Editing Festival Speech Synthesis System (1997) (section)

=== New Method for Concatenate Synthesis ===
In terms of concatenate synthesis, except for the conventional approach of using a single-instance diphone-based method<ref>Beutnagel, M. C., & Conkie, A. (1999, September). Interaction of units in a unit selection database. In ''EUROSPEECH'' (Vol. 99, pp. 1063-1066).</ref> with an inventory comprising one recording of each diphone type, Festival's speech synthesis system introduced a novel "clunits" technique. This method involves an inventory of units recorded within natural sentences and employs a limited form of unit selection.<ref>Clark, R. A. J., Richmond, K., & King, S. (2007). Multisyn: Open-domain unit selection for the Festival speech synthesis system. ''Speech Communication'', ''49''(4), 317–330. <nowiki>https://doi.org/10.1016/j.specom.2007.01.014</nowiki></ref>

To be more specific, the fundamental strategy involves organizing units within a specific category, like a particular sound, based on considerations about the context of pronunciation and intonation. This categorization relies on questions related to linguistic elements, such as whether the unit appears at the end of a phrase or is stressed within a syllable. Then the Festival system constructs decision trees for each sound in the database, with each leaf representing a list of database units determined by the questions leading to that specific leaf. During the synthesis process, Festival utilizes the appropriate decision tree to identify the optimal cluster of potential units for each targeted specification. A search is then conducted to determine the most suitable path through the potential units, taking into account the distance of each unit from its cluster center and the cost associated with joining adjacent units.<ref>Black, A. W., & Taylor, P. A. (1997). ''Automatically clustering similar units for unit selection in speech synthesis.'' <nowiki>https://era.ed.ac.uk/handle/1842/1236</nowiki></ref>

This new technique can avoid the issue of estimating weights in a feature-based target distance measure while maintaining sensitivity to prosodic and phonetic distinctions. Additionally, it effectively manages variability in unit sparseness, with the tree-building algorithm initiating a cluster split only when significant variations are identifiable.