Investigating the Language Independence of Voice Activity Projection Models through Standardization of Speech Segmentation Labels

Voice Activity Projection (VAP) is crucial for ensuring natural turn-taking in dialogue systems. One known limitation in VAP is that, according to an experiment in a previous study, a model trained in one language cannot be applied to other languages. However, in that study, specific languages were associated with specific datasets, making it unclear whether the performance of the VAP model was influenced by the language or by the dataset. In this study, utilizing the Transformer-based VAP model, we conducted training and evaluation of the model using multiple datasets across several languages in order to determine what is affecting the performance. The results indicated that the performance of the VAP model is more dependent on the dataset than on the language. Specifically, differences in how speech segmentation labels are provided significantly impacted the performance. The results also showed that a model trained in one language can be applied to other languages, suggesting that the VAP model captures common features independent of language.