Unsupervised machine learning for the accurate classification of the discourse marker like in code-switching utterances

Abstract

Spanish-English bilinguals use the discourse marker like in English, Spanish, and code-switching utterances. An acoustic analysis found tat the [l] and diphthong in like is produced differently depending on the type of utterance in which it occurs. To investigate how exactly these differences are manifested in the acoustic signal, we built a logistic-polynomial regression model to classify like tokens based on acoustic data. The model first projects F1 and F2 values onto a space of time-dependent polynomials. We then apply multinomial logistic regression to classify these polynomials as English, Spanish, or code-switching. The area under the curve was 0.75, showing classification was significantly greater than random. This model outperforms a model that rely on static values for F1 and F2 and a model based on independent component analysis. The superiority of the polynomial model suggests that the time-dependent progression of F1 and F2 values, rather tan absolute formant values, is useful for predicting an imminent code-switch. By building and comparing additional models to investigate other acoustic aspects of the signal we can be more informed when building perception experiments to see what aspects of the signal listeners are more likely to use anticipating an upcoming code-switch. © 2013 Acoustical Society of America.