Speaker identification from film dialogues

Abstract

A considerable research on speaker identification from speech dialogues has been conducted. However, the present article reports speaker identification task from film dialogue which includes textual features only. We have used a text corpus of film scripts, extracted from the IMSDb archive and annotated with speakers. As different characters in a script exhibit various linguistic styles while speaking, we have incorporated the stylistic features that are extracted from the turns of various characters. These features include frequency based information commonly used in author identification task as well as the part-of-speech (POS) information of turn text. The k-nearest neighbor (k-NN), Naïve Bayes' (NB) and Conditional Random Field (CRF) classifiers have been employed for identifying the speakers from the dialogue scripts at turn levels. All the supervised classifiers outperform the baseline system that was developed using a simple randomization technique.