This paper describes a hidden Markov model (HMM)-based text-to-speech (TTS) system developed at the Nagoya Institute of Technology (NITECH) for the Blizzard Challenge 2015. The tasks of the Blizzard Challenge 2015 are speech synthesis for six Indian languages and multilingual involving one Indian language and English. In this challenge, only Indian language speech data and text are provided as training data. Therefore, pronunciation information, such as phoneset, phoneme sequences, and lexicon, for each language is needed to construct TTS systems. Standard methods for constructing a TTS system of a new language take a huge cost because it requires a special knowledge of the target language. In this paper, we focus on automatic construction of a TTS system without the special knowledge of the target language. The results of a large-scale subjectivity evaluation are discussed.
@inproceedings{sawada15_blizzard, title = {The NITECH HMM-based text-to-speech system for the Blizzard Challenge 2015}, author = {Kei Sawada and Kei Hashimoto and Keiichiro Oura and Keiichi Tokuda}, year = {2015}, booktitle = {The Blizzard Challenge 2015}, pages = {23--28}, doi = {10.21437/Blizzard.2015-3},}Cite as:Sawada, K., Hashimoto, K., Oura, K., Tokuda, K. (2015) The NITECH HMM-based text-to-speech system for the Blizzard Challenge 2015. Proc. The Blizzard Challenge 2015, 23-28, doi: 10.21437/Blizzard.2015-3