Phase-Aware Signal Processing for Speech Transmission

Everyday life applications highly depend on successful speech transmission and speech communication, to name a few: smart homes with voice commands, hands-free mobile telephony, and speech recognition with machines. In all these applications it is quite important to guarantee a high performance robust to the background noise or reverberation in the room. A pre-processing stage in the form of signal enhancement is very important in order to remove the undesired background noise sources. While state-of-the-art technology for speech transmission mainly focuses on filtering the signal in the amplitude domain, we aim to push the limits of the achievable performance by extending the enhancement solution to both the amplitude and phase components leading to the new concept of phase-aware signal processing. The contributions in this proposal are divided in three lines: i) develop methods for estimating phase information about the desired source signal observed in noise, ii) develop novel speech enhancement algorithms in the complex domain to circumvent the limits of the conventional methods, iii) extension of single-channel source separation and artificial bandwidth extension using the concept of phase-aware signal processing, and iv) derive new quality estimators to reflect how humans perceive speech in noise and reduce or avoid tedious listening tests.

Everyday life applications heavily depend on successful speech transmission and speech communication, to name a few in voice command smart home or in handsfree mobile telephony and robust automatic speech recognition. All these applications are required to perform with high enough accuracy in presence of background noise or reverberation. A pre-processing stage in the form of speech enhancement or source separation is required to remove the undesired interfering component from the desired one before the target application. While state-of-the-art technology mainly focuses on filtering the signal in amplitude domain, we aim to push the limits of the achievable performance by extending the enhancement problem to both amplitude and phase parts leading to the new concept of phase-aware signal processing. The contributions in this proposal are in three folds: i) find solutions for estimating phase of desired signals from their mixed observation. The knowledge about phase is then further utilized in order to improve the amplitude estimation accuracy leading to a full complex signal enhancement framework, ii) The second goal of the proposal will be to extend the phase-aware signal processing tools to a speaker-dependent scenario. While previous works trained statistical models learned on spectral amplitude extracted from training sentences, we extend the dictionary learning idea to phase information as well. To this end, a new harmonic model plus phase distortion is employed to obtain a useful representation of phase, iii) Finally the last research question addressed in this proposal is to find reliable predictor of performance evaluation for the proposed phase-aware signal processing tools developed in previous steps. This is important to quantify how successful are the proposed methods in terms of achieving an accurate estimation of amplitude and phase parts when used for different applications: speech enhancement, source separation and robust automatic speech recognition.