Pitch, loudness, and segmental duration correlates: towards a model for the phonetic aspects of Finnish prosody

Neural networks are used widely today for modeling a variety of different aspects of spoken language. We use them to model Finnish lexical prosody with an aim to shed light on the interaction between the main prosodic parameters: segmental durations, loudness and pitch. We have analyzed the performance of a group of networks which were all trained to generate values for a different prosodic parameter given similar input information. The experiments were performed on speech material contained within our Finnish speech database.