Theoretical and experimental bases of a new method for accurate separation of harmonic and noise components of speech signals

In this paper, the problem of separating the harmonic and aperiodic (noise) components of speech signals is addressed. A new method is proposed, based on two specific processes dedicated to better take into account the non-stationarity of speech signals: first, a period-scaled synchronous analysis of spectral parameters (amplitudes and phases) is done, referring to the Fourier series expansion of the signal, as opposed here to the typically used Short-Term Fourier Transform (STFT). Second, the separation itself is based on a low-pass time-filtering of the parameters trajectory. Additionally to presenting the theoretical basis of the method, preliminary experiments on synthetic speech are provided. These experiments show that the proposed method has the potential to significantly outperform a reference method based on STFT: Signal-to-error ratio gains of 5 dB are typically obtained in the presented experiments. Conditions to go beyond the theoretical framework towards more practical applications on real speech signals are discussed.