Published techniques for computing line spectral frequencies (LSFs) generally avoid rootfinding methods because of concerns about convergence and complexity. However, this paper shows that stable predictor polynomials have properties that make rootfinding an attractive approach. It is well known that the problem of finding the LSFs for an N'th order predictor polynomial can be reduced to the problem of finding the roots of a pair of polynomials of order N/2 with real roots. The author extends this result by showing that these polynomials have the following properties: 1. It is possible to select starting points for a Newton's rootfinding method such that the iteration will converge monotonically to the largest root. 2. The Newton iteration can be modified to speed up the process while still maintaining good convergence properties. In this paper, the author presents the rootfinding procedures with proofs of their good convergence properties. Finally, he presents experimental results showing that this procedure performs well on speech signals, and that it can be implemented on fixed-point DSPs.

This work describes a new version of the decimation-in-degree (DID) transformation used by Wu and Chen (see ibid., vol.5, p.105-15, 1997) as part of their procedure for computing line spectrum pair (LSP) coefficients. This new version eliminates all nontrivial multiplications, requires fewer total operations (additions and multiplications), and is performed in place-eliminating a memory buffer.

A new speech enhancement scheme based on a generalized Wiener filter formulation is proposed. A ternary-valued parameter is derived empirically based on the likelihood of the input signal vector being classified as speech. This parameter controls the Wiener filter coefficient in order to obtain an improved speech spectral estimate. This "ternary-decision" concept renders a logical compromise between the simple, hard, "binary speech/noise decision" filtering and the elaborate "soft-decision" filtering approaches in terms of practicality and performance. An important feature in our scheme is that we exploit the interframe spectral relationship to reinforce the assessment of the likelihood of weak speech components. This feature prevents many weak formants from being disproportionally attenuated as in most previous schemes. Other important features of our scheme include a novel speech/noise classifier and a robust noise median amplitude tracker, both of which make the estimate of noise spectrum more reliable. A preliminary evaluation of this new scheme is reported here.

A high-quality speech coder at 400 b/s has been developed. This coder is designed to provide reliable communication over an HF channel in a hostile environment. The major techniques utilized for coding include vector quantization, interpolation, and trellis coding. The test results in the presence of speaker variations, background noise, and acoustic impairments are discussed. The three speaker average DRT was 80 for this coder.>

A real time low rate voice coder/decoder has been developed and is being tested using speakers in simulated operational environments. The unit uses the TMS32010 to perform the signal processing functions, and is packaged in a manner suitable for an office or benign field environment. The software implements the following algorithms:\bullet2400 bps LPC using the United States Government standard LPC-10 algorithm.\bullet800 bps using vector quantization.\bullet400 bps using frame repeat vector quantization with trellis coding of gain and pitch. The hardware configuration is described, along with a brief explanation of the available test configurations and execution options. A review of the coding algorithms is presented, with a summary of the operational parameters. Test results include measured execution times, program and data memory sizes. Preliminary listening tests indicate that the voice quality at 800 bps is very close to that achieved at 2400 bps with the standard algorithms.

Algorithms for a multiple rate (2400, 800, and 400 bps) CODEC are described. The U.S. Government standard LPC-10 algorithm is used for the 2400 bps rate. Lower rates are obtained by applying the source coding techniques of vector quantization and trellis coding to the LPC analysis output. In this paper, the algorithms used are described, along with data rate control and formatting techniques for use with a fixed rate modem. Simulation test results for different vector quantizer codebook sizes and structures are presented, along with DRT test results showing the voice quality achieved at each rate. Fortran simulations of the algorithms have been performed. Real time implementation of the CODEC is in progress, with completion expected by the end of 1985.

A new method of implementing filter banks for subband coding of speech is presented. First, Esteban's Quadrature Mirror Filter principle is extended to allow the direct synthesis of filter banks with any number of equal size filter bands. Then, by combining the quadrature filter characteristic with the polyphase network implementation of filter banks, a new filter bank structure is obtained which requires 35% fewer computations than existing designs. In this paper the theory of operation of the Polyphase Quadrature Filter is presented and techniques for its efficient implementation are described. Then, examples of filter banks using this approach are shown and compared with other designs, and a simulation of a 16 Kbits/s coder using these filters is presented.