Ricardo Sousa

The harmonic and noise information of the glottal pulses in speech

Abstract This paper presents an algorithm, in the context of speech analysis and pathologic/dysphonic voices evaluation, which splits the signal of the glottal excitation into harmonic and noise components. The algorithm uses a harmonic and noise splitter and a glottal inverse filtering. The combination of these two functionalities leads to an improved estimation of the glottal excitation and its components. The results demonstrate this improvement of estimates of the glottal excitation in comparison to a known inverse filtering method (IAIF). These results comprise performance tests with synthetic voices and application to natural voices that show the waveforms of harmonic and noise components of the glottal excitation. This enhances the glottal information retrieval such as waveform patterns with physiological meaning.

Accurate analysis and visual feedback of vibrato in singing

Vibrato is a frequency modulation effect of the singing voice and is very relevant in musical terms. Its most important characteristics are the vibrato frequency (in Hertz) and the vibrato extension (in semitones). In singing teaching and learning, it is very convenient to provide a visual feedback of those two objective signal characteristics, in real-time. In this paper we describe an algorithm performing vibrato detection and analysis. Since this capability depends on fundamental frequency (F0) analysis of the singing voice, we first discuss F0 estimation and compare three algorithms that are used in voice and speech analysis. Then we describe the vibrato detection and analysis algorithm and assess its performance using both synthetic and natural singing signals. Overall, results indicate that the relative estimation errors in vibrato frequency and extension are lower than 0.1%.

Non-iterative frequency estimation in the DFT magnitude domain

The accurate estimation of the frequency of sinusoids is a frequent problem in many signal processing problems including the real-time analysis of the singing voice. In this paper we rely on a single DFT magnitude spectrum in order to perform frequency estimation in a non-iterative way. Two new frequency estimation methods are derived that are matched to the time analysis window and that reduce the maximum absolute estimation error to about 0.1% of the bin width of the DFT. The performance of these methods is evaluated including the parabolic method as a reference, and considering the influence of noise. A combined model is proposed that offers higher noise robustness than that of a single model.

DFT-based frequency estimation under harmonic interference

In this paper we address the accurate estimation of the frequency of sinusoids of natural signals such as singing, voice or music. These signals are intrinsicly harmonic and are normally contaminated by noise. Taking the Crame¿r-Rao Lower Bound for unbiased frequency estimators as a reference, we compare the performance of several DFT-based frequency estimators that are non-iterative and that use the rectangular window or the Hanning window. Tests conditions simulate harmonic interference and two new ArcTan-based frequency estimators are also included in the tests. Conclusions are presented on the relative performance of the different frequency estimators as a function of the SNR.

PLL based detector for coherent beacon receivers using DSP techniques

Amplitude and relative phase measurement of noisy CW signals require appropriate techniques, such as coherent detection, for maximum dynamic range and accuracy. We hereby describe a two channel system, developed for satellite microwave propagation beacon experiments, that comprises a user developed card, a DSP kit and dedicated software. The solution, that uses digital signal processing programmable chips and a DSP software managed PLL offers considerable advantages over a classical analogue system. Several prototype experimental results, tested with realistic signals, are presented and discussed.

Multi-label classification from high-speed data streams with adaptive model rules and random rules

Multi-label classification is a methodology that tries to solve classification problems where multiple classes are associated with each data example. Data streams pose new challenges to this methodology caused by the massive amounts of structured data production. In fact, most of the existent batch mode methods may not support this condition. Therefore, this paper proposes two multi-label classification methods based on rule and ensembles learning from continuous flow of data. These methods are derived from a multi-target regression algorithm. The main contribution of this work is the rule specialization for subsets of class labels, instead of the usual local (individual models for each output) or a global (one model for all outputs) methods. Prequential evaluation was conducted where global, local and subset operation modes were compared against other online classifiers found in the literature. Six real-world data sets were used. The evaluation demonstrated that the subset specialization presents competitive performance, when compared to local and global approaches and online classifiers found in the literature.

Online Semi-supervised Learning for Multi-target Regression in Data Streams Using AMRules

Most data streams systems that use online Multi-target regression yield vast amounts of data which is not targeted. Targeting this data is usually impossible, time consuming and expensive. Semi-supervised algorithms have been proposed to use this untargeted data (input information only) for model improvement. However, most algorithms are adapted to work on batch mode for classification and require huge computational and memory resources.

Co-training study for online regression

This paper describes the development of a Co-training (semi-supervised approach) method that uses multiple learners for single target regression on data streams. The experimental evaluation was focused on the comparison between a realistic supervised scenario (all unlabelled examples are discarded) and scenarios where unlabelled examples are used to improve the regression model. Results present fair evidences of error measure reduction by using the proposed Co-training method. However, the error reduction still is relatively small.

Co-training Semi-supervised Learning for Single-Target Regression in Data Streams Using AMRules

In a single-target regression context, some important systems based on data streaming produce huge quantities of unlabeled data (without output value), of which label assignment may be impossible, time consuming or expensive. Semi-supervised methods, that include the co-training approach, were proposed to use the input information of the unlabeled examples in the improvement of models and predictions. In the literature, the co-training methods are essentially applied to classification and operate in batch mode.

Online Multi-label Classification with Adaptive Model Rules

The interest on online classification has been increasing due to data streams systems growth and the need for Multi-label Classification applications have followed the same trend. However, most of classification methods are not performed on-line. Moreover, data streams produce huge amounts of data and the available processing resources may not be sufficient. This work-in-progress paper proposes an algorithm for Multi-label Classification applications in data streams scenarios. The proposed method is derived from multi-target structured regressor AMRules that produces models using subsets of output attributes (output specialization strategy). Performance tests were conducted where the operation modes global, local and subset approaches of the proposed method were compared to each other and to others online multi-label classifiers described in the literature. Three datasets of real scenarios were used for evaluation. The results indicate that the subset specialization mode is competitive in comparison to local and global approaches and to other online multi-label classifiers.