IEEE Signal Processing - May 2018 - 103

(and nonintrusively without a reference) evaluate the quality of transmitted speech. For example, the recent perceptual objective listening quality assessment (POLQA)
method [65], standardized as ITU-T Rec. P.863, includes a
perceptual model based on both the spectral and temporal
masking effects of human hearing and cognitive modeling.
While the perceptual model based on a gammatone filter
bank determines which distortions can be perceived by listeners, the cognitive model predicts the level of those distortions. The output of the cognitive model is the absolute
category rating (i.e., overall quality score), which reflects
the opinion of an average listener who is used to using commercial telephony services.

Production model-based compression
Perceptual audio coding uses models of human auditory perception, whereas speech coding is traditionally based on the
human vocal tract (speech production) model. Though the
band-limited wired and wireless communication systems have
changed dramatically from analog to digital, the paradigm of
speech coding has remained the same (i.e., based on the waveform and the linear prediction model) [5]. Linear predictive
coding (LPC) is used in the majority of standardized higher
bit rate [3], [59] and lower bit-rate speech coding [66-71].
LPC also uses certain models of human auditory perception,
i.e., perceptual weighting of the residual quantization error
and adaptive postfiltering [57] to minimize different types of
auditory distortion. LPC is usually realized as an analysis-bysynthesis system that selects an excitation signal from among a
large set of candidates in a closed-loop manner. In other words,
speech coders include decoded feedback during encoding.
As introduced in the "Human Cognitive Speech Processing"
section, cortical speech-processing mechanisms also include
feedback decoding processes. Here, the similarity with human
cortical speech production ends. Rather, machine-speech compression is inspired by the physiological process of speech production based on the source-filter theory.
Further connections with human production exist in sparse
representations that contribute significantly to a low computational complexity. In 1986, sparse excitation was proposed for
CELP [72] as a complexity reduction method; the speech source
defined as a codebook populated with pseudo-random white
sequences (Gaussian excitation vectors) was sparse, in terms of
the number of nonzero pulses for voiced speech. In a typical
5-ms frame period (or subframe period depending on the CELP
variant), only about 10% of the pulses were set to a value other

than zero. In 1990, Laflamme et al. introduced sparse algebraic codes (with few nonzero components) for fast searching of
the codebooks [73] to get a minimum variance residual with
an analysis-by-synthesis scheme. Most current standardized
speech coders are based on this ACELP. Alternatively, a sparse
LP residual is defined within a compressive sampling framework [74], introduced in the "Machine CSC" section.

Switched audio/speech coding
Current hybrid coding approaches offer real-time switching
between perceptual coding for audio sources, and ACELP
coding for speech sources. Today, two switching coders are
popular: 1) Opus [75], the open-source codec of the Internet
Engineering Task Force that includes speech coding technology from Skype's SILK codec and audio coding technology
from CELT codec (http://celt-codec.org), and 2) the EVS [3]
codec of the 3GPP. Recent subjective evaluation demonstrates
that the 3GPP EVS codec, compared to Opus, provides the
same quality at about half of the bit rate in low bit rates [60].
Recent codecs offer new features and improvements for
low-delay, realtime communication systems, higher quality for
both clean and noisy speech, and mixed content and music,
including wideband, superwideband, and fullband content support. However, as proposed by Adoul et al. in 1987 [4], ACELP
is the core compression method. Thus, the compression paradigm has not change significantly in the last 30 years.

Machine CSC
Human cognitive speech processing involves transforming sensory inputs in both feed-forward (bottom-up) and feedback (topdown) processes. LPC also involves open-loop (feed-forward)
processing for calculation of the gross spectral shape 1/A (z) and
closed-loop (feedback) processing for calculation of the excitation signal U (z) that models the fine spectral structures, as the LP
model represents the speech signal S (z) as a linear time-invariant
system with the following transfer function:
S (z) =

U (z)
=
A (z)

U (z)
p

1 - / ai z

.

(1)

-i

i =1

The input to the system is U (z), and p pairs of complex-conjugate poles of A (z) represent formant frequencies (the spectral peaks of the sound spectrum).
Table 1 compares LPC and CSC. The main differences lie
in the underlying model used: LPC is a linear time-invariant

Table 1. A comparison of LPC and CSC.
Condition

Linear Predictive Coding

CSC

Speech representation

Formants and vocal tract excitation

Articulatory, auditory, and somatosensory targets [38]

Models

Electrical circuits and digital filters

Deep and spiking NNs

Temporal context

Synchronous frame-subframe structure

Asynchronous streams [13]

Sparsity

Excitation signal [73]

Whole speech representation

IEEE Signal Processing Magazine

|

May 2018

|

103
http://www.celt-codec.org
Table of Contents for the Digital Edition of IEEE Signal Processing - May 2018
