Wavelet speech enhancement with reference signal based inference in application to personal active noise control systems

In many working industrial environments speech is significantly corrupted by acoustic noise. This makes communication between people difficult, limits working efficiency, enhances production costs, may be a source of misunderstanding, and finally lead to accidents. In other cases, people moving in industry halls or even on rush streets and making phone calls are barely understood by the far-end users. Therefore, the problem of enhancing intelligibility of transmitted or recorded speech has been gaining a significant interest for recent decades. The general difficulty is that the speech is nonstationary, and the noise is also nonstationary in many cases. Moreover, their frequency contents are overlapped, and thus employment of a simple ad-hoc filtering is unjustified. Therefore, more sophisticated algorithms are of interest. They take advantage of a rapid development of signal processing and microelectronics. There are, e.g. the family of spectral subtraction methods, Wiener or Kalman filtering, sub-space methods, neural networks, and wavelet decomposition. Most of them are based on a single measurement of the speech-plus-noise signal. However, in the applications mentioned above, a reference signal highly correlated with the noise disturbing the speech, and uncorrected with the speech itself can be acquired. Such structure is addressed in this paper. The wavelet decomposition method is accordingly adopted. The idea of wavelet coefficients thresholding is used. However, instead of calculating the thresholds, the voiced and unvoiced segments are classified with an inference mechanism based on the reference signal analysis. Efficiency of the proposed algorithm is proven on real-world data. The algorithm is found useful for improving speech intelligibility in active noise control and / or communication systems.