Speech Signal Enhancement by Information Combining (Aachener Beiträge zu digitalen Nachrichtensystemen)

Mobile phones as well as tablets are omnipresent and belong to everyday life. Today audiovisual communication takes place at different locations and in a large variety of acoustic environments. In consequence, the intelligibility as well as the quality of speech may significantly be degraded by ambient background noise. In order to improve speech intelligibility and to ensure a convenient communication with high audio quality, speech enhancement techniques are required. In this thesis all critical components contributing to the enhancement of the up-link signal are addressed:
• signal capturing at the acoustic front-end with a new near field beamformer
• new codebook based speech and noise estimation procedure generating and exploiting reliability information, and
• actual noise reduction exploiting spectral dependencies of human speech.
For the acoustic front-end of the digital processing chain a novel concept for the filter optimization of a near field beamformer is introduced. The optimization scheme allows to closely approximate a predefined reception characteristic which can be freely chosen according to the application. The output of the beamformer provides a pre-enhanced signal with improved SNR for subsequent single-microphone based speech enhancement.
Single-microphone noise reduction usually relies on statistical properties of speech and noise. In general, the noise is assumed to be stationary or only slightly time-varying, which is in practice often not fulfilled. Due to imprecise noise estimation, single-microphone systems are prone to unpleasant artifacts that are called musical tones. In this context different Information Combining methods, merging various estimates, are presented which address specifically the problem of non-stationary noise signals, leading to a significant improved estimation accuracy. On the one hand, the proposed Information Combining is used with respect to spectral dependencies of human speech. On the other hand, it merges the best of several speech and noise estimates depending on their reliability. The necessary estimates are provided by a new statistical noise estimator as well as a codebook driven speech and noise estimation algorithm. The achieved estimation quality opens up the possibility to close the gap between the conflicting goals of high noise attenuation, low speech distortion, and the prevention of undesired musical tone artifacts. Finally, the practical aspects of the proposed enhancement systems are considered and discussed with two implemented real-time demonstrators.