A New Method for Speech Enhancement Based on Harmonic Distortion

A near-end listening enhancement (NELE) method that combines harmonic (non-linear) distortion with linear distortion for improving speech intelligibility in noisy conditions without altering the signal-to-noise ratio (SNR) is proposed. Parameters were optimized across noise types, SNRs, and reverberation using a Genetic Algorithm with Speech Intelligibility in Bits (SIIB) as the loss function. Subjective experiments confirmed that, on average, the odds of correctly recognizing a word treated with the proposed method were 18.8 higher relative to plain speech when presented at $-9$ dB SNR, averaged over the results of Cafeteria and Speech-Shaped Noise (SSN). In Cafeteria noise, the odds of correct recognition were 16.4 times higher compared to plain speech, while in SSN they were about 8.4 times higher, averaged across SNRs ranging from $-9$ to $-3$ dB. Objective evaluation further confirmed its superiority to alternative approaches in reverberant maskers. These results demonstrate that combining harmonic and linear distortions can substantially improve intelligibility in adverse listening conditions.