Discrete Fourier transform based noise reduction in audio signals

Audio signal quality degradation due to background noise remains a persistent challenge across communication, broadcasting, and recording industries. This research presents a spectral subtraction approach for noise reduction utilizing the Discrete Fourier Transform to separate noise components from desired audio content in the frequency domain [1]. The methodology transforms time-domain signals into frequency representations where noise characteristics can be estimated during speech pauses and subsequently subtracted from the composite spectrum. Implementation employed 2048-point DFT windows with 50% overlap and Hann windowing to minimize spectral leakage effects [2]. Testing across six distinct noise types revealed signal-to-noise ratio improvements ranging from 8.4 dB for impulsive noise to 29.4 dB for periodic hum contamination at 50/60 Hz power line frequencies. Processing efficiency achieved real-time capability on standard computing hardware, requiring only 12.4 ms to process one second of audio using optimized radix-2 FFT algorithms [3]. Perceptual quality evaluation using PESQ and STOI metrics confirmed that the noise reduction maintained speech intelligibility above 82% while achieving artifact suppression scores exceeding 71%. The spectral flooring technique with β = 0.02 effectively prevented musical noise artifacts that commonly plague spectral subtraction methods [4]. Comparative analysis against Wiener filtering demonstrated competitive performance with 23% lower computational requirements, making the DFT-based approach suitable for resource-constrained embedded applications. The research establishes practical guidelines for parameter selection based on noise characteristics and quality requirements [5].