Speech compensation using Stereo Based Stochastic Vector Mapping based on Full Covariance Models
|Author(s)||Randa Al-Wakeel, Mahmoud Shoman, Magdy Aboul-Ela, Sherif Abdo|
|Issue Date||August, 2014|
|Publishing Date||August, 2014|
|Keywords||speech compensation, noisy speech recognition, Stochastic vector Mapping.|
Speech compensation techniques aim to provide speech recognition systems with the robustness against sources of noise existing in the real environments. These sources of noise cause the recognition performance to deteriorate dramatically.
In this paper, we are interested in Stereo based Stochastic Vector Mapping (SSM) speech compensation technique introduced in . In , in the experimental work, it was assumed that the speech features are uncorrelated (independent). This assumption simplifies the estimation of the enhancement parameters and it reduces the needed implementation time. In this paper, we will extend the experimental work to the case when the speech features are correlated (dependent). We aim to clarify the effect of considering the correlation on the efficiency of SSM. We considered the two estimators Maximum A Posteriori (MAP) estimation and Minimum Mean Square Error (MMSE) estimation used in . A part of the experimental work was dedicated to test the SSM with Multi Style Trained (MST) recognition models and also with recognition models trained using SSM compensated speech.
However, considering the correlation between features introduces better performance, it cannot be applied in real time applications without a way to reduce the complexity of the implementation and the time needed.