Synthesis of missing units in a Telugu text-to-speech system

Synthesis of complex consonant clusters is a challenge in TTS systems, as it is difficult to ensure proper coarticulation. This problem surfaces in the synthesis of missing units in unit concatenation systems. Back-off techniques are used to synthesize these missing units. The current back-off techniques, which use smaller sub-word units to synthesize the missing unit, are found unsuitable for frequent usage causing severe degradation in synthesis quality. However missing units are a frequent occurrence when synthesizing unrestricted text. Hence to build open domain TTS systems, efficient back-off techniques are necessary. This problem is further aggravated when working with under-resourced languages as they lack the manually annotated speech corpora which enable the development of these high quality synthesis techniques. In this thesis we propose a two pronged strategy to improve the quality of back-off techniques. The two pronged strategy consists of refining the boundaries of units in the acoustic inventory to reduce phone insertion errors during back-off and use of a novel back-off technique to ensure coarticulation consistency. A boundary refinement algorithm which removes necessity for manually labelled data by relying on a knowledge base of cost functions derived from landmark specific acoustic cues was proposed, to increase the quality of unit boundaries. However increase in accuracy of unit boundaries was insufficient when context-dependent units are not available in the database. Hence a back-off technique which synthesizes complex consonant clusters even when the required context-dependent units are not available, was proposed. This technique is motivated from a phenomenon observed during second language acquisition and emulates native speaker intuition in synthesis of these clusters. The technique was proven to be better than the conventional backoff techniques. Further it was proven to be robust to segmentation errors, reducing the reliance on high quality segmentation techniques which require manually labelled data. Thesis Supervisor: Dr. Kishore S. Prahallad Title: Assistant Professor