Investigation of intra-speaker spectral parameter variation and its prediction towards improvement of spectral conversion metric

In spectral conversion of statistical voice conversion (VC), distance-based measures between the converted and target spectral parameters are often used as evaluation or training criteria. However, even if the same speaker utters the same sentence, the spectral parameters vary utterance by utterance, and thus, spectral distance between utterances still exists. Moreover, the original prosodic features of input speech are often kept unchanged in some VC systems, such those that function in realtime. In such cases, prosody of converted and target speech samples are different, and these differences increases spectral distance. These potential spectral variations are not considered in the conventional evaluation/training criterion. Thus, by constructing criteria that consider this spectral difference improvements in sound quality can be expected. In this paper, we investigate intra-speaker spectral variation between utterances of the same sentence. We also propose a method for predicting this variation from prosodic parameter differences between the corresponding utterances. We conduct experimental evaluations using many speech samples of the same sentence uttered by a single speaker, with results demonstrating that the proposed method effectively predicts the intra-speaker spectral variation from the observed prosodic changes.