Time-domain isolated phoneme classification using reconstructed phase spaces

Abstract

This paper introduces a novel time-domain approach to modeling and classifying speech phoneme waveforms. The approach is based on statistical models of reconstructed phase spaces, which offer significant theoretical benefits as representations that are known to be topologically equivalent to the state dynamics of the underlying production system. The lag and dimension parameters of the reconstruction process for speech are examined in detail, comparing common estimation heuristics for these parameters with corresponding maximum likelihood recognition accuracy over the TIMIT data set. Overall accuracies are compared with a Mel-frequency cepstral baseline system across five different phonetic classes within TIMIT, and a composite classifier using both cepstral and phase space features is developed. Results indicate that although the accuracy of the phase space approach by itself is still NOT THE PUBLISHED VERSION; this is the author's final, peer-reviewed manuscript. The published version may be accessed by following the link in the citation at the bottom of the page.

Sub-banded reconstructed phase spaces for speech recognition

“…This is because the time complexity of the Viterbi algorithm [1], which is used for the recognition of speech, is far greater for the RPS approach, due to the amount of data [30]. For RPS based methods to become useful, this issue must be solved.…”

“…Again, GMMs are used to model the RPS features, and are learned using binarysplit EM. The number of mixtures used, which was determined empirically in [30], is 128. The classification accuracy for an RPS of dimension 10, with delta dimensions is 38.81%.…”

“…Since each point in the RPS is treated as a frame, with likelihoods calculated for each model every 6.25 us for a sampling rate of 16,000 Hz, there are many more frames in this approach than in the standard spectral feature-based methodology. This leads to a recognition time complexity that is on the order of 100 times greater than automatic speech recognition systems based on spectral features [30].…”

Classificação de sinais de vozes saudáveis e patológicas por meio da combinação entre medidas da análise dinâmica não linear e codificação preditiva linear

“…Recentes pesquisas relacionadas às séries temporais, geradas a partir dos mecanismos de produção da voz humana, têm sido realizadas considerando-se as técnicas da dinâmica não linear e da teoria do caos com objetivos variados, dentre os quais podem ser destacados: classificação de fonemas (Johnson et al, 2005;Kokkinos e Maragos, 2005), reconhecimento automático de locutor (Petry, 2002), discriminação entre vozes saudáveis e patológicas, diagnóstico de patologias laríngeas e avaliação de efeitos de tratamentos clínicos (Dajer, 2006;Henríquez et al, 2009;Jiang et al, 2006;Scalassara et al, 2008;Torres et al, 2003;Zhang e Jiang, 2008).…”

Analysis and classification of speech signals by generalized fractal dimension features

“…The anal-4 ysis may be followed by measurement of invariant quantities on the reconstructed space. Early works in the field employing phase-space reconstruction include [21,22,26,27,17,18,28], whereas recently there has been increasing interest in the area [19,29,30]. These employ concepts on Lyapunov exponents [18,19,29], density models of the phase-space [30], correlation dimension measurements [18,28], especially for fricative consonants [17], or surrogate analysis on the nonlinear dynamics of vowels [31].…”