Thomas Esch

Model-Based Dereverberation Preserving Binaural Cues

The ability of the human auditory system for sound localization mainly depends on the binaural cues, especially interaural time and level differences (ITD and ILD). In the context of digital hearing aids and binaural audio transmission systems, these cues can be severely degraded by independent bilateral signal processing such as dereverberation or noise reduction. This contribution presents a novel two-stage binaural dereverberation algorithm which explicitly preserves the binaural cues. The first stage is based on a statistical model of the room impulse responses (RIR) and comprises a spectral subtraction rule which reduces late reverberation only. It includes a smoothing process of the spectral gains to reduce musical tones. In a second stage, the residual reverberation is attenuated by a dual-channel Wiener filter. This is derived from a coherence model of the reverberant sound field taking into account shadowing effects of the head. The overall binaural-input binaural-output structure efficiently reduces both early and late reverberation. In experiments as well as informal listening tests using measured binaural room impulse responses, the proposed algorithm significantly improves speech quality according to objective and subjective measures.

Speech enhancement using a modified Kalman filter based on complex linear prediction and supergaussian priors

This paper presents a modified Kalman filter operating in the frequency domain for single channel speech enhancement. The proposed scheme uses a two step approach. In the first step, information from previous, enhanced speech DFT coefficients is exploited to perform an estimation of the current speech coefficients. Investigations show that the highest prediction gain is achieved by modeling the temporal trajectory of the speech DFT coefficients as a complex autoregressive (AR) process. In the second step, the first prediction is updated using three alternative spectral estimators, including the conventional Kalman filter gain. Instrumental measurements show the improvement of the proposed scheme compared to purely statistical weighting rules.

Model-based speech enhancement using SNR dependent MMSE estimation

This contribution presents a modified Kalman filter approach for single channel speech enhancement which is operating in the frequency domain. In the first step, temporal correlation of successive frames is exploited yielding estimates of the current speech and noise DFT coefficients. This first prediction is updated in the second step applying an SNR dependent MMSE estimator which is adapted to the (measured) statistics of the speech prediction error signal. Objective measurements show consistent improvements compared to estimators which do not take into account the temporal correlation or the influence of the input SNR on the statistics of the prediction error signal.

Wideband noise suppression supported by artificial bandwidth extension techniques

This contribution presents a wideband (50Hz – 7 kHz) speech enhancement system that is operating in the frequency domain. As a novel feature, techniques known from artificial bandwidth extension (BWE) are used to improve the spectral estimation process by exploiting the statistical dependencies between the low band (50Hz – 4kHz) and the high band (4–7kHz). Conventional noise suppression is used in the low band, while a novel approach is applied to the high band. Features from the processed (enhanced) low band signal are extracted and used to estimate subband energies of the high band. The weighting gains determined from these energy estimates are adaptively combined with conventional gains obtained in addition for the high band. The performance of the proposed method is shown to be consistently better than the conventional approach, especially at low input SNR values.

Combined reduction of time varying harmonic and stationary noise using frequency warping

Speech enhancement under non-stationary environments is still a challenging problem. This contribution presents a noise reduction system that is capable of tracking and suppressing both time varying harmonic noise and stationary noise. In a first stage, the harmonic noise power is estimated and attenuated using a modified Minimum Statistics approach that performs frequency warping according to the harmonics fundamental frequency. A conventional noise estimation technique is applied in a second stage in order to reduce the random components of the noise spectrum. The performance of the proposed noise suppression system is shown to be consistently better than conventional approaches.

An information theoretic view on Artificial Bandwidth Extension in noisy environments

Artificial Bandwidth Extension (ABWE) exploits spectral dependencies of speech signals and aims at recovering missing high frequency components if only the narrowband speech signal is available. This contribution provides an information theoretic view on ABWE when used in noisy conditions. Based on the results of [1], a performance bound of ABWE is formulated if the narrowband signal is disturbed by additive noise. The performance bound is evaluated using real entropy measurements and the influence of noise suppression prior to ABWE is investigated.