Digital Signal Processing Reference
In-Depth Information
section 4.3.4, and solutions might, for example, be based on iterative methods such as
Expectation-maximize (EM) or Markov chain Monte Carlo (MCMC) which are both
powerful and computationally intensive. This is, however, likely to be in accord with
continual increases in speed and capacity of computational devices.
To conclude, the range of problems encountered in audio signals from all sources,
whether from recorded media or communications and broadcast channels, present
challenging statistical estimation problems. Many of these have now been solved
successfully, but there is still significant room for improvement in achieving the highest
possible levels of quality. It is hoped that the powerful techniques which are now
practically available to the signal processing community will lead to new and more
effective audio processing in the future.
5
DIGITAL AUDIO SYSTEM
ARCHITECTURE
Mark Kahrs
Multimedia Group
CAIP Center
Rutgers University
P.O. Box 1390
Piscataway, NJ 08855-1390
Notes
1.
the 'Packburn' unit achieved masking within a stereo setup by switching between channels
2.
With acknowledgement to Mr. B.C. Breton, Scientific Imaging Group, CUED
3. provided that no samples are missing from the first
P
elements of s; otherwise a correction must be
made to the data covariance matrix (see [Godsill, 1993])
4. the approximation assumes that the parameter likelihood for the first
P
data samples is insignificant
[Box and Jenkins, 1970]
5.
These techniques are also often referred to as 'spectral subtraction'. We will not use this terminology
in order to avoid ambiguities between the general principle and the particular technique described in [Boll,
1979], nor will we use the term 'spectral estimation' as quite a number of the STSA techniques are not based
on a statistical estimation approach.
6.
This suppression rule is derived by analogy with the well-known Wiener filtering formula replacing
the power spectral density of the noisy signal by its periodogram estimate.
7.
Strictly speaking, this effect could still be perceived for longer window durations when the relative
signal level approaches 1. However, it is then perceived more like an erratic fluctuation of the sinusoid level
which is hardly distinguishable from the phenomenon to be described in section 4.17.
8. More precisely, the quantity displayed is the signal power estimated from 10ms frames. As the
power spectral densities of the two types of noise exhibit a strong peak at the null frequency, the two noises
were pre-whitened by use of an all-pole filter [Cappé, 1991]. This pre-processing guarantees that the noise
autocorrelation functions decay sufficiently fast to obtain a robust power estimate even with short frame
durations [Kay, 1993].
kahrs @caip.rutgers.edu
Abstract:
Audio signal processing systems have made considerable progress over
the past 25 years due to increases in computational speed and memory capacity. These
changes can be seen by examining the implementation of increasingly complex algorithms
in less and less hardware. In this chapter, we will describe how machines have been
designed to implement DSP algorithms. We will also show how progress in integration
has resulted in the special purpose chips designed to execute a given algorithm.
5.1
INTRODUCTION
Audio signal processing systems have made considerable progress over the past 25
years due to increases in computational speed and memory capacity. These improve-
ments are a direct result of the ever increasing enhancements in silicon processing
technologies. These changes can be demonstrated by examining the implementation
of increasingly complex algorithms in less and less hardware. In this chapter, we will
describe how sound is digitized, analyzed and synthesized by various means. The
chapter proceeds from input to output with a historical bent.
