The Speech Alignment Paradox

Applying a recently presented text-independent speech alignment technique based on unit selection to the training of a voice conversion system suggested that the more training data was available, the less speaker-specific information was learned. This paradoxical effect contradicts experience we have from other corpus-based applications as speech recognition or synthesis. There, the performance usually gains with increasing amount of data. In this paper, we investigate this paradox by means of several experiments and derive a mathematical proof for a special case of the speech alignment paradox. Index Terms: speech processing, speech alignment