Speech probability distribution

Phonemic segmentation using the generalised Gamma distribution and small sample Bayesian information criterion

“…central limit theorem). Previous work in speech processing has demonstrated that Laplacian (LD) and Gamma (ΓD) distributions are more suitable than GD for approximating active voice segments for many frame sizes (Gazor and Zhang, 2003;Martin, 2005). More specifically, LD fits well the highly correlated univariate space of speech amplitudes as well as the uncorrelated multivariate space of feature values after a Karhunen-Loeve Transformation (KLT) or Discrete Cosine Transformation (DCT).…”

“…More specifically, LD fits well the highly correlated univariate space of speech amplitudes as well as the uncorrelated multivariate space of feature values after a Karhunen-Loeve Transformation (KLT) or Discrete Cosine Transformation (DCT). It is also asserted that DCT coefficients of clean speech samples are better modelled using the generalised Gaussian distribution (GGD) since they have a distribution which is more peaky than a LD (Gazor and Zhang, 2003). Nakamura (2000) proposes a generalised Laplace distribution (GLD) for speech recognition, demonstrating significant improvement in word accuracy when MFCC is used.…”

“…Once the parameters for each of the Q components have been estimated with the online gradient ascent algorithm, the likelihood ratio can be calculated. Here, as in (Gazor and Zhang, 2003), we are making the assumption that the noisy speech signal has uncorrelated components in the MFCC domain.…”

Voice Activity Detection with Generalized Gamma Distribution

“…In specific, LD fits well the highly correlated univariate space of the speech amplitudes as well as the uncorrelated multivariate space of the feature values after a Karhunen-Loeve Transformation (KLT) or Discrete Cosine Transformation (DCT) [9]. While some reports attest that LD offers only a marginally better fit than GD, this is not valid when silence segments are absent from the testing [8]. The reason is that while clean speech segments best exhibit LD or ΓD properties the silence segments are Gaussian random processes.…”

“…Using a transformed feature space, it is possible to assume that these two Gaussian random processes are independent of each other and maximum a posteriori estimators can be used to determine the signal parameters. Nevertheless, previous work in speech processing has demonstrated that Laplacian (LD) and Gamma (ΓD) distributions are more suitable than GD for approximating active voice segments for most frame sizes [8]. In specific, LD fits well the highly correlated univariate space of the speech amplitudes as well as the uncorrelated multivariate space of the feature values after a Karhunen-Loeve Transformation (KLT) or Discrete Cosine Transformation (DCT) [9].…”