493

High-Performance IIR QMF Banks For Speech Subband Coding

Zhongnong Jiang, Abeer Alwan and Alan [Link], Jr.

Electrical Engineering Department, University of California, Los Angeles
Los Angeles, CA 90024

Abstract - In this paper, two high-performance imple- In certain applications, system delay is an impor-
mentations of IIR QMF banks for speech coding are tant issue. For example, the total delay for one-way
proposed. The first implementation involves a perfect communication should be as short as 1 ms (e.g., net-
reconstruction (PR) QMF bank while the second is a work telephony with no echo control). The current
tree-structured filter bank. In contrast to existing imple- CCITT standard sets the delay requirement at 5 ms; no
mentations of PR QMF banks, our approach does not existing FIR QMF bank can meet this requirement
require the transmission of initial conditions nor the (even if no signal quantizers and channel
design of synthesis filters which are more complex than coders/encoders are used). For example, Cox et al. [7]
their analysis counterparts. In systems with delay con- proposed a 4-channel FIR QMF bank which has a
straints, the tree structured IIR QMF bank can be used; delay of 7.5 ms at an 8 KHz sampling rate. In contrast,
a 4-channel scheme shows a low delay of about 1.9 ms the 4-channel tree-structured IIR QMF bank proposed
(15 samples) at an 8 kHz sampling rate. No existing here has a delay of only 1.9 ms. Although the use of
H R QMF banks can achieve this low delay. Moreover, conventional IIR QMF banks is accompanied with a
the phase distortion of the low-delay filter bank does price: phase non-linearity, it is possible to minimize
not appear to affect the perceptual quality of the pro- phase distortion such that it does not affect the percep-
cessed speech signals. Subjective tests were conducted tual quality of the speech signals.
to evaluate the speech quality.
2. Design of PR IIR QMF Banks
1. Introduction An M-channel PR IIR QMF bank consisting of
Although FIR quadrature mirror filter (QMF) allpass functions A, ( z ) and A, ( z - l ) , for 0 Ik IM-1 is
banks [ l ] are widely used for signal subband coding, depicted in Fig. 1 [4, 51 where M -point IDFT and DFT
IIR QMF banks are more efficient [2-51. For two- units are also included. Perfect reconstruction for this
channel QMF banks, stable and causal IIR filters can be IIR QMF bank is achieved based on the following iden-
used without causing aliasing errors and amplitude dis- tity
tortion, but phase distortion does occur [3]. If anti-
A, ( z ) A , ( z - ' ) = 1 0 Ik 5 M-1 (1)
causal IIR filters are included, perfect reconstruction
(PR) IIR QMF banks can be achieved. This approach
Obviously, if A ( z ) is stable and causal (we have
was first investigated for image compression [4]. More
dropped the subscripts for simplicity), A (z-') becomes
recently, efficient PR IIR QMF banks using all-pass
unstable and causal. However, we can also view A ( z - ' )
functions as building blocks for processing infinite
as stable and anti-causal. In principle, the implementa-
length speech signals have been proposed [5]. How-
tion of the anti-causal A ( z - ' ) can be accomplished by
ever, some problems arise. First, initial conditions of
reversing the input signal sequence in the time domain,
IIR synthesis filters must be transmitted; this increases
processing it through A ( z ) , and reversing the output
the transmission bit rate. Second, because of segmented
signal sequence. For finite-length signals, such as
processing the implementation uses anti-causal syn-
image signals, this filtering procedure can easily be
thesis filters which are considerably more complex than
realized by using memory circuits; but the reversion of
the analysis filters. In applications such as digital radio
infinite length signals is not feasible. Since A ( z ) is an
broadcasting, however, simpler decoders, rather than
allpass filter, the processing of infinite length signals,
encoders, are preferred. In this paper, a modified
such as speech, can be precisely implemented by using
scheme of the anti-causal IIR filtering approach investi-
a special segmentalized approach proposed in [SI. This
gated initially by Czamch 161 is proposed lo overcome
method is illustrated by the following steps:
these problems.
I) Assume that the input sequence ( x ( n ) ) is non-
zero for 0 I n IK-1 (a length-K sequence).
*This work was supported by the National Science
Foundation grants MIP 92-01 104 and IRI 93-0941 8
 494

