CS6422
CS6422
characteristics than white noise because of its qua-
si-periodic nature.
efit of suppressing the spurious taps mentioned in
Section 4.1.1.2.1, “Pre-Emphasis”.
Research at Crystal has shown that quasi-periodic
signals cause the formation of spurious non-zero
coefficients within the adaptive filter at tap inter-
vals determined by the periodicity of the signal.
The Microcontroller Interface allows four settings
for graded beta: none, 0.19 dB/ms, 0.38 dB/ms,
and 0.75 dB/ms. Use 0.75 dB/ms for acoustically
dead rooms or cars, and 0.19 dB/ms or no grading
This results in small changes in period being very of beta for large, or acoustically live rooms.
destructive to the adaptive filter’s performance.
4.1.1.3 Update Control
One mechanism the CS6422 uses to prevent this
filter corruption with speech is to pre-emphasize
As mentioned in Section 4.1.1.1, “Theory of Oper-
ation”, the update control algorithms are the heart
the signal sent to the adaptive filter so that much of
of any useful echo canceller implementation. Aside
the low frequency content is removed.
from telling the adaptive filter when to adapt, they
The CS6422 works very well with a speech training
are responsible for correcting performance when
signal because of the pre-emphasis filter. White
the path changes more quickly than the filter can
noise training signals, however, result in sub-opti-
respond. For example, if the adaptive filter is actu-
mal performance, so when testing with white noise,
ally adding signal power instead of cancelling it,
it is recommended that the pre-emphasis filters be
the update control algorithms will reset the adap-
disabled.
tive filter to cleared coefficients, forcing it to re-
start.
4.1.1.2.2 Graded Beta
4.1.1.4 Speech Detection
The update gain of an adaptive filter, sometimes
called the “beta”, is the rate at which the filter co-
efficients can change. If beta is too low, the adap-
tive filter will be slow to adapt. Conversely, if it is
too high, the filter will be unstable and will create
unwanted noise in the system.
The CS6422 detects speech by using power estima-
tors to track deviations from a background noise
power level. The power estimators filter and aver-
age the raw incoming samples from the ADC.
A background noise level is established by a regis-
ter that increases 3 dB at intervals determined by
NseRmp (Register 2, bits 11 and 10). When the
power estimator level rises, the background noise
level will slowly increase to try to match it. When
the power estimator level is below the background
noise level, the background noise level adjusts
quickly to match the power estimator level. This
method allows significant flexibility in tracking the
background noise level.
In most echo canceller implementations, the beta is
a fixed value for all the filter coefficients. In some
situations, though, through knowledge of the char-
acteristics of echo path response, the beta can be
varied for groups of coefficients. This preserves
stability by allowing the beta to be higher for some
coefficients and compensating by reducing beta be-
low nominal for others.
For example, acoustic echo tends to decay expo-
nentially, so the first taps need to be larger than the
later taps. Having a beta profile that matches the
expected response path enhances the echo cancel-
ler’s ability to correctly and accurately model the
acoustic path. Furthermore, this has an added ben-
Speech is detected when the power estimator level
rises above the background noise level by a given
threshold. The half-duplex receive speech detector
threshold is set by RHDet (Register 2, bits 15 and
14), the half-duplex transmit speech detector
threshold is set by THDet (Register 1, bits 15 and
34
34
DS295F1