Harmonics extraction based speech recovery for underdetermined mixing systems

摘要

It is an intractable task to achieve high-efficiency and high-quality speech recovery for the existing underdetermined systems. To solve this problem, this paper proposes a harmonics extraction based underdetermined speech recovery algorithm, which consists of 4 stages. In the 1st stage, spectrum correction technique is adopted to extract the harmonic components from the mixtures’ short time Fourier transform (STFT); In the 2nd stage, a phase-coherence criterion is applied on these harmonic components to identify the single source components; In the 3rd stage, these single source patterns are further categorized into multiple groups by means of the adaptive k -means clustering, from which the mixing matrix is estimated; In the last stage, this estimated matrix is further combined with the subspace projection algorithm, which resultantly yields the final source recovery. The high efficiency lies in that the harmonics extraction is properly combined with single source component identification. The speech recovery experiment demonstrated that, compared to the original subspace projection algorithm, the proposed method can acquire a higher recovery quality, which presents a potential application in other harmonic related fields.