VERY LOW DELAY AND HIGH QUALITY CODING OF
20 HZ - 15 KHZ SPEECH SIGNALS AT 64 KBIT/S
C. Murgia
1
, G. Feng
1
, A. Le Guyader
2
& C. Quinquis
2
murgia@icp.grenet.fr
1
Institut de la Communication Parlée, URA CNRS n° 368, INPG/ENSERG,
Université Stendhal, GRENOBLE, FRANCE
2
CNET Lannion A, TSS/CMC, 22301 LANNION, FRANCE
ABSTRACT
In this paper, an algorithm for coding 20 Hz - 15 kHz speech
signals at 64 kbit/s with a very low delay (frame of 0.16 ms) is
presented. To achieve a quality near to transparency, we propose
adapting the Low-Delay CELP coder [1] to the 15 kHz bandwidth
and suggest a new noise shaping method based on a psycho-
acoustic model. In this way we take advantage of linear predictive
coding and masking properties of the human perception system.
Finally, an algebraic codebook is proposed, allowing an important
reduction of coder computational complexity, without decreasing
the perceived quality of signals.
1. INTRODUCTION
High quality speech coding plays an important role in modern
telecommunication systems, especially in videoconference and
teleconference systems. Compared with narrow band
(300 Hz - 3400 Hz), or wideband (50 Hz - 7 kHz) speech, the
expansion to the 20 Hz - 15 kHz bandwidth increases significantly
the quality of the signals, and improves the sensation of remote
speaker presence. To ensure good interactivity in the
communication systems, the one-way delay should not exceed
100 ms including processing, sound recording and transmission
delays, leaving only about 20 ms for the encoder-decoder delay.
To achieve these aims, we propose a high quality coding
algorithm at 64 kbit/s for the 20 Hz - 15 kHz speech signal. The
starting point of our research was the Low-Delay Code-Excited
Linear Prediction method (LD-CELP), originally developed for
narrow band signals [1]. Several problems (filter stability, coding
noise, etc.) arise from the extension of this algorithm to the
20 Hz - 15 kHz bandwidth [2]. In this paper, we suggest a new
noise-shaping method based on a psycho-acoustic model of the
human perceptual system. Moreover, in order to reduce the
complexity of this algorithm, we propose the use of an ternary
algebraic codebook.
2. SYSTEM OVERVIEW
The proposed coding system is based on backward adaptive CELP
coding [1]. The basic principles of this algorithm are summarised
as follows. The LPC coefficients are updated at the encoder (local
decoder) and at the decoder by backward adaptive linear
prediction on the previously synthesised signal. Only the
excitation parameters are transmitted to the decoder. The
excitation signal is a 5-sample vector, selected from a 7 bits shape
codebook using the analysis by synthesis method and scaled by a
gain factor quantized on 3 bits. With this backward LPC method,
the buffering delay is only 5 samples, i. e. the excitation frame
length. The direct extension of the LD-CELP scheme to the
20 Hz - 15 kHz bandwidth gives rise to degradations, especially
coding noise and some defaults due to filter instabilities. A
detailed analysis of these problems shows that the role of some
coder components has to be reconsidered. The necessity for the re-
optimisation of the coder parameters was confirmed by informal
listening tests.
3. SYNTHESIS FILTER ORDER
SELECTION
An important parameter of the system is the synthesis filter order.
In the LD-CELP coding system, its value was set to 50 so that the
spectral envelope and also the harmonic structure of high pitched
signals could be modelled. To have the same prediction gain for a
sampling rate of 32 kHz, the filter order should be equal to 200. In
addition to the computation complexity caused by such a high
order, we must solve the problem of the filter instability. In order
to determine the optimal filter order, we measure the prediction
gain depending on the filter order. Figure 1 shows an example of
the experiment results. Generally, for male voices saturation
occurs for a filter order greater than 60, although for the female
voices this behaviour is less marked.
[dB]
prediction order
males
mean
females
Figure 1: Behaviour of the prediction gain versus the synthesis
filter order for a corpus of six male and female voices.
