We describe the methodology for the collection and annotation of a large corpus of emotional speech data through crowdsourcing. The corpus offers 187 hours of data from 2,965 subjects. Data includes non-emotional recordings from each subject as well as recordings for five emotions: angry, happy-low-arousal, happy-high-arousal, neutral, and sad. The data consist of spontaneous speech elicited from subjects via a web-based tool. Subjects used their own personal recording equipment, resulting in a data set that contains variation in room acoustics, microphone, etc. This offers the advantage of matching the type of variation one would expect to see when exposing speech technology in the wild in a web-based environment. The annotation scheme covers the quality of emotion expressed through the tone of voice and what was said, along with common audio-quality issues. We discuss lessons learned in the process of the creation of this corpus.

Growth in voice-based applications and personalized systems has led to increasing demand for speech- analytics technologies that estimate the state of a speaker from speech. Such systems support a wide range of applications, from more traditional call-center monitoring, to health monitoring, to human-robot interactions, and more. To work seamlessly in real-world contexts, such systems must meet certain requirements, including for speed, customizability, ease of use, robustness, and live integration of both acoustic and lexical cues. This demo introduces SenSay AnalyticsTM, a platform that performs real-time speaker-state classification from spoken audio. SenSay is easily configured and is customizable to new domains, while its underlying architecture offers extensibility and scalability.

We are interested in Greek folk music genre classification by resorting to canonical correlation analysis (CCA). Here, the genre is related to the place of origin of the song. The CCA learns a linear transformation of the song lyrics descriptors that is highly correlated with their genre labels as well as another linear transformation of the audio features extracted from music recordings, which is maximally correlated with their genre labels. In the latter task, thanks to the deep CCA (DCCA), deep nonlinear transformations of the audio features are learnt, which are maximally correlated with the genre labels. Experimental findings are disclosed for a two-class genre recognition problem, employing folk songs originated from Pontus and Asia Minor. It is demonstrated that the CCA achieves an average accuracy of 97.02% across the 5 folds, when the term frequency-inverse document frequency features model the song lyrics. By modeling the music signal of each song with 28 mel-frequency cepstral coefficients (MFCCs) extracted from each frame and averaged over all frames, the average accuracy of the CCA drops to 72.9% across the 5 folds. The DCCA yields an accuracy of 69% for audio-based genre recognition.

The noise in musical audio recordings is assumed to obey an α-stable distribution. A sparse linear regression framework with structured priors is elaborated. Markov Chain Monte Carlo is used to infer the clean music signal model and the α-stable noise distribution parameters. The musical audio recordings are processed both as a whole and in segments by using a sine-bell window for analysis and overlap-and-add reconstruction. Experiments on noisy Greek folk music excerpts demonstrate better denoising under the α-stable noise assumption than the Gaussian white noise one, and when processing is performed in segments rather than in full recordings.

A novel method is proposed for updating an already trained asymmetric and symmetric probabilistic latent semantic analysis (PLSA) model within the context of a varying document stream. The proposed method is coined online PLSA (oPLSA). The oPLSA employs a fixed-size moving window over a document stream to incorporate new documents and at the same time to discard old ones (i.e., documents that fall outside the scope of the window). In addition, the oPLSA assimilates new words that had not been previously seen (out-of-vocabulary words), and discards the words that exclusively appear in the documents to be thrown away. To handle the new words, Good-Turing estimates for the probabilities of unseen words are exploited. The experimental results demonstrate the superiority in terms of accuracy of the oPLSA over well known PLSA updating methods, such as the PLSA folding-in (PLSA fold.), the PLSA rerun from the breakpoint, the quasi-Bayes PLSA, and the Incremental PLSA. A comparison with respect to the CPU run time reveals that the oPLSA is the second fastest method after the PLSA fold. However, the better accuracy of the oPLSA than that of the PLSA fold. pays off the longer computation time. The oPLSA and the other PLSA updating methods together with online LDA are tested for document clustering and F1 scores are also reported.

Probabilistic latent semantic analysis is enhanced with long distance bigram models in order to improve word clustering. The long distance bigram probabilities and the interpolated long distance bigram probabilities at varying distances within a context capture different aspects of contextual information. In addition, the baseline bigram, which incorporates trigger-pairs for various histories, is tested in the same framework. The experimental results collected on publicly available corpora (CISI, Cran field, Medline, and NPL) demonstrate the superiority of the long distance bigrams over the baseline bigrams as well as the superiority of the interpolated long distance bigrams against the long distance bigrams and the baseline bigram with trigger-pairs in yielding more compact clusters containing less outliers.

A novel system for speaker diarization is proposed that combines the eigengap criterion and cluster ensembles. No explicit assumptions on the number of speakers are made. Two variants of the system are developed. The first variant does not cluster the speech segments that are detected as outliers, while the second one does. The aforementioned system variants are assessed with respect to various metrics, such as the overall classification error, the average cluster purity, and the average speaker purity. Experiments are conducted on two-person dialogue scenes in movies as well as on news broadcasts from MDE RT-03 Training Data Speech Corpus released by the U.S. National Institute of Standards and Technology. In the latter case, the diarization error rate is also reported. It is demonstrated that the clustering performance does not degrade when outliers are present. Moreover, thanks to the eigengap criterion, the evaluation metrics are improved.

Face detection is a key problem in building systems that perform face recognition/verification and model-based image coding. Two algorithms for face detection that employ either support vector machines or backpropagation feedforward neural networks are described, and their performance is tested on the same frontal face database using the false acceptance and false rejection rates as quantitative figures of merit. The aforementioned algorithms can replace the explicitly-defined knowledge for facial regions and facial features in mosaic-based face detection algorithms.