Temporal coherence is a feature-binding mechanism that ensures features that evolve together in time belong to the same object or event. Coherence has been extensively studied in biological systems, demonstrating how our brain leverages this mechanism to perform complex tasks in real environments and facilitate segregation of complex sensory signals (or wholes) into individual objects (or parts), following Gestalt principles. Although intuitive and computationally tractable, these concepts have rarely been leveraged in audio technologies. Audio event detection is an application that specifically deals with identifying sound events in an audio recording; hence is a natural avenue to explore principles of temporal coherence. In this study, we propose coherence-based learning, formulated as a contrastive loss, to train event detection models whereby embeddings driven by acoustic events are coherently constrained to maximize discriminability across events. This approach results in improved detection performance with no additional computational cost and a very small overhead during the training procedure.

Synthesis of music content generally leverages the underlying statistical structure of music to develop generative models, able to create new musical expressions within the same genre. In this work, we explore the statistical structure of a musical corpus and its effect on modulating the attention of listeners. The study specifically explores listeners' engagement to newly synthesized music and tests the hypothesis that maximizing statistical surprisal would result in increased auditory salience. The study employs a dynamical statistical model to estimate melodic line surprisal and develops an optimization procedure using parametrized codebooks to synthesize musical segments that maximize statistical surprisal. A behavioral experiment with a dichotic listening task is designed to probe salience of the synthesized melodies against original melodies by measuring listeners' engagement in a continuous-fashion. Results indicate that we can control the salience of sounds by manipulating the statistical surprisal, guided by the complexity of the temporal structure of the musical corpus. This work suggests that future work in automated music synthesis could leverage statistical models of music beyond musical aesthetics to also manipulate the degree of engagement.

Sound event detection is a challenging task, especially for scenes with multiple simultaneous events. While event classification methods tend to be fairly accurate, event localization presents additional challenges, especially when large amounts of labeled data are not available. Task4 of the 2018 DCASE challenge presents an event detection task that requires accuracy in both segmentation and recognition of events while providing only weakly labeled training data. Supervised methods can produce accurate event labels but are limited in event segmentation when training data lacks event timestamps. On the other hand, unsupervised methods that model the acoustic properties of the audio can produce accurate event boundaries but are not guided by the characteristics of event classes and sound categories. We present a hybrid approach that combines an acoustic-driven event boundary detection and a supervised label inference using a deep neural network. This framework leverages benefits of both unsupervised and supervised methodologies and takes advantage of large amounts of unlabeled data, making it ideal for large-scale weakly la-beled event detection. Compared to a baseline system, the proposed approach delivers a 15% absolute improvement in F-score, demonstrating the benefits of the hybrid bottom-up, top-down approach.

When an automatic speech recognition (ASR) system is deployed for real-world applications, it often receives only one utterance at a time for decoding. This single utterance could be of short duration depending on the ASR task. In these cases, robust estimation of speaker normalizing methods like feature-space maximum likelihood linear regression (FMLLR) and i-vectors may not be feasible. In this paper, we propose two unsupervised speaker normalization techniques-one at feature level and other at model level of acoustic modeling-to overcome the drawbacks of FMLLR and i-vectors in real-time scenarios. At feature level, we propose the use of deep neural networks (DNN) to generate pseudo-FMLLR features from time-synchronous pair of filterbank and FMLLR features. These pseudo-FMLLR features can then be used for DNN acoustic model training and decoding. At model level, we propose a generalized distillation framework, where a teacher DNN trained on FMLLR features guides the training and optimization of a student DNN trained on filterbank features. In both the proposed methods, the ambiguity in choosing the speaker-specific FMLLR transform can be reduced by augmenting i-vectors to the input filterbank features. Experiments conducted on 33-h and 110-h subsets of Switchboard corpus show that the proposed methods provide significant gains over DNNs trained on FMLLR, i-vector appended FMLLR, filterbank and i -vector appended filterbank features, in real-time scenario.