Digital Signal Processing Reference
In-Depth Information
2.
EMPLOYED SPEECH ENHANCEMENT
SYSTEM
The employed speech enhancement system is an oversampled
Generalized DFT (GDFT) filterbank using a subband processing block
consisting of a Subband Adaptive Filter (SAF) and a Wiener filtering sub-
block to reduce noise in each subband. A Voice Activity Detector (VAD) is
used to control both the adaptation in the SAF, and the noise spectrum
estimation in the Wiener filter. The complete speech enhancement system
(the SAFs, the Wiener filter, and the VAD) is efficiently implemented on an
ultra-low resource oversampled WOLA filterbank detailed in [5]. We now
describe various components of the enhancement system.
2.1
Subband Adaptive Filters for Noise Cancellation
Subband Adaptive Filters have become a viable choice for adaptive noise
and echo cancellation. The SAF approach employs filterbanks to split time-
domain inputs into a number of frequency bands, each serving as input to an
adaptive filter. Subband signals possess reduced spectral dynamics and, due
to their narrower bandwidth, may be decimated. Subband decomposition and
decimation thus result in much “whiter” signals as input to a parallel bank of
much shorter adaptive filters with better convergence behavior [6]. If critical
sampling is used, aliasing distortion occurs that may be eliminated by
employing either adaptive cross-filters between adjacent subbands or gap
filters [6,7]. Systems with cross-filters generally converge slower and have
higher computational cost, while systems employing gap filters produce
significant signal distortion. OverSampled SAF (OS-SAF) systems, on the
other hand, offer a simplified structure that, without employing cross-filters
or gap filters, significantly reduce the aliasing level in subbands. In an
attempt to minimize additional computational cost, a non-integer
oversampling factor close to one is sometimes used. For many low-
delay real-time applications including adaptive noise and echo cancellation,
however, higher oversampling factors permit wide-range subband gain
adjustments and the use of shorter windows with less stringent requirements
[5]. Consequently, to avoid aliasing and other distortions, the solution of
choice is to use an oversampling factor of two or more. However,
oversampling degrades the convergence behavior of SAF systems (due to
coloration of the subband signal) when the Normalized Least Mean Square
(NLMS) algorithm is employed.
To improve the convergence rate and computation complexity of OS-
SAF systems, we have proposed convergence improvement techniques [8]
