A scalable frequency domain-based linear convolution architecture for speech enhancement

Most speech enhancement algorithms consist of a time-varying filter which is applied to the signal in the frequency domain. One of the motivations for filtering in the frequency domain compared to convolution in the time domain is to reduce the computational complexity. Filtering in the frequency domain, however, can introduce distortion if the linear convolution condition is not fulfilled. Although there are standard approaches to linear convolution in the frequency domain, they tend to be computationally prohibitive for small terminals. In this paper, we present a new and more efficient approach. The proposed approach is derived from the equivalence of zero-padding and interpolation in time and frequency domains. A distinct advantage of the approach proposed in this paper relates to its scalability which is exploited to manage computational complexity with only moderate degradation in speech quality.