Signal processing and methods for advanced acoustic applications

From a loudspeaker manufacturer perspective, the intelligibility of sound can be significantly affected on both downstream and upstream side of a loudspeaker production chain, differently. On the downstream side, the sound intelligibility can be affected by characteristics of typical acoustic environments: sound waves are partially reflected by the physical boundaries of the environment leading to reverberation, echo and feedback problems. On the upstream side, the quality of the sound is directly correlated to the quality of the speaker: inspections and quality control protocols are conducted during pre-production, production, and pre-shipment stage to reduce the amount of damaged drivers.;The research presented in this thesis deals with signal processing algorithms in order to develop both robust downstream and upstream solutions of a loudspeaker production chain, providing increased performance and sound quality for advanced acoustic systems in realistic conditions.;On the downstream side of a loudspeaker production chain, the acoustic feedback problem is considered and a novel algorithm for the adaptive feedback cancellation in a single acoustic MIMO array is proposed. When a microphone is too close to the loudspeaker or the amplification is too large, acoustic feedback can occur where acoustic effects are perceived as howling and ringing, degrading sound intelligibility and sound quality.;While the canonical methods (automatic gain control, notch filtering, phase modulation) provide a reactive solution with limited performance, gain and high computational complexity, the new framework namely, Partitioned Block Frequency Domain (PBFD) based Adaptive Feedback Cancellation (AFC) method, is able to tackle the acoustic feedback problem in large acoustic spaces. The results of the proposed framework is compared with the state of the art using real acoustic data showing superior performance with up to 18dB of Maximum Stable Gain (MSG) and 30 seconds less convergence time.;On the upstream side of a loudspeaker production chain, the use of radar micro-Doppler for loudspeaker analysis is introduced for the first time. This approach offers the potential benefits to characterize the mechanical motion of a loudspeaker and identify defects. Increasing quality checks at various stages of production (with limited costs) can provide substantial benefits to loudspeaker manufacturers. Compared with acoustic based approaches, the use of a radar allows reliable measurements in an acoustically noisy end of production line. In addition, when compared to a laser vibrometric approach the use of radar micro-Doppler reduces the number of measurements required and provides direct access to the information of the metallic components of the loudspeaker.;Following the modelling of the radar return from the loudspeaker, a procedure to extract mechanically impaired features of the loudspeaker motion is introduced. Results show the ability to detect the linear and harmonic frequency responses of both good and defected speakers. These are used as features of a Bidirectional Long Short-Term Memory (BiLSTM) Recurrent Neural Network (RNN) classifier, leading to classification accuracy above the 98% on real data.From a loudspeaker manufacturer perspective, the intelligibility of sound can be significantly affected on both downstream and upstream side of a loudspeaker production chain, differently. On the downstream side, the sound intelligibility can be affected by characteristics of typical acoustic environments: sound waves are partially reflected by the physical boundaries of the environment leading to reverberation, echo and feedback problems. On the upstream side, the quality of the sound is directly correlated to the quality of the speaker: inspections and quality control protocols are conducted during pre-production, production, and pre-shipment stage to reduce the amount of damaged drivers.;The research presented in this thesis deals with signal processing algorithms in order to develop both robust downstream and upstream solutions of a loudspeaker production chain, providing increased performance and sound quality for advanced acoustic systems in realistic conditions.;On the downstream side of a loudspeaker production chain, the acoustic feedback problem is considered and a novel algorithm for the adaptive feedback cancellation in a single acoustic MIMO array is proposed. When a microphone is too close to the loudspeaker or the amplification is too large, acoustic feedback can occur where acoustic effects are perceived as howling and ringing, degrading sound intelligibility and sound quality.;While the canonical methods (automatic gain control, notch filtering, phase modulation) provide a reactive solution with limited performance, gain and high computational complexity, the new framework namely, Partitioned Block Frequency Domain (PBFD) based Adaptive Feedback Cancellation (AFC) method, is able to tackle the acoustic feedback problem in large acoustic spaces. The results of the proposed framework is compared with the state of the art using real acoustic data showing superior performance with up to 18dB of Maximum Stable Gain (MSG) and 30 seconds less convergence time.;On the upstream side of a loudspeaker production chain, the use of radar micro-Doppler for loudspeaker analysis is introduced for the first time. This approach offers the potential benefits to characterize the mechanical motion of a loudspeaker and identify defects. Increasing quality checks at various stages of production (with limited costs) can provide substantial benefits to loudspeaker manufacturers. Compared with acoustic based approaches, the use of a radar allows reliable measurements in an acoustically noisy end of production line. In addition, when compared to a laser vibrometric approach the use of radar micro-Doppler reduces the number of measurements required and provides direct access to the information of the metallic components of the loudspeaker.;Following the modelling of the radar return from the loudspeaker, a procedure to extract mechanically impaired features of the loudspeaker motion is introduced. Results show the ability to detect the linear and harmonic frequency responses of both good and defected speakers. These are used as features of a Bidirectional Long Short-Term Memory (BiLSTM) Recurrent Neural Network (RNN) classifier, leading to classification accuracy above the 98% on real data