查看文章

In this article development of a PUEPS corpus is described. This dataset contains recordings of the acted emergency telephone conversations. Speakers that participated in the experiments reported crime scenes that were presented to them in a form the earlier prepared movies. Recording sessions were performed in the laboratory conditions. To each conversation metadata that summarize information about the speaker, conversation, and the reported event were added. Moreover, manually prepared transcriptions enriched with tags describing paralinguistic phenomena are also a part of the described corpus. These transcriptions were made using tools prepared by the authors for fast and convenient work due to: prompting, annotation, and data management mechanisms. The transcription experiments showed substantial improvement of the work efficiency and speed. Final multilevel speaker recognition experiments proved that the accuracy of the speaker recognition is noticeably improved due to the use of transcriptions and the linguistic level analysis.