7th International Conference on Spoken Language Processing
September 16-20, 2002
An important topic in unit selection based speech synthesis is the scalability of such systems. Related to this problem is the question regarding the optimal size of a unit selection database. An ideal system should produce ever better synthesis results when more data is added to the system, but for a practical system this might not be the case. The unit selection criteria are generally not sufficiently developed to ensure that a system makes an optimal use of the data that it has available.
In this paper we propose a database reduction technique based on the statistical behaviour of unit selection. We investigate the effect of scaling down the database by objective and subjective criteria. We compare the proposed reduction technique with a technique that simply limits the size of unit lists to a fraction of their original size (random removal).
The results show that the proposed technique is far better than random removal, and that we can remove a significant portion of our database without causing any severe quality loss.
Bibliographic reference. Rutten, Peter / Aylett, Matthew P. / Fackrell, Justin / Taylor, Paul (2002): "A statistically motivated database pruning technique for unit selection synthesis", In ICSLP-2002, 125-128.