Speech Enhancement Approach Based on Gravitational Search Algorithm (GSA)

Abstract

In this chapter, we propose a novel heuristic algorithm (Rashedi et al. in Inf Sci 2232–2248, 2009) called gravitational search algorithm (GSA) to speech enhancement. All the previously proposed stochastic approaches to dual-channel speech enhancement are based on the variants of particle swarm optimization (PSO). PSO has become popular optimization technique over the last decades for its simplicity of concept and implementation. However, the standard version of PSO tends to suffer from premature convergence by easily getting trapped into local minima. This is due to a decrease of diversity in the problem search space and it leads to the fitness stagnation of the swarm. In order to increase the diversity in search space and to improve the local searching capability, another heuristic algorithm GSA is proposed to speech enhancement in the present chapter. The search process in all the PSO and variants of PSO algorithms is based on bird flocking or fish schooling. GSA is another class of optimization technique with different scheme of search process and has an advantage of considering the distance between the neighbour agents to update the position of the agent. The inertia mass parameter considered for updating the agent movement in GSA algorithm is against the motion of the agent. A bigger inertia mass provides a slower motion of agents in the search space, and hence, a more precise local search is possible with increasing diversity in search space. On the other hand, a bigger gravitational mass causes a higher attraction of agents which permits a faster convergence. Hence, as an alternative approach to standard PSO-based enhancement technique, GSA is introduced to adaptive noise cancellation in speech enhancement systems with dual microphones. GSA is mainly constructed on the basis of law of gravity (Halliday et al. 2001; Serway and Jewett 2004) and the notion of mass interactions. The proposed algorithm is studied for real-world noise condition called babble noise, at three different input SNR levels. As per the present literature, there is no analysis about the intelligibility of enhanced speech using optimization techniques. In the present study, the proposed algorithm is compared with the standard PSO algorithm for dual-channel speech enhancement and the intelligibility analysis is also reported.