Prosodic Reading Style Simulation for Text-to-Speech Synthesis

摘要

The simulation of different reading styles (mainly by adapting prosodic parameters) can improve the naturalness of synthetic speech and supports a more intelligent human machine interaction. The article exemplarily investigates the reading styles News and Tale. For comparison, all examined texts contained the same genre-neutral paragraphs which have been read without a specific style instruction: Normal but also faster, slower, rather monotone or more emotional which led to corresponding artificial styles. The measured original intonation and durations style patterns control a diphone synthesizer (mapped contours). Additionally, the patterns are used to train a neural network (NN) model. Within two separate listening tests, different stimuli presented as original signal/style, respectively, with mapped or NN generated prosodic contours have been evaluated. The results show that both, original utterances and artificial styles are basically perceived in their intended reading styles. Some reciprocal confusions indicate the similarities between different styles like News and Fast, Tale and Slow as well as Tale and Expressive. The confusions are more likely for synthetic speech. To produce e. g. the complex style Tale, different features of the prosodic variations Slow and Expressive are combined. The training method for the synthetic styles requires a further improvement.