This paper proposes a hybrid speech enhancement estimator that integrates the Perceptually-motivated Karhunen–Loève Transform (PKLT) with the Dual-Masking Harmonic-based (DMH) algorithm in a unified framework termed PKDMH. The main novelty lies in combining perceptual subspace projection with harmonic-residual suppression, enabling the system to jointly remove noise while preserving speech-relevant spectral cues. PKLT first performs perceptual subspace projection and suppresses inaudible components, after which DMH eliminates remaining broadband and harmonic residuals. The proposed PKDMH system was evaluated using the TIMIT dataset contaminated with five noise types: White, Pink, F16, Airport, and Car noise—across five SNR levels (−10 dB, −5 dB, 0 dB, +5 dB, +10 dB). Objective evaluation used the standard perceptual and signal-level measures of PESQ, STOI, SNRseg, Csig, Cbak and Covl. Results show that the enhanced quality of separation and speech signal ratio between enhanced signals and original target binary mask cause obvious improvements in quantity, with average PESQ gains of 1.099, 0.888 and 0.824 for White, Pink and F16 noise, respectively. These results bring out the subjective benefit of the PKDMH cascade, in terms of being a more robust enhancement approach under low SNR and acoustically varying cases.
