Dionysis E. Tsoukalas

Speech enhancement based on audible noise suppression

A novel speech enhancement technique is presented based on the definition of the psychoacoustically derived quantity of audible noise spectrum and its subsequent suppression using optimal nonlinear filtering of the short-time spectral amplitude (STSA) envelope. The filter operates with sparse spectral estimates obtained from the STSA, and, when these parameters are accurately known, significant intelligibility gains, up to 40%, result in the processed speech signal. These parameters can be also estimated from noisy data, resulting into smaller but significant intelligibility gains.

Blind Estimation and Suppression of Late Reverberation Utilising Auditory Masking

A new method for blind estimation and suppression of late reverberation of speech signals is presented. The proposed algorithm consists of two steps. In a first step, the reverberation time is blindly determined from the reverberant signal. Then, an approximation of the power spectrum of late reverberation is subtracted from the power spectrum of the reverberant signal. Hence, a preliminary estimation of the anechoic speech spectrum is derived. In a second step, the auditory masking threshold of the clean spectrum estimation is calculated and used to define the coefficients for a nonlinear filter for the reverberant signal, which produces the final enhanced speech signal. The performance of the algorithm is demonstrated on artificially generated signals. Subjective tests are conducted and their results indicate that the quality of the speech signals obtained by the proposed method is superior when compared to previous methods.

Time-frequency mapping based on non-uniform smoothed spectral representations

For audio and acoustic system responses, the auditory system tends to smooth high frequency spectral regions and to register, mainly, low-frequency components of late echoes in the time domain. To model such processing, a theoretical model for non-uniform spectral smoothing is introduced allowing the choice of arbitrary frequency resolution profiles to match such auditory features. This generalized framework is extended to allow mapping of such smoothing spectra into the time domain.

Low bit-rate speech coding by perceptually optimized noise excitation modulation

Abstract A novel low bit-rate high-quality speech coding technique is presented based on a perceptually optimized signal reconstruction method. According to this parametric speech model, the signals spectral envelope is reconstructed from non-linear spectral filtering of an excitation signal, which is a combination of a random broadband noise signal with a number of discrete spectral pulses extracted from the original speech using a perceptual model. This general coding platform allows variable bit-rate implementations, starting from 1.9kbit/s, at which sufficient intelligibility (more than 92%) was measured, while at higher bit-rates (2.8 kbit/s) intelligibility scores were better than 94% with sufficient naturalness in the coded speech. In all cases, the complexity of the proposed system is very low.