According to these results, we carried out a series of simulations
of the coder using four different synthesis filter orders. Figure 2
illustrates the evolution of the segmental signal-to-noise-ratio
(SNRseg) depending on the filter order. At an order of greater
than 64 the SNRseg continues to increase slowly for music
signals, but we observe a slight decrease for speech.
Thus the synthesis filter order is fixed at 64, which appears to be a
fair compromise between a good spectral resolution and a
reasonable algorithm complexity.
22
23
24
25
40 50 64 80 100
speech
music
mean
prediction order
[dB]
Figure 2: Behaviour of segmental signal to noise ratio depending
on synthesis filter order for a corpus of ten speech and music
signals.
4. RESTITUTION OF SIGNAL
PERIODICITY
In our coder, the size of the excitation vector is 5 samples, which
corresponds to a temporal resolution of 0.16 ms for a 32 kHz
sampling frequency. This choice allows, by adapting the gain and
selecting the optimum excitation vectors, the reproduction of the
details and the periodicity of the linear predictive residual, without
requiring a too high filter order. Therefore, all the fundamental
speech frequencies can be correctly modelled. This phenomenon
is illustrated in figure 3 for a female voice signal. We notice that
the short term excitation presents periodic peaks fairly
synchronised with the original signal pitch.
In our coder, the modelisation of the spectral envelope is provided
by the synthesis filter, whereas the pitch modelisation is mainly
ensured by the choice of the optimum vector and their gains.
10 ms
Figure 3: Modelisation of the filter synthesis excitation. From top
to bottom : the original signal (F
0
≈ 280 Hz), the synthesised
signal, the excitation of synthesis filter (order 64).
5. OPTIMUM PERCEPTUAL FILTER FOR
NOISE SHAPING
In CELP coders, the use of a perceptual filter allows the shaping
the coding noise spectrum, so that it could be perceptually masked
by the signal. This filter has the form:
()
()()
Wz Az Az= // /γγ
12
(1)
where 1/A(z) is the synthesis filter. By varying γ
1
and γ
2
, it is not
possible to control both the shape and the tilt of the perceptual
filter spectrum on a wide bandwidth with a large dynamic range.
For this reason, an improvement of the spectral modelisation has
been suggested: W(z) is used in cascade with another filter
controlling the spectrum tilt [3, 4]. In addition, this form of filter
is only partially able to take into account the characteristics of
human hearing system since it comes from an AR model of the
speech production process.
[dB]
Hz
Figure 4: Speech signal spectrum and its optimum masking shape.
To solve this noise shaping problem for the 20 Hz - 15 Hz
bandwidth, we propose a new optimal method for noise shaping
based on a psycho-acoustic model [6, 7]. In order to avoid
introducing any additional delay, we determine the perceptual
filter from the previous input signal samples. At the beginning of a
new LPC frame (every 2.5 ms), a masking shape is calculated on
the last 1024 samples. This block length is necessary to guarantee
a good spectral resolution. The samples are weighted by a hybrid
analysis window, in which the more recent are weighted by a
cosine window and the others by an exponential window. This
increases the importance of the most recent samples in
determining the filter coefficients. Furthermore, its spectral
characteristics are comparable to those of the Hamming window
[5]. The power spectrum of this weighted signal is calculated
using the FFT algorithm. Then the optimum masking shape is
obtained by the convolution (in the Bark domain) of this spectrum
with the basilar membrane spreading function [6, 7]. The
autocorrelation coefficients of the masking shape are obtained by
an inverse FFT. Finally, the AR model is determined using the
Levinson-Durbin algorithm.
a)
c)
b)
[dB]
Hz
Figure 5: Behaviour of LPC spectra: a) synthesis filter;
b) perceptual weighting filter determined in the classical way with
tilt correction; c) perceptual filter determined from the optimum
masking shape.
It should be noted that the proposed method does not introduce
any additional delay, since the filter coefficients are determined
from the input signal preceding the current frame.
Figure 4 shows a speech signal spectrum and its optimum masking
shape obtained using the algorithm described in [6] and [7].
