IEEE Signal Processing - May 2018 - 98

Introduction
Speech coding is an essential technology in information transmission and communication systems. Human cognitive processing operates at about 50 bits/s, which corresponds roughly
to the speech production semantics as the rate of phonemic
information in speech (e.g., most languages have approximately 32 phonemes encoded with 5 bits, and 1 s of speech
has perhaps ten phonemes), and the sensory system is known
to encode nonredundant structures [1]. Efficient coding maximizes the amount of information conveyed about the sensory
signal to the rest of the brain. The incoming acoustic signal is
transmitted mechanically to the inner ear and undergoes a
highly complex transformation before it is encoded efficiently
by spikes at the auditory nerve. This great efficiency in information representation has inspired speech engineers to incorporate aspects of cognitive processing when developing efficient
speech technologies.
Speech coding is a field in which research has slowed considerably in recent years. This has occurred not because it has
achieved the ultimate in minimizing the bit rate for transparent speech quality, but because recent improvements have been
small and commercial applications (e.g., cell phones) have been
mostly satisfactory for the general public-by the same token,
the growth of available bandwidth has reduced requirements to
compress speech even further. However, better compression is
always desirable (e.g., in large archival systems).
Historically, the mechanisms of speech perception have led
to perceptual speech coding [2], primarily suitable for digital
audio. Substantial progress in this context incorporates mechanisms to optimize coder performance for the human ear in

the context of subband (transform) coders. On the other hand,
the most common speech coding for medium-to-low bit rates
is based on models of human speech production, realized as
linear predictive vocoders, and analysis-by-synthesis linear
predictive coders. Unified audio and speech coding is usually
realized with realtime switching according to the input signal
type. For example, the Enhanced Voice Services (EVS) coder
standardized in 2015 by the 3rd Generation Partnership Project
(3GPP) offers new features and improvements for low-delay,
real-time communication systems, higher quality for clean/
noisy speech, mixed content, and music, including support for
wideband, superwideband, and fullband content [3]. However,
the core speech compression method is algebraic code-excited
linear prediction (ACELP) proposed in 1987 by Adoul et al.
[4]. Thus, the compression paradigm has not changed significantly in the last 30 years and has its roots in linear predictive
theory dating back to the early 1940s [5], [6].
Research studies conducted in the last decade have incorporated additional aspects of speech processing, i.e., the
functional and temporal organization of human speech and
language processing. Figure 1 shows the overall speech perception process. Speech signals are treated indiscriminately
by subcortical structures from other types of acoustic input.
Such processes have now been fairly well characterized,
specifically the cochlea, which is decomposed into different
frequency channels forming the auditory spectrogram [7].
Automatic speech processing is derived much less from the
subsequent stages in the auditory cortex and, in particular,
omits how this continuous representation is transformed into a
discrete representation (i.e., a lexicon). Such a transformation

Speech Sound

Auditory Spectogram

Cochlea

- Continuous Signal
- Speaker Sensitive
- Speech-Rate Sensitive
Auditory Cortex
* Speech perception must deal with different
timescales corresponding to phonemic, syllabic,
and prosodic temporal modulations.

Lexicon
/k/ə/m/h/o/m/r/a/j/t/ə/w/e/
- Discrete Representation
- Speaker Insensitive
- Speech-Rate Insensitive

Figure 1. The overall speech perception process carried out by the peripheral (cochlea) and the central auditory (primary auditory cortex) systems. The
green and blue boxes with corresponding vertically spaced lines represent syllabic and phonetic speech segmentation, respectively.

IEEE Signal Processing Magazine

May 2018

Table of Contents for the Digital Edition of IEEE Signal Processing - May 2018