Data pruning using confidence measures for concatenative synthesis system built using automatically transcribed audio

Today, we can record and store large amounts of single speaker audio data, and also download it from the web. Generally, these data are prosodically rich and can therefore act as excellent candidates for building concatenative text-to-speech (TTS) systems. But transcritpions for these audio data are often not available and automatic transcriptions are error prone. In addition, these audio data contain bad acoustic (poorly articulated, noisy, inaudible, unintelligible, clipped) regions. Both above reasons can damage the resulting synthesized voice. So, pruning bad data becomes necessary. In this paper, we describe the development of two concatenative TTS systems using a lecture speech downloaded from Coursera and an audiobook downloaded from Librivox. Confidence measures such as phone posterior probability and unit duration obtained from the ASR system are used to remove bad data. Voices built using automatic transcripts are compared with those built using reference transcripts, and the effect of data pruning is investigated in terms of intelligibility and naturalness with the help of perceptual evaluation on Blizzard 2013 test corpus.