By transforming these phonetic posteriorgrams into clustered phoneme embeddings, we predict natural fundamental frequency patterns without requiring transcriptions. Our experiments demonstrate that ...