Predicting Questions in Swedish Intonation Data : Examining the use of a final rising intonation pattern in Swedish questions

Sammanfattning

Question specific intonation patterns in Swedish are often described as having a phrase-final rise. The use of this marker is optional, and some studies link it to certain question types. This study investigates how often such patterns occur in interview dialogues, and their impact on how machine learning models trained on different features can predict questions in speech. The models were trained on sequences of normalized f0-values from 1522 sentences. They were also trained on the semantic embeddings of each sentence using a Swedish Sentence-BERT model. A phonetic analysis investigating the frequency of a perceptible rise in the phrase-final syllable of 189 questions from the same data found that it only occurs in 8.6% of questions, with no significant difference in frequency between question types. In the computational analysis, none of the models accurately classified questions based on intonation alone. The Logistic Regression model achieved the highest accuracy at 54.5%, marginally above the chance level. Models trained on semantic embeddings achieved better accuracy (66-72.4%), but were not optimal. These findings suggest that question-specific intonation patterns are rare in Swedish, limiting machine learning models’ ability to predict questions from intonation alone.