Figure 5 shows the LPC spectrum of the synthesis filter, the
spectrum of the classical noise-weighting filter, with tilt
correction, and finally the spectrum of the perceptual filter,
determined from the optimum masking shape. We observe that the
new LPC perceptual filter is a good model of the optimum
masking shape. We also notice that the classical weighting filter
spectrum is very far from the optimum coding noise shape, in the
perceptual sense. In fact, between 1 kHz and 3 kHz the frequential
response of the classical weighting filter is above the optimum
perceptual shape, which makes the quantization noise more
audible. Moreover, between 6 kHz and 8 kHz, and beyond 11
kHz, its frequency response is well below the auditory threshold.
A re-optimisation of classical filter parameters (γ
1
, γ
2
) and tilt
control filter parameters does not achieve optimum noise shaping.
6. BIT ALLOCATION AND CODEBOOK
OPTIMISATION
This procedure is fully backward adaptive, so only the excitation
indexes are transmitted to the decoder. The shape codebook index
is coded with 7 bits and the gain codebook index with 3 bits. With
a 5-sample frame and at the sampling frequency of 32 kHz the bit
rate is 64 kbit/s.
The shape codebook has been designed using a closed-loop
iterative procedure [8]. Figure 6 shows the results of the codebook
design. Even if the W-SNRseg gain is only 0.5 dB - the starting
codebook has already high performances - the increase in
perceived quality is very significant. Informal listening confirms
these results. The original signal is first presented to the listeners
and followed - in a random order - by two synthesised signals
coded with or without the optimised codebook. The listeners
found the signal coded with the newly designed codebook as the
nearest to the original one in 70% of cases.
-0,5
-0,4
-0,3
-0,2
-0,1
0
0123456789
[dB]
loops
Figure 6: Results of the shape codebook optimisation design:
behaviour of average weighted mean-square error depending on
the optimisation loops.
7. CODER COMPLEXITY REDUCTION
In CELP coders the most computationally complex block is the
analysis-by-synthesis research of the optimum code word in the
codebook. The index of the optimum code vector is obtained by
minimising the following distortion expression:
( )
D c Hc t Hc
k k k
= −
2
2 ,
(2)
where c
k
is the vector of index k, t the target vector, H the lower
triangular matrix of the truncated impulse response of the cascade
of the synthesis and perceptual filters and the symbol < , >
represent the inner product. In this research procedure, for every
LPC frame, the codebook vectors are filtered and in this way
adjusted depending on the signal characteristics. This filtering
operation implies an important computational load in the research
procedure. The introduction of a ternary algebraic codebook
reduces the search computational complexity [9]. This algebraic
code is generated using modulo-3 arithmetic with five digits and
by using the following transposition: {0, 1, 2} ⇔ {-1, 0, +1}. In
this way the code-vectors with index k = 3
i
, only have a non-zero
pulse of -1 in the position i. If S(k) is the last component of the
vector Hc
k
, it can be demonstrated that:
( )
( )
( )
( ) ( )
( )
( ) ( )
S h i
S j h i S j
S j h i S j
i
i
i
3
3
3
= −
− = − −
+ = − +
(3)
where 1 ≤ j≤ (3
i
-1)/2 and 0 ≤ i < 5. We observe that the
code-vectors are obtained by shifting to the left and by a link with
one of the ternary symbols. Thus, the energy E
k
= ||Hc
k
||
2
can be
calculated using the expression:
()
EE k
kk
S
=+
+int( / .5)30
2
(4)
where int(x) is the smallest integer that is less than or equal to x
and for 1 ≤ k ≤ (3
5
-1)/2. To complete the computation of the
distortion the following equation can be applied:
()
()
()
() ( )
()
() ( )
corr v i
corr j v i corr j
corr j v i corr j
i
i
i
3
3
3
=−
−=− −
+=− +
(5)
where 1 ≤ j≤ (3
i
-1)/2 and 0 ≤ i < 5 and corr(k) = <t, Hc
k
> =
<H
T
t, c
k
> and v=H
T
t. The computational gain obtained applying
this algorithm is 28% of the total coding time.
Table I summarises the SNRseg measure on a corpus of speech
and music signals when we use the optimised compared to the
algebraic codebook in the analysis-by-synthesis procedure. We
notice that the SNRseg is slightly degraded in the case of the
ternary algebraic codebook, but informal listening tests show that
the perceptual difference is insignificant.
