CONF Sarfjoo_SPSC_2021/IDIAP Graph2Speak: Improving Speaker Identification using Network Knowledge in Criminal Conversational Data Fabien, Mael Sarfjoo, Seyyed Saeed Madikeri, Srikanth Motlicek, Petr https://publications.idiap.ch/index.php/publications/showcite/Fabien_Idiap-RR-01-2023 Related documents 1st ISCA Symposium on Security and Privacy in Speech Communication 2021 10--13 10.21437/SPSC.2021-3 doi Criminal investigations mostly rely on the collection of speech conversational data in order to identify speakers and build or enrich an existing criminal network. Social network analysis tools are then applied to identify the central characters and the different communities within the network. This paper introduces a new method, Graph2Speak, to re-rank individuals after applying a speaker identification step, by leveraging the frequency of previous interactions extracted from a graph. We deploy our method on two candidate datasets for criminal conversational data, Crime Scene Investigation (CSI), a television show, and the ROXANNE simulated data. We demonstrate that our method can reduce the error rates of the speaker identification baseline by up to 12% (relative). REPORT Fabien_Idiap-RR-01-2023/IDIAP Graph2Speak: Improving Speaker Identification using Network Knowledge in Criminal Conversational Data Fabien, Mael Sarfjoo, Seyyed Saeed Madikeri, Srikanth Motlicek, Petr EXTERNAL https://publications.idiap.ch/attachments/reports/2020/Fabien_Idiap-RR-01-2023.pdf PUBLIC Idiap-RR-01-2023 2023 Idiap 19 rue Marconi, 1920 Lausanne January 2023 Criminal investigations mostly rely on the collection of speech conversational data in order to identify speakers and build or enrich an existing criminal network. Social network analysis tools are then applied to identify the most central characters and the different communities within the network. We introduce two candidate datasets for criminal conversational data, Crime Scene Investigation (CSI), a television show, and the ROXANNE simulated data. We also introduce the metric of conversation accuracy in the context of criminal investigations. By re-ranking candidate speakers based on the frequency of previous interactions, we improve the speaker identification baseline by 1.2% absolute (1.3% relative), and the conversation accuracy by 2.6% absolute (3.4% relative) on CSI data, and by 1.1% absolute (1.2% relative), and 2% absolute (2.5% relative) respectively on the ROXANNE simulated data. https://arxiv.org/abs/2006.02093 URL