A Single Channel Speech Enhancement Algorithm for Long Distance Scene

Research on speech enhancement has primarily focused on transmission distances of up to 8 meters. However, there is a growing demand for algorithms that can cater to greater distances, particularly in the context of mega-space transmission and security applications. This paper aims to address this gap by creating a long-distance Chinese speech dataset and proposing a single-channel speech enhancement algorithm. The algorithm incorporates log-MMSE (Minimum Mean Square Error) noise reduction, Mel acoustic parameters for dereverberation, pre-emphasis high-frequency enhancement, and parametric equalizer tone optimization to mitigate noise interference, high-frequency attenuation, and sound aliasing in large-space sound transmission. The proposed algorithms have been shown to significantly improve both subjective and objective measures of speech quality, as evidenced by improvements in speech spectrogram, PESQ (Perceptual Evaluation of Speech Quality), STOI (Short-Time Objective Intelligibility), and MUSHRA (Multiple Stimuli with Hidden Reference and Anchor) scores. Additionally, the equalizer enhances intelligibility and naturalness while compensating for timbral stiffness in intermediate processing steps.