In this study artificial neural nets (ANN), relating articulation to acoustic features, are used as tools to investigate the ability of different vocal-tract models to describe the vocal tract. The vocal tract models are compared by means of a test material common to all evaluations. Different ANN configurations are used and investigated. Aspects of analysing natural speech signal with ANNs, trained on synthetic speech, are discussed. In addition, some effects of the speaker-dependency of the ANNs are investigated. The models are tested on synthetic speech (based on X-ray data of Swedish vowels) and natural speech from different speakers.

Keywords: Artificial neural networks, speech signal inversion, vocal tract model, speaker-dependency.