Underdetermined Speech Blind Identification Based on Spectrum Correction and Phase Coherence Criterion

The existing underdetermined speech blind identification (BI) algorithms can hardly possess both high recovery quality and high efficiency. This limitation may lie in the neglection of the phase extraction of speeches, which requires technical innovation of the phase-coherence identification. This paper proposes a BI scheme characterized by a combination of the ratio-interpolation-based spectrum corrector and a phase-coherence criterion (involving the operations of frequency merging, effective candidate pattern screening, and single-active-source (SAS) pattern recognition). Its high recovery quality is due to the combination that yields a set of SAS patterns with accurate harmonic parameters. Its high efficiency arises from two aspects: first, the phase-coherence criterion condenses the original patterns into a small quantity of SAS patterns; and second, an efficient density-based clustering algorithm is adopted to classify these SAS patterns. Essentially, the performance enhancement owns to the fact that the sources' phase information can be effectively extracted from the observations by means of the above technique combinations. Both the theoretical analysis and simulation verified that the proposed BI method outperforms the existing BI algorithms in recovery quality, efficiency, and anti-noise performance, which presents a vast potential in other harmonics-related BSS fields, such as mechanical vibration analysis, and channel estimation in communication.