|
|
Esfandiar Zavarehei PhD
Student, |
 |
|
||||||||||||||
|
|
|
What We Do
It is well known that the more information about the signal we have, the better
we can estimate the signal in noisy conditions. Most speech enhancement
methods that are practically used these days, do not use speech models and
exclusive characteristics such as harmonic+noise
models, Formant trajectories, Pitch trajectories, etc. efficiently. With the
ever growing computation power, it is time we looked at more complex models
of speech for speech processing purposes. In Multimedia Signal Processing (MSP) group we are working on speech
enhancement in real-world noisy environments such as car and train. At this
time, we are mainly concerned with modelling the inter-frame correlation of
successive frames. Most speech enhancement methods proposed in the literature
are frame-based which means the estimation method processes each frame individually
and independent of its adjacent frames. Several methods are proposed which
smooth the estimate over several frames, but
most of these methods do not statistically optimize the inter-frame
correlation. A Kalman filter based method, which employs an autoregressive
(AR) filter to model the spectral fluctuations, is being developed in MSP
group. Also we are working on HNM model for reconstruction of speech after de-noising it.
|
|||||||||||||||
|
Contact me at esfandiar dot zeverehei at brunel dot ac dot uk |
|||||||||||||||||
