![]() |
EUROSPEECH 2003 - INTERSPEECH 2003
|
![]() |
This work comes within the scope of concatenative speech synthesis. We propose a method to evaluate the criteria used in units selection methods. Usually criteria are evaluated in a comparative black-box way : the performance of a criterion are measured relatively to other criteria performances, thus evaluation is not always discriminant or formative. We present a glas-box method to measure the performances of a criterion in an absolute way. The principle is to explore the possible sequences of units able to synthesize a given target utterance, to assign to each sequence a value X of objective quality and a value Y related to the tested criterion ; then mutual information I(X;Y) is calculated to measure the explicative power of the criterion Y in relation to the quality variable X. Results are encouraging concerning criteria associated to units types, but combinatorial problems weigh heavy for criteria related to units instances.
Bibliographic reference. Francois, Hélène / Boeffard, Olivier (2003): "Evaluation of units selection criteria in corpus-based speech synthesis", In EUROSPEECH-2003, 1325-1328.