Session X: SIGNAL PROCESSING FORUM - PART 1
Sunday, May 12, 14:00 18:30 h Digital audio systems are unlike conventional analog systems in which signals can be any value between a minimum to maximum and occur continuously in time. Digital audio systems use finite precision in representing signals and coefficients and in performing arithmetic operations. Consequently, system performance is determined by the precision that is used throughout the system. This paper discusses the factors that influence the performance of Infinite Impulse Response filters in high performance audio applications using fixed point arithmetic. High performance DSPs have been used extensively for the implementation of signal transforms in transform based audio coders. While the first generation of DSPs featured single stage MAC (Multiple Accumulate) blocks, the current generation of DSPs feature dual-MAC hardware blocks. Though fast algorithms are available for implementation of transforms, a re-look at the algorithms from this semi-parallel architecture point of view is beneficial as it leads to more efficient implementations. This paper looks specifically at the family of lapped transforms, and quantifies the implementation efficiency of traditional fast optimizations on architectures with this type of semi-parallel computing capability, and derives algorithmic methods of increasing this efficiency. Just recently, a new type of delta-sigma converter, Trellis Noise Shaping Converter, has been introduced. When it is used to generate a bit stream in 1-bit digital audio format, its trellis structure with Viterbi algorithm enables more efficient use of data bits, which yields better performance in stability, signal to noise ratio and tonal behavior. It solves most of the performance problem caused by harsh non-linearity inherent in 1-bit quantization. New Sigma Delta Modulators (SDMs) topologies for use in Super Audio CD (SACD) applications are introduced, called Sigma Delta Pre-Correction (SDPC), which allow the generation of ultra-high quality DSD. Spurious peaks, which are known theoretically to exist in SDMs, are present at levels well below -165 dB, even if undithered. Already a slight amount of dither, will further reduce these signals to levels which are with standard numerical precision undetectable. A method is presented for improving current coding efficiency in DSD signals. The goal of this work is to explore new compression techniques which are tailored to the DSD format and which are meant to complement the current lossless DST compression practice used for SACD. The new technique builds on principles illustrated in previous papers. The method makes use of the highly oversampled character of DSD. Example implementations and results have been obtained. Losses to stability and signal-to-noise ratio have been measured and their audio effects have been minimized and quantified. Lower bounds are established on the compression ratio of these methods. This is viewed as a first step for a potentially constant bit rate compression scheme. Time quantization and noise shaping applied to linear frequency modulation (LFM) can form an alternative although unconventional means of generating 1-bit uniformly sampled code that is similar in structure to a feedback sigma-delta modulator (SDM). Fundamental insight into the SDM process and base-line coding spectrum emerges, where specifically linearity of signal conversion is studied and compared to that of linear pulse code modulation (LPCM). Time dispersive limiters both within and outside the noise shaper are investigated and their consequence on linearity explored. A noise averaging simulation reveals intrinsic distortion and noise modulation to be low when appropriate dither is used. This paper clarifies some of the confusion, which has arisen over the efficacy of dither in PCM and Sigma-Delta Modulation (SDM) systems. It presents a means of analyzing "in-band" idle tone structure and describes a fair means of comparison between PCM and SDM. It presents results, which show that dither can be effective in S-D systems. This is Part 3 of our ongoing investigation into the behavior of 1-bit sigma-delta modulators. It addresses the following topics: The method of noise shaping consisting in minimization of quantization noise in neighborhood of harmonics of digital test sinusoidal signal is considered. This method provides quantization noise power on harmonics frequencies considerably smaller, than methods of noise shaping with uniform power spectral density. For example for 16 bit, 44100 Hz sampling frequency, -60 dB level, 1378 Hz test signal this method provides total harmonic distortion (THD) 0.00067% (without noise shaping THD=1.1%, with dither in a frequency band (20000-22050) Hz and a standard deviation 0.46 quantum THD=0.043%). |
|