| Title | : | Speaker Modelling with i-Vectors |
| Speaker | : | Srikanth M.R (IDIAP, Switzerland) |
| Details | : | Tue, 9 Jan, 2018 3:00 PM @ A M Turing Hall |
| Abstract: | : | The demand for robust speech technology applications has been exponentially increasing, thanks especially to the booming smart phone and smart devices market. A common problem faced in systems developed for speech processing is speaker recognition, that is to automatically identify who is speaking, which is helpful to provide and improve personalized services (e.g. hotword based authentication). In case of multiple speakers, it is also necessary to find out who speaks when, also called speaker diarization, which is often the case in recordings of business meetings. This talk will address these two research problems with focus on i-vector based speaker modeling techniques. In recent years, i-vector extraction from audio, which produces a fixed length low-dimensional representation of a variable length sequence, has received tremendous attention due to its success in speaker recognition. In this talk, a fast and scalable training algorithm based on Probabilistic Principal Component Analysis for rapid prototyping of speaker recognition systems is presented. Next, i-vector modeling is applied for text dependent speaker recognition and diarization. We also analyze the integration of Automatic Speech Recognition systems in the i-vector paradigm. Finally, the problem of fusion of speaker recognition with other speech technologies (e.g. language recognition), which is useful for intelligence gathering systems used by law enforcement agencies, is addressed. Brief Biography of Speaker: Srikanth got his Ph.D. in Computer Science and Engineering from Indian Institute of Technology Madras in 2013. During his Ph.D., he worked on automatic speaker recognition and spoken keyword spotting. He is currently working as a Postdoctoral Researcher at IDIAP, Switzerland, since August, 2013. At IDIAP, Srikanth worked on speaker diarization in the DIMHA project, speaker recognition for the Speaker Identification Integrated Project as a lead developer and researcher and, more recently, low-resource Automatic Speech Recognition for the SARAL project. His current research interests include - automatic speaker recognition, speaker diarization, low-resource speech recognition and summarization. |