[dB]
Speech
Music Total
Optimised
23.41
24.34 23.79
Algebraic
22.90
23.96 23.33
Gain
-0.51
-0.38 -0.46
Table I: Segmental SNR obtained on a corpus of speech and
music signals using two optimised versions of the coder with:
optimised and algebraic codebook.
8. RESULTS
Two versions of the coder have been tested and all the parameters
(synthesis filter order, excitation vector size, shape codebook)
have been optimised. In the first version, we use the classical
weighting filter, and in the second, the perceptual filter
determined from the optimum masking shape. The segmental SNR
values for a corpus of speech and music signals, synthesised by
the two coders are shown in table II. By using the new perceptual
filter, the segmental SNR increases by about 1.5 dB. This gain
appears to come from a better noise shaping, especially in low
frequencies, as illustrated in figure 4 and figure 5. Our informal
listening tests confirm the quality improvement due to this new
optimum-shaping method. In particular, a significant increase in
the clearness of the synthesised signals is obtained.
[dB]
Speech
Music Total
A)
23.41
24.34 23.78
B)
24.62
26.18 25.24
Gain
+1.21
+1.84 +1.46
Table II: Segmental SNR obtained using two optimised versions
of the coder with: A) classical weighting filter and tilt correction
filter; B) perceptual filter using the optimum masking shape.
CONCLUSION
We have proposed a new 64 kbit/s coding algorithm of
20 Hz - 15 kHz speech, with a very low delay of 0.6 ms. The
quality obtained is quasi-transparent. We have used a new
optimum-noise shaping filter based on a psycho-acoustic model of
human perception system. This new filter improve significantly
the speech quality. Informal listening tests on a large speech
corpus and some music signals show that the coding quality is
close to transparency. Finally, a modified algorithm has been
proposed, allowing an important reduction of coder computational
complexity, without decreasing the perceived quality of signals.
CD-ROM SOUND LINKS
Original female speech signal: [SOUND A217S1.WAV].
Coded female speech signal: [SOUND A217S2.WAV].
Original male speech signal: [SOUND A217S3.WAV].
Coded male speech signal: [SOUND A217S4.WAV].
REFERENCES
1. J.-H. Chen, R. V. Cox, Y.-C. Lin, N. Jayant & M. J.
Melchner, " A Low-Delay CELP Coder for the CCITT
16 kb/s Speech Coding Standard " - IEEE Jour. on Selec.
Areas in Comm., vol. 10, n° 5, pp. 830-849, 1992.
2. C. Murgia, G. Feng, C. Quinquis & A. Le Guyader, " Very
Low Delay an High Quality Coding of 20 Hz - 15 kHz
speech at 64 kbit/s" - Proc. of EUROSPEECH ‘95, pp.
37-40.
3. E. Ordentlich & Y. Shoham, " Low-Delay Code-Excited
Linear-Predictive Coding of Wideband Speech at 32
kbps " - Proc. of ICASSP ‘91, pp. 9-12, 1991.
4. O. Gottesman & Y. Shoham, " Real -Time Implementation
of High-Quality 32 kbps Wideband LD-CELP Coder " -
Proc. of EUROSPEECH ‘93, pp. 1115-1118, 1993.
5. J.-H. Chen, Y.-C. Lin & R. V. Cox, " A Fixed - Point
16 kb/s LD-CELP Algorihm " - Proc. of ICASSP ‘91, pp.
21-24, 1991.
6. Y. Mahieux & J. P. Petit, " Transform Coding of Audio
Signals at 64 kbit/s " - Globecom ‘90, pp. 518-522, 1990.
7. Y. Mahieux & J. P. Petit, " High-Quality Audio Transform
Coding at 64 kbps " - IEEE Trans. on Comm., vol. 42, n°
11, pp. 3010-3019, 1994.
8. G. Davidson, M. Yong & A. Gersho, " Real-Time Vector
Excitation Coding of Speech at 4800 bps " - Proc. of
ICASSP ‘87, pp. 2189-2192, 1987.
9. R. Di Francesco, " Codage Algébrique de la parole :
prédiction linéaire à excitation par code ternaire " - Ann.
Télécommun. 47, n° 5-6, pp. 214-226, 1992.
