Nonlinear modeling for audio and acoustic systems

The market for technology incorporating voice-based interfaces is experiencing rapid expansion. The main reason for this growth is people's desire to use devices without their hands. A diverse range of applications such as voice-controlled gadgets, systems that can recognize voices or speakers, hands-free communication devices, systems for talking in noisy places without needing to speak directly into a microphone, and tools to help people with hearing problems. Nevertheless, in most scenarios, these systems suffer from acoustic feedback and ambient noise, which significantly reduce their effectiveness. Because of this, the people who design speech processing systems are very interested in finding ways to eliminate echoes and reduce background noise. In telecommunications remote conference is used to facilitate hands-free communication between users in multi-source environments without a body-attached mic or a wired microphone. As a result, this advancement brings challenges, as interfering sounds may disrupt both the quality and comprehensibility of your target speech. In hands-free communication, a conversation typically occurs between speakers at the near-end and far-end points. The near-end microphone picks up the intended speech signal and two interfering signals: the echo from a loudspeaker recreating the far-end signal and background noise. This acoustic coupling between the loudspeaker and microphone can reduce speech clarity at the far-end due to the echo. These issues arise from linear or nonlinear processes hence requiring similar adaptive solutions. Designing adaptive filters for this purpose presents various challenges related to modeling the acoustic pathway. This investigation deals with these modeling concerns by recommending new adaptive algorithms that enhance robustness against interferences on audio processing systems. With this novel technique, conventional design approaches can be bypassed leading to more efficient and effective adaptive filters. In this thesis, we illustrate our framework by focusing on audio applications where we developed a type of Linear-in-parameter nonlinear adaptive filter that is suitable for nonlinear acoustic echo cancellation applications. We set: Develop a new nonlinear model for efficient online processing in real-time applications, focusing on LIP filters implemented as Functional Link Adaptive Filters (FLAF). It will explore a novel class of FLAF using transformed domain adaptive filters, particularly Frequency Domain Adaptive Filters (FDAF) and their Partitioned Block variant (PBFDAF). The research will design a new algorithm for Nonlinear Acoustic Echo Cancellation (NAEC), investigate various expansion methods for the Functional Expansion Block (FEB), conduct experiments using simulated and real audio data, compare performance and computational complexity of different adaptation algorithms. Introduces a new class of FLAFs based on nearest Kronecker product (NKP) decomposition. These filters are designed for nonlinear system identification, capable of handling unknown degrees of nonlinearity by updating the most influential coefficients. Recognizing that not all weights significantly contribute to the model, the approach leverages the low-rank nature of weight vectors to improve convergence and tracking performance compared to traditional FLAFs. The proposed NKP-based FLAFs also aim to enhance noise mitigation, particularly in nonlinear acoustic echo cancellation scenarios, offering improved adaptability and performance for practical nonlinear system identification applications. Introduces robust filtering architectures based on a novel framework using an Adaptive Exponential Functional Link (AEFL) that is preceding spline nonlinearity. This approach aims to improve performance consistency for various nonlinear systems. Simulations show that this method outperforms traditional FLAF models in tackling nonlinear distortions and enhancing signal quality. These filters have been extensively tested under different situations of acoustic audio applications ranging from nonlinear acoustic echo cancellation to reverberation detection. The results obtained using our method surpass those based on other techniques. The evaluation will encompass both computational efficiency and performance metrics. The discourse will conclude by presenting potential avenues for future research and development in this field.