A model based approach to tonal interference suppression; an LPC and cepstral filtering framework for alarm noise reduction in single channel speech signals in the context of emergency calls

Abstract

When alarm call center operators receive emergency calls, the caller is often located close to a nearby activated fire alarm or smoke detector. In many cases, this means that alarm noise has a strong presence in the call received by the operator. This is an issue for several reasons. First, it exposes the operator to the risk of hearing damage, second, it degrades the quality of the speech, thereby increasing the listening effort and the cognitive demand. This thesis acknowledges this problem and applies it to the context of a real emergency call center. Therefore, alarm noise levels in the headsets were measured at the Oslo Fire and Rescue Service in Oslo, Norway. The noise levels in the headsets were found to be above the Norwegian action limit values. In an attempt to propose a solution to this problem, two model-based methods for suppressing tonal noise in emergency calls were proposed. The techniques used in this work are grounded in the source-filter separation approach: Cepstral liftering, and linear predictive coding (LPC). Based on these techniques, two filter frameworks were programmed in MATLAB. In the first proposed method, Proposal A, tonal interference suppression was applied solely by cepstral liftering. In Proposal B, this technique was iterated and combined with LPC-based analysis. The performance of the proposed filters was discussed, based on analytical results computed in MATLAB, including waveforms, pitch detection, and spectograms. The filters were further evaluated by objective quality measures (WSS, LLAR and fwSNRseg. To extend the analysis, a subjective comparative listening test was conducted. The results of the listening test indicated that even though the proposed models were partially successful in suppressing tonal noise components, this was accomplished at the cost of a decreased speech quality. This proved to have a negative impact on the overall perception and the original degraded audio signal was preferred in most cases over the filtered one. The filters were also compared to the built-in filter in the videoconferencing software Webex, which uses AI technology, based on data. Compared to the proposed filters, Webex was found to be superior in suppressing the tonal noise and, at the same time, it managed to preserve the speech quality to a great extent. More work is needed, to achieve effective suppression of tonal interference while preserving speech quality through a model-based approach. Large data sets are required for the fine-tuning of parameters to ensure efficiency and adaptability to different scenarios. It is suggested that further work investigates the implementation of data-driven technologies and the use of trained neural networks.