Direct line spectral frequency adaptation in second order cascade sections

We propose the adaptive cascade recursive least squares with subsection adaptation (CRLS-SA) algorithm for direct estimation of line spectral frequencies (LSF) for use in speech coding. The least squares based CRLS-SA algorithm has a low computational complexity since each section is adapted independently from other sections. Using CRLS-SA, the LSF can be obtained directly from the coefficients of each section of the CRLS-SA structure. Two CRLS-SA approaches can be used for estimating LSF: use CRLS-SA as an AR inverse filter and obtain the LSF by converting the coefficients of each second order filter using a look-up table or, alternatively, construct a cascade of second order LSF polynomials P(z) and Q(z), and adapt each using CRLS-SA. The results show that the LSF estimates from CRLS-SA are competitive with the results obtained by the popular autocorrelation method followed by approximation using Chebyshev polynomials.