Sound source separation has become a topic of intensive research in the last years. The research effort has been specially relevant for the underdetermined case, where a considerable number of sparse methods working in the time-frequency (T-F) domain have appeared. In this context, although binary masking seems to be a preferred choice for source demixing, the estimated masks differ substantially from the ideal ones. This paper proposes a maximum a posteriori (MAP) framework for binary mask estimation. To this end, class-conditional source probabilities according to the observed mixing parameters are modeled via ratios of dependent Cauchy distributions while source priors are iteratively calculated from the observed histograms. Moreover, spatially smoothed posteriors in the T-F domain are proposed to avoid noisy estimates, showing that the estimated masks are closer to the ideal ones in terms of objective performance measures.

This paper considers 2-D Fourier analysis of local time-frequency regions of wideband spectrograms, a representation we refer to as the wideband Grating Compression Transform (WGCT). We develop frequency-dependent, speech-production-based models of speech signals for the WGCT, building on previous work in modeling narrowband-based GCT representations (NGCT). Comparisons show important distinctions, including dual behavior, between the wideband and narrowband models, and distinct ways in which vocal tract/formant content is distributed redundantly throughout the NGCT and WGCT spaces. Our results motivate a novel taxonomy of speech-signal behavior as an interpretative framework (i.e., in relation to speech-production characteristics) for 2-D processing of speech using the GCT, as well as for other 2-D approaches and time-frequency distributions such as the auditory spectrogram. We demonstrate and evaluate the ability of the model to represent real speech content through demodulation techniques for analysis/synthesis of wideband spectrograms. Finally, we develop a co-channel speaker separation method, using prior and estimated pitch information, based on the WGCT, as well as through fusion with the NGCT. These GCT-based separation systems are compared against and further fused with a reference sinusoidal separation system.

We present a new system for the harmonic analysis of popular musical audio. It is focused on chord estimation, although the proposed system additionally estimates the key sequence and bass notes. It is distinct from competing approaches in two main ways. First, it makes use of a new improved chromagram representation of audio that takes the human perception of loudness into account. Furthermore, it is the first system for joint estimation of chords, keys, and bass notes that is fully based on machine learning, requiring no expert knowledge to tune the parameters. This means that it will benefit from future increases in available annotated audio files, broadening its applicability to a wider range of genres. In all of three evaluation scenarios, including a new one that allows evaluation on audio for which no complete ground truth annotation is available, the proposed system is shown to be faster, more memory efficient, and more accurate than the state-of-the-art.

Human communication is more than just the transmission of information. It also involves complex interaction dynamics that reflect the roles and communication styles of the participants. A novel approach to studying human communication is to view conversation as a coupled time series and apply analysis techniques from dynamical systems to the recurring topics or concepts. In this paper, we define a set of metrics that enable quantification of the complex interaction dynamics visible in conceptual recurrence. These multi-participant recurrence (MPR) metrics can be seen as an extension of recurrence quantification analysis (RQA) into the symbolic domain. This technique can be used to monitor the state of a communication system and inform about interaction dynamics, including the level of topic consistency between participants; the timing of state changes for the participants as a result of changes in topic focus; and, patterns of topic proposal, reflection, and repetition. We demonstrate three use studies applying the new metrics to conversation transcripts from different genres to demonstrate their ability to characterize individual communication participants and intergroup communication patterns.

It has been possible to reproduce point sound sources between listeners and a loudspeaker array by using the focused-source method. However, this method requires physical parameters of the sound sources to be reproduced, such as source positions, directions, and original signals. This fact makes it difficult to apply the method to real-time reproduction systems because decomposing received signals into such parameters is not a trivial task. This paper proposes a method for recreating virtual sound sources in front of a planar or linear loudspeaker array. The method is based on wave field synthesis but extended to include the inverse wave propagator often used in acoustical holography. Virtual sound sources can be placed between listeners and a loudspeaker array even when the received signals of a microphone array equally aligned with the loudspeaker array are only known. Numerical simulation results are presented to compare the proposed and focused-source methods. A system was constructed consisting of linear microphone and loudspeaker arrays and measurement experiments were conducted in an anechoic room. When comparing the sound field reproduced using the proposed method with that using the focused-source method, it was found that the proposed method could reproduce the sound field at almost the same accuracy.

This paper compares a set of large scale support vector machine (SVM) training algorithms for language and speaker recognition tasks. We analyze five approaches for training phonetic and acoustic SVM models for language recognition. We compare the performance of these approaches as a function of the training time required by each of them to reach convergence, and we discuss their scalability towards large corpora. Two of these algorithms can be used in speaker recognition to train a SVM that classifies pairs of utterances as either belonging to the same speaker or to two different speakers. Our results show that the accuracy of these algorithms is asymptotically equivalent, but they have different behavior with respect to the time required to converge. Some of these algorithms not only scale linearly with the training set size, but are also able to give their best results after just a few iterations. State-of-the-art performance has been obtained in the female subset of the NIST 2010 Speaker Recognition Evaluation extended core test using a single SVM system.