2) Pass { x ( n ) } through A ( ? ) and obtain the output being processed. The systematic errors, however, can
sequence {U (n )]. be diminished if a modified segmentation is introduced
Make { v ( n ) = u ( K - n - l ) ) , and save the state- (see Fig. 3(@3(g)). Now we assign K+KI samples in
3)
variable vector p ( K ) of the firstA ( z ) . each data block with the first K1 samples taken from
the next data block (this can be done by invoking addi-
4) Pass { v ( n ) }through the second A ( z ) using the tional memory circuits). After each reversed data block
identical realization form as the first and the fol- passes through the filter, the first K , output samples are
lowing initial state-variable vector simply dropped (i.e., each output data block still has
q(0) = TP(K) (2) only K samples). This modification can greatly reduce
the systematic errors, since the most corrupted first K
where T is an exchange matrix with the main samples are eliminated. An obvious penalty of using
anti-diagonal elements being 1 (i.e., = 1) the above segmentalized processing is that an increase
and other elements equal to 0. in computation is necessary. Now we discuss how to
The above filtering procedure allows the output select K I to make systematic errors small. Assume that
sequence { Z ( n ) }to be a reversed replica of the input, A ( z ) is stable and the input signal sequence x ( n ) is
i.e., F(n) = x (K-n -1). For processing infinite-length bounded by some positive value Q . We then follow
sequences, a double-buffer structure to segment the the analysis given in [6] to find the upper bound for
input sequence into K-point data blocks can be used, systematic errors e (n) as
following the steps outlined above. Though this
appears to be an efficient realization, some obvious le(n)l S Q Ih(k)l (4)
problems exist. The values of p ( K ) in the first A ( z ) k=K,-l
must be saved and transmitted to the second. Further- As expected, the errors caused by the above modified
more, because A (z-') is involved in the synthesis stage, segmented processing approach are closely related to
the decoder is more complex than the encoder. the property of the impulse response of A ( z ) . The per-
To overcome these problems, we suggest a dif- centage of the overall increased computations in a PR
ferent anti-causal filtering approach, originally exam- IIR QMF bank will be 100Kl/2K. For example with
ined in [6]. We propose a modified (properly reversed) K ,= K / 2 , this percentage will be only 25.
PR IIR QMF bank, as shown in Fig. 2. The correspond- In many cases, efficient first-order allpass func-
ing perfect reconstruction conditions become a + z-'
tions of the form A ( z ) =~ with la1 < 1 can be
Ak (z-I)Ak(z)= 1 0 5 k w - 1 (3) 1+aZ-'
used [4]. The upper bound of the systematic errors is
Note that the anti-causal A ( z - I ) , realized as in Fig.
3(a), is now used in the analysis filters rather than the Upper Bound = (1 + a ) l a l K ' Q (5)
synthesis filters. Zero initial conditions are employed a2
for both analysis and synthesis filters. If the simple seg- As an example, consider an 8-channel IIR QMF
mentation from [5] and zero conditions are used, "cor- bank given in [6]. The all-pass filter coefficients are:
rupted" output signals will occur. This problem, how- % = 0.1038, a,= 0.2078, Or, = 0.3144,
ever, can greatly be alleviated by using a modified
a3 = 0.4260, a, = 0.5458, = 0.6774, 06 = 0.8264
scheme of the anti-causal IIR filtering approach from
[6], although systematic errors still exist. Specifically, Note that channel 7 is a direct connection. For channel
the input signal sequence x (n ) with infinite length can 6, if we want the systematic errors to be smaller than
be partitioned into a series of data blocks denoted as 104Q, K, should be chosen as 64 and the upper bound
x j ( n ) , where j is the block index and each data block value is 1 . 3 ~ 1 0 - ~ QIf. all other channels choose
consists of K signal samples. Fig. 3(b)-3(e) show that K , = 32, the upper bound value is merely 1 . 4 ~ 1 0 - ~ Q .
three data blocks xj-,(n), x j ( n ) and x j + , ( n ) are
reversed. They are then fed into A ( z ), the resulting out- 3. Design of Low-Delay IIR QMF Banks
put data blocks, uj-'(n), u j ( n ) and uj+'(n),are also A conventional IIR QMF bank usually uses
segmentally reversed. For simplicity we have made the half-band IIR filters as building blocks [4]. A tree-
input and output data blocks exhibit the same shape. structured 4-channel IIR QMF bank with an efficient
After carefully studying this filtering process, we find realization is shown in Fig. 4, where Ao(z) and A , ( z )
that the resulting output data blocks are corrupted by are allpass filters and the output X ( z ) has the following
systematic errors which result from the fact that the relationship with X ( z ) [ 11
contribution of future data blocks, ~ ~ + ~~( ~n +) ,~ ...,
(n) ,
are simply ignored because they are not available when %z ) = D ( 2 ) X ( Z ) (6)
the reversed data block xi (-n ) resulting from xi ( n ) is
where D ( z ) = -A0(z2)A
z -3 I(z2)Ao(z4)A1 ( z 4 )is referred
4
 495

