Digital Signal Processing Reference
In-Depth Information
estimate
a.
However, now
A
(
k
) is no longer constant for all
k.
This means
that each sub-band has to be used. Unfortunately, the separation will perform
better in some bands than the others. This is a reality that cannot be escaped.
In summary, the steps involved in the BSS of underdetermined convolutive
mixture are:
1.
2
.
Compute the FT (short) of the observed mixtures.
For each frequency sub-band
(a) Obtain an initial estimate of
A
matrix using the procedure
described above.
(b)Apply the “dual update” algorithm iteratively to refine the
estimates of
A
and
S
.
3.
Find the appropriate permutation of sources in each frequency band
using the final estimates of
A
and
S
that are obtained after the “dual
update” algorithm converged.
Obtain the separated source signals and the final estimate of
a
by
applying the inverse FT.
4.
A block diagram of the proposed algorithm for BSS of underdetermined
convolutive mixtures is provided in Figure 18-2.
3.
IN-VEHICLE DATA COLLECTION DETAILS
A linear microphone array of five microphones was built by CSLR,
University of Colorado and was used to collect in-vehicle speech data (see
Chapter 2). This array was placed on the visor of the driver side. Another
linear microphone array built by Andrea Electronics was placed on the visor
of the front passenger.
A reference microphone was placed behind the driver seat facing the rear
passenger. Using an eight channel digital audio recording device, seven
channels that correspond to five microphones of the linear array, one output
of the Andrea microphone array (separate microphone outputs are not
available in the case of Andrea microphone array) and one reference
microphone data was recorded. A Sport Utility Vehicle (SUV) was employed
for this data collection task. Navigation related speech data including the
phrases “how far is the airport from Malibu?”, “Can you give me directions
to the airport?”, etc. was collected in side the vehicle for two speakers (one
male and one female) under three conditions - quiet (window was up, radio
and A/C were off and there was no presence of cross talk), radio on and cross
talk present.
