Citation Context

...ions in pitch which is otherwise assumed to be perfectly periodic. While this complicates analysis of a speech signal, it does allow for better synthesis as less information about the signal is lost. =-=[7]-=- STRAIGHT also makes better use of F0 information in analysis to create smooth spectral envelopes. [14] Very high quality and natural speech can therefore be produced using STRAIGHT. For intellectual ...

Citation Context

...ore rightly called HSMMs: semi-Markov, as transition probabilities are governed by a Gaussian trained on observed statistics rather than by the probability of a particular number of self-transitions. =-=[14]-=- For the dialectal modelling that is our goal, we expect spectral characteristics to be the most relevant as they correspond to pronunciation. Average overall pitch is fairly irrelevant, as it varies ...

Citation Context

...ne in the utterance to be synthesized, and average between the prediction from each model to produce the final set of statistics to be used. This minimizes the amount of time and processing required. =-=[10, 17]-=- An arbitrarily large number of voices can be used, as the final step would simply involve finding more statistics and averaging them appropriately, but we will use two here for theoretical simplicity...

Citation Context

...lly Gaelic-speaking area, unlike the Scots influence of the other two, and the Borders area where Jedburgh is located has historically had a very distinct accent to the Central Belt influence of Ayr. =-=[3, 6]-=- Second, as can be seen in figure 6.1, these three cities cover between them the largest area of Scotland of any three cities in the survey. This allows more space for mi26 Figure 6.1.: VOYS recording...

Citation Context

...ne in the utterance to be synthesized, and average between the prediction from each model to produce the final set of statistics to be used. This minimizes the amount of time and processing required. =-=[10, 17]-=- An arbitrarily large number of voices can be used, as the final step would simply involve finding more statistics and averaging them appropriately, but we will use two here for theoretical simplicity...

Citation Context

...quality. A model trained from data should learn the correct surface realization of a particular phone in a particular setting, regardless of what the symbol attached to that phone during training is. =-=[8]-=- Some Scottish models, for example, should learn to weaken or drop /l/at the end of a word in certain contexts and to realize it as dark otherwise. Considering the very 1Located at http://hts.sp.nitec...

Citation Context

... and what shifts should be made. Unfortunately there is relatively little research on regional variation in Scottish pronunciation, and not enough to implement such complex distinctions at this time. =-=[9]-=- Indeed, this is part of why dialectal voices must be trained from data. Under these circumstances, it is appropriate to fall back on this assumption to make it possible to model the dialect geography...

Citation Context

...the paucity of resources for Scottish accents. Recordings of a wide variety of adolescent speakers will allow evaluation of physical and sociolinguistic effects on their voices and manners of speech. =-=[2]-=- However, for the purposes of this project, the ages are primarily interesting as a note that voices may have different qualities to adult speech (for example, higher F0), as well as a possible explan...

Citation Context

...ue in two regards: it uses adolescents instead of trained adults, with the associated decrease in professionalism, and collection of recordings is decentralized, with some effects on quality control. =-=[3]-=- It is worthwhile to explore whether such corpora can still produce quality voices, as considering the relative ease of data collection they are more attractive to the prospective corpus builder, and ...

Citation Context

...istening to speech. The key point is that the MGCCs of a particular speech window are a series of numbers which indicate the vocal tract shape, essentially the phone being uttered aside from voicing. =-=[12]-=- • Duration, Traditionally when modelling with hidden Markov models, the duration spent in any particular state is modelled via the self-transition probability on that state. However, this is not appr...