Distributed representation of melodic contours

Abstract

We introduce a new computational model for melodic contours-melody embeddings. It is based on the approach of distributional semantics where embeddings represent units as continuous vectors in a multi-dimensional space based on hypothesis that units with similar meaning are used in similar contexts. This paradigm is applied to melodic contours and their segments. Melodic contours are represented by vectors of the same dimensionality independent on their length and shape. We successfully evaluated the ability of the proposed model to measure the distance between melodic contours. The results of applying the model for a task of prominent words detection have not showed the improvement over traditional prosodic features. Nevertheless we assume the model to be very promising. The possible applications for the proposed unsupervised prosodic model include processing of speech of underresourced languages, modelling prosodic variability for textto-speech synthesis, recognition and classification of prosodic events by means of deep-learning algorithms.

Language-Dependent Melody Embeddings

“…Thus we propose to apply the embeddings approach to the task of language comparison. The melody embedding model was introduced in [7]. The novelty of the approach is to train multidimensional melody embeddings purely on melodic infor-mation.…”