A Quantitative Model of Yorùbá Speech Intonation Using Stem-ML

We present a quantitative model of Standard Yorùbá (SY) intonation; it is designed to have parameters that are linguistically interpretable. The model is built and trained on speech data from a native speaker of SY. The resulting model reproduces the data well: its Root Mean Square prediction error (RMSE) is 14.00 Hz on a test set. We find that intonation is used to mark sentence and phrase boundaries: beginning syllables are systematically stronger, while ending syllables are systematically weaker than the medial syllables. The M tone is the strongest and the H tone is the weakest, though the differences are modest. We see comparable amounts of carry-over and anticipatory co-articulation. The resulting model for SY shows similar characteristics when compared to Mandarin and Cantonese intonation models.