An Analysis of Sound Event Detection under Acoustic Degradation Using Multi-Resolution Systems

The Sound Event Detection task aims to determine the temporal locations of acoustic events in audio clips. In recent years, the relevance of this field is rising due to the introduction of datasets such as Google AudioSet or DESED (Domestic Environment Sound Event Detection) and competitive evaluations like the DCASE Challenge (Detection and Classification of Acoustic Scenes and Events). In this paper, we analyze the performance of Sound Event Detection systems under diverse artificial acoustic conditions such as high- or low-pass filtering and clipping or dynamic range compression, as well as under an scenario of high overlap between events. For this purpose, the audio was obtained from the Evaluation subset of the DESED dataset, whereas the systems were trained in the context of the DCASE Challenge 2020 Task 4. Our systems are based upon the challenge baseline, which consists of a Convolutional-Recurrent Neural Network trained using the Mean Teacher method, and they employ a multiresolution approach which is able to improve the Sound Event Detection performance through the use of several resolutions during the extraction of Mel-spectrogram features. We provide insights on the benefits of this multiresolution approach in different acoustic settings, and compare the performance of the single-resolution systems in the aforementioned scenarios when using different resolutions. Furthermore, we complement the analysis of the performance in the high-overlap scenario by assessing the degree of overlap of each event category in sound event detection datasets.

Real-time Noise-suppressed Wide-Dynamic-Range Compression in Ultrahigh-Resolution Neuronal Imaging

SummaryWith a limited dynamic range of an imaging system, there are always regions with signal intensities comparable to the noise level, if the signal intensity distribution is close to or even wider than the available dynamic range. Optical brain/neuronal imaging is such a case where weak-intensity ultrafine structures, such as, nerve fibers, dendrites and dendritic spines, often coexist with ultrabright structures, such as, somas. A high fluorescence-protein concentration makes the soma order-of-magnitude brighter than the adjacent ultrafine structures resulting in an ultra-wide dynamic range. A straightforward enhancement of the weak-intensity structures often leads to saturation of the brighter ones, and might further result in amplification of high-frequency background noises. An adaptive illumination strategy to real-time-compress the dynamic range demands a dedicated hardware to operate and owing to electronic limitations, might encounter a poor effective bandwidth especially when each digitized pixel is required to be illumination optimized. Furthermore, such a method is often not immune to noise-amplification while locally enhancing a weak-intensity structure. We report a dedicated-hardware-free method for rapid noise-suppressed wide-dynamic-range compression so as to enhance visibility of such weak-intensity structures in terms of both contrast-ratio and signal-to-noise ratio while minimizing saturation of the brightest ones. With large-FOV aliasing-free two-photon fluorescence neuronal imaging, we validate its effectiveness by retrieving weak-intensity ultrafine structures amidst a strong noisy background. With compute-unified-device-architecture (CUDA)-acceleration, a time-complexity of <3 ms for a 1000×1000-sized 16-bit data-set is secured, enabling a real-time applicability of the same.

Does the Speech Cue Profile Affect Response to Amplitude Envelope Distortion?

Purpose A broad area of interest to our group is to understand the consequences of the “cue profile” (a measure of how well a listener can utilize audible temporal and/or spectral cues for listening scenarios in which a subset of cues is distorted. The study goal was to determine if listeners whose cue profile indicated that they primarily used temporal cues for recognition would respond differently to speech-envelope distortion than listeners who utilized both spectral and temporal cues. Method Twenty-five adults with sensorineural hearing loss participated in the study. The listener's cue profile was measured by analyzing identification patterns for a set of synthetic syllables in which envelope rise time and formant transitions were varied. A linear discriminant analysis quantified the relative contributions of spectral and temporal cues to identification patterns. Low-context sentences in noise were processed with time compression, wide-dynamic range compression, or a combination of time compression and wide-dynamic range compression to create a range of speech-envelope distortions. An acoustic metric, a modified version of the Spectral Correlation Index, was calculated to quantify envelope distortion. Results A binomial generalized linear mixed-effects model indicated that envelope distortion, the cue profile, the interaction between envelope distortion and the cue profile, and the pure-tone average were significant predictors of sentence recognition. Conclusions The listeners with good perception of spectro-temporal contrasts were more resilient to the detrimental effects of envelope compression than listeners who used temporal cues to a greater extent. The cue profile may provide information about individual listening that can direct choice of hearing aid parameters, especially those parameters that affect the speech envelope.