, responsive to the output of the interpolator to convert the LSP coefficients into CEP coefficients; and    a cepstral weighting filter for smoothing the output of said speech recognition system.    
   
   
       4 . The automatic speech recognition system of  claim 2  wherein the second set of bits within each frame are is divided into subframes, with an adaptive gain coefficient and a fixed gain coefficient calculated for each subframe.  
   
   
       5 - 6 . (canceled)  
   
   
       7 . A method of performing feature extraction on a continuous bitstream of encoded speech, the method comprising the steps of: 
 determining the line spectrum product coefficients from a predetermined set of bits in each frame in the continuous bitstream;    converting the coefficients into spectral envelope information in the form of a first set of feature recognition parameters;    determining adaptive codebook gain coefficients and fixed codebook gain coefficients from another set of bits in each frame of said continuous bitstream; and    calculating voiced and unvoiced information in the form of a second set of feature recognition parameters from the adaptive and fixed codebook gain coefficients.    
   
   
       8 . The method as defined in  claim 7  wherein in determining the line spectrum product coefficients, performing the steps of: 
 performing an inverse quantization on the predetermined set of bits in each frame; and    performing an LSP to CEP conversion on the inverse quantized results.    
   
   
       9 . The method as defined in  claim 8  wherein in performing the LSP to CEP conversion, the LSP coefficients are first converted to LPC coefficients and the LPC coefficients are converted to CEP coefficients.  
   
   
       10 . The method as defined in  claim 8  wherein the method comprises the additional step of weighting the CEP coefficients.  
   
   
       11 - 13 . (canceled)