Time filter

Source Type

Anandha Mala G.S.,Eswari Engineering College Chennai
International Journal of Applied Engineering Research | Year: 2015

Man has the ability to hear multiple speech, recognize and respond to each speech separately. He/she can organize the time–frequency energy of the same speaker across time into a single stream where as it is not possible for a system to do. Hence a new system is be devised with superior methodology to do so. Many research works have been proposed earlier are falling under supervisory techniques. The real challenge lies in designing an algorithm for un-supervisory technique. Unsupervised Co-channel speech separation aims to separate two speech signals from a single mixture without using any pre-trained model. This paper presents a hybrid technique to separate the given speech mixture using Vector Quantization based Heuristic Clustering Algorithm (VQ-HCA) with high accuracy of speech separation. The algorithm first splits the given mixture as voiced and unvoiced speech segments using onset-offset algorithm. Then from the voiced speech segments, the pitch values are found out using tandem algorithm. These pitch values are grouped into two clusters using a Heuristic clustering algorithm. Finally the voiced segments of a person, unvoiced segments and complementary of voiced segment of second person (unvoiced segment) are grouped as a single stream to get the separated speech. Evaluations demonstrate that our approach outperforms even for two speakers with same gender combinations. Our algorithm produces significant accuracy level of the separated signals and the convergence time taken to separate the signals are less when compared to other supervised and unsupervised methods. © Research India Publications. Source

Discover hidden collaborations