Language-Dependent Melody Embeddings

Daniil Kocharov, Alla Menshikova

The paper explores the perspectives of applying the distributional approach to prosodic typology of languages. The method discussed here is an adaptation of the distributional semantics approach, as suggested by Mikolov, to melodic features of speech. The paper contains a detailed description of the new method, as well as a comparison of five European languages (English, Czech, German, Russian and Finnish) in terms of melody embeddings. The total amount of speech data was over 500 hours. The experimental results show that melody embeddings are language dependent. The proposed melody embedding model has shown reasonable results in language comparison.

 DOI: 10.21437/Interspeech.2018-1962

Cite as: Kocharov, D., Menshikova, A. (2018) Language-Dependent Melody Embeddings. Proc. Interspeech 2018, 2167-2170, DOI: 10.21437/Interspeech.2018-1962.

  author={Daniil Kocharov and Alla Menshikova},
  title={Language-Dependent Melody Embeddings},
  booktitle={Proc. Interspeech 2018},