On Open-Set Speaker Identification with I-Vectors

Kevin Wilkinghoff

Open-set speaker identification systems first need to decide if an utterance belongs to one of the known so called blacklist speakers and second identify the exact blacklist speaker. In this paper, an open-set speaker identification system based on i-vectors is presented. The system consists of an outlier detector in combination with a classical closed-set speaker identification chain and utilizes an effective preprocessing technique for i-vectors, called linear alignment. Its overall structure is justified both theoretically and experimentally by comparing multiple outlier detectors. In experimental evaluations, our proposed system reaches an improvement of 37.5% for the top-S Equal Error Rate (EER) and a 50% lower top-1 EER over the baseline system of the 1st Multi-target speaker detection and identification Challenge Evaluation and improves upon all other published results obtained on this dataset.

 DOI: 10.21437/Odyssey.2020-58

Cite as: Wilkinghoff, K. (2020) On Open-Set Speaker Identification with I-Vectors. Proc. Odyssey 2020 The Speaker and Language Recognition Workshop, 408-414, DOI: 10.21437/Odyssey.2020-58.

  author={Kevin Wilkinghoff},
  title={{On Open-Set Speaker Identification with I-Vectors}},
  booktitle={Proc. Odyssey 2020 The Speaker and Language Recognition Workshop},