Challenges in spoken language diarization in code-switched scenario

Spoken language diarization (SLD) is a task of automatically annotating the monolingual segments in a given code-switched (CS) speech signal. Most of the SLD methods in the literature are linked to the task like automatic speech recognition and no independent frameworks exist. One attractive direction to evolving an independent framework is to explore speaker diarization (SD). However, there are several challenges in doing the same due to the production of the CS speech by a single speaker. This work uses the speaker change detection (SCD) framework of SD to illustrate the challenges. The initial experiments demonstrate that the performance of language change detection (LCD) is significantly poor compared to SCD. The degradation is mostly due to the significant overlapping of the features belonging to two languages in the feature space. Trained adaptation models are used to compute the Gaussian mixture model posterior (GP) features and observed that the overlapping is significantly lower compared to the acoustic features. Hence the GP feature provides a relative improvement of 20% for LCD in terms of identification rate (IDR). Thus using existing SD frameworks, to begin with, and modifying them to increase the discrimination between languages in the CS scenario seems to be a good way to evolve independent frameworks.