TY  - JOUR
T1  - Multichannel Speech Processing and Separation Using Hybridized
K-Superset Heuristic Algorithm
AU - , Logeshwari AU - Mala, Anandha 
JO  - Asian Journal of Information Technology
VL  - 15
IS  - 23
SP  - 4770
EP  - 4782
PY  - 2016
DA  - 2001/08/19
SN  - 1682-3915
DO  - ajit.2016.4770.4782
UR  - https://makhillpublications.co/view-article.php?doi=ajit.2016.4770.4782
KW  - Speech separation
KW  -multichannel
KW  -heuristic
KW  -K-superset
KW  -voiced
KW  -unvoiced
KW  -binary mask
AB  - The presence of competing speakers signal input mixture in a noisy and fluctuated environment
greatly degrades the performance. Noise and fluctuations hinders the process of speech separations. Very few
research works accounts for these factors in their supervisory and unsupervisory estimation methods. The
limitation of libraries or known speech samples and its recognition in a new environment makes speech signal
processing a cumbersome exercise. An intelligent unsupervisory method would be a better choice for such
requirements. Keeping view of this challenge, this research study proposes a novel method to separate
multichannel speech signal from a single mixture captured in both stationary and non-stationary noisy and
fluctuating environment using Multichannel B Hybridized K-superset Heuristic Speech separation Algorithm
(MC-HKHSA). MC-HKHSA estimates pitch values for voiced and fluctuated voiced segments and forms
supersets for multiple speakers. Noisy segments are filtered and fluctuated voice segments are grouped as a
unique stream. This approach involves coarse and fine level speech processing and segregation mechanism
making the overall process hybridized. Aim of MC-HKHSA is to segregate individual speech signals retaining
its intelligibility, quality and naturalness. It removes excess background residual noise while retaining the
positive features of enhanced speech. Simultaneous process management reduces the error rate due to interalgorithmic
value conversions. Simulation and experimental evaluations demonstrate that our approach
outperforms other existing schemes with various energy levels. The improvement was due to our fine
algorithmic analysis towards inter and intra superset evaluation and fine-tuning their overlapping coefficients
effectively through hybridized mechanism. The convergence time taken to separate the signals was
comparatively less when compared to other supervised and unsupervised methods. Results show the proposed
scheme consistently reduces background noise with no further apparent speech damage.
ER  -