to as the distortion function of an IIR QMF bank. noticeable. Hence, it seems that our filter bank designs
In designing practical QMF banks for speech can be used for high-performance speech applications.
subband coding, it is desirable to have filters with at
least 40 dB stopband attenuation [7]. Linear-phase FIR 5. Summary
filters with this performance have between 24 and 32 Two high-performance IIR QMF bank structures
taps. After designing several 7th-order half-band IIR for speech subband coding have been investigated. The
filters with stopband edge frequency w, varying up to first filter bank structure is a new implementation of a
0 . 7 2 ~(wp = x - q),we found that phase distortions PR QMF bank, which requires lower transmission rate
change slowly if w,$ > 0.61~.Hence, our filter banks and uses simpler encoders than existing PR IIR QMF
were designed around w, =0.61~. We designed four banks. Another filter bank scheme employs conven-
IIR QMF banks with w.r =0.581~, 0.6oJl, 0.62n, 0.641~ tional half-band IIR filters in a tree structure so as to
and stopband attenuations varied from 48 to 62 dB. achieve very low system delays. Although phase distor-
tion is unavoidable, the subjective tests show that phase
System delay of an IIR filter is difficult to define
distortion does not affect the perceptual quality of the
since the delay is signal dependent. We decided to
processed speech. We plan to investigate further the
adopt two measures for the system delay: the partial
use of these filter banks in speech subband coding
energy of the impulse response and the position of the
applications and consider the effects of vector quanti-
peak value of the impulse response. The partial energy zation on the system performance.
measure E ( n ) was first introduced in [8]:
n
E(n)= c. Id("*,
m=U
(7) References

where d ( n ) is the impulse response of D ( z ). To be

consistent with delay measurement of linear-phase FIR P. P. Vaidyanathan, Multirate Systems and
Filters Banks. Prentice Hall, Englewood Cliffs,
filters, we define the delay of an IIR filter to be
NJ, 1992.
equivalent to the number of samples at which 50% of
the impulse response energy appears. An impulse con- T. A. Ramstad, "Sub-band coder design using
sists of all frequency components with the same magni- quadrature mirror filters," Proc. EUSIPCO, pp.
tude and, thus, E ( n ) appears to be a measure of the 747-752, 1980.
average delay of the system. R. Ansari, "Elliptic filter design for a class of
Another measure of system delay is the location generalized halfband filters," IEEE Trans.
of the peak value of the impulse response. This meas- Acoust. , Speech, Signal Processing, vol.
ure is used for linear-phase FIR filters and is similar to ASSP-33, pp. 1146-1 150, Oct. 1985.
the 50% energy delay measured by E ( n ) since the J. Husoy and T. Ramstad, "Application of an
impulse response of a linear-phase FIR filter is either efficient parallel IIR filter bank for image sub-
symmetrical or anti-symmetrical. band coding," Signal Processing, vol. 2, pp.
System delays of our filter banks, measured with 279-292, Aug. 1990.
the two functions described above, were less than 15 S. K. Mitra et al., "A novel implementation of
samples (1.9 ms, 8 kHz sampling rate). If phase distor- perfect reconstruction QMF banks using IIR
tion does not affect the quality of the processed speech, filters for infinite length signals," Proc. IEEE
then these filter banks are optimal in terms of computa- Symp. Circuits and Systems, pp. 2312-2315,
tional efficiency and system delay. San Diego, May 1992.
4. System Performance R. Czarnch, "Recursive processing by noncausal
digital filters," IEEE Trans. Acoust. , Speech,
Subjective tests were conducted to evaluate the Signal Processing, vol. ASSP-30, pp. 363-370,
perceptual quality of the processed speech signal. June 1982.
Three sentences, spoken by two female talkers and a
male talker, were sampled at 8 kHz using 16-bit preci- R. V. Cox et al., "New directions in subband
sion. The sentences were processed with and without coding," IEEE J. Select. Areas Commun. (Spe-
noise. In the 'noisy' simulations, white Gaussian noise cial Issue on Voice Coding for Communica-
was added to the sentences prior to the filtering. Two tions). vol. 6, pp. 391-409, Feb. 1988.
SNRs were used: I O and 15 dB. A. V. Oppenheim and R. W. Schafer, Discrete-
The subjective tests show that the systematic Time Signal Processing. Prentice Hall, Engle-
errors of the PR QMF bank and the phase distortion of wood Cliffs, NJ, 1989.
the low-delay filter banks are not (perceptually)
 496

I - I

e I

Fig. 1. An M-channel perfect reconstruction IIR QMF bank.

I e
I a a I

I T 2-l

K- 1 0

Fig. 2. A modified M-channel PR IIR QMF bank. Fig. 3. A modified implementation scheme for A(z").

Fig. 4. A tree-structured 4-channel conventional IIR QMF bank.