Duration modelling

This disparity has been known for some time, and several attempts were made in ASR to correct this, the most notable proposal being the hidden semi-Markov model (HSMM) [282]. In this model, the transition probabihties are replaced by an explicit Gaussian dmation model. It is now known that this increase in durational accuracy does not in fact improve speech recognition to any significant degree, most probably because duration itself is not a significant factor in discrimination. In synthesis, however, modelling the duration accurately is known to be important and for this reason there has been renewed interest in hidden semi-Markov models [503], [513]. 

---

Finally, Graversen (2004) shows that the order in which the instruments of the activation are used is significant to the effect. Graversen makes use of a so-called duration model, where persons who have received an offer are compared with persons who have not received an offer yet. This analysis concludes that in cases where more than one offer has been given in an unemployment period, education must be offered before job training. 

We generally find statistically significant effects, of the expected sign, in the duration models. While we do not report the estimated shape parameters of the Weibull distribution in Tables 3.4 or 3.5, they indicate negative duration dependence as claim duration increases, the rate of exit from claimant status falls. Hence, the longer a claimant stays on a workers compensation claim, the less likely he is to leave it. 

Table 3.3 Duration Model Using Claimant Data Descriptive Statistics... 

HMM synthesis started with the now classic paper by Tokuda et al [453] which explained the basic principles of generating observations which obey the dynamic constraints. Papers explaining the basic principle include [452], [454] [305]. From this, a gradual set of improvements have been proposed, resulting in today s high quality synthesis systems. Enhancements include more powerful observation modelling [509] duration modelling in HMMs [504], [515] trended HMMs [134], [133], trajectory HMMs [515], HMM studies on emotion and voice transformation [232], [429], [404], [505]. 

A number of refinements to the basic HMM technique have been proposed, including more realistic duration modelling and accounting for global variance. 

Krishna, N. S., Talukdar, P. P., Bali, K., and Ramakrishnam, A. G. Duration modeling for Hindi text-to speech synthesis system. In Proceedings of the International Conference on Speech and Language Processing 2004 (2004). 

This result shows us how to generate from HMMs while also obeying the dynamic constraints of the delta coefficients. In further papers, Tokuda, his colleagues and others have extended this basic result to include any number of delta coefficients (e.g. acceleration coefficients, third-order deltas and so on) [453], exphcit duration modelling [514], use ofmore-sophisticated trajectory modelling [133], [134], [454], [513], use of multiple Gaussian mixtures and use of streams, which will be explained below. 

Davies MR, Mistry HB, Hussein L, PoUard CE, Valentin J-P, Swinton J, Abi-Gerges N (2012). An in siUco canine cardiac midmyocardial action potential duration model as a tool for early drug safety assessment. Am J Physiol Heart Circ Physiol 302 H1466-H1480. 

---