Statistical parametric synthesis of budgerigar songs

In this paper we present the synthesis of budgerigar songs with Hidden Markov Models (HMMs) and the HMM-based Speech Synthesis System (HTS). Budgerigars can produce complex and diverse sounds that are difficult to categorize. We adapted tech- niques that are commonly used in the area of speech synthe- sis so that we can use them for the synthesis of budgerigar songs. To segment the recordings, the songs are broken down into phrases, which are sounds separated by silence. Complex phrases furthermore can be subdivided into smaller units and then be clustered to identify recurring elements. These ele- ment categories along with additional contextual information are used together to enhance the training and synthesis. Over- all, the aim of the process is to offer an interface that gener- ates new sequences and compositions of bird songs based on user input, consisting of the desired song structure and contex- tual information. Finally, an objective evaluation comparing the synthesized output to the natural recording is performed, and a subjective evaluation with human listeners shows that they pre- fer resynthesized over natural recordings and that they perceive no significant differences in terms of naturalness between natu- ral, resynthesized, and synthesized versions.