Patrick J Kenny

Timeline

1. 2004

   Factor analysis for speaker verification

   In 2004, he and Pierre Dumouchel report speaker verification experiments that use factor analysis likelihood ratios on NIST 1999 and NIST 2000 test sets at the Odyssey workshop. On a matched NIST 2000 setup they reach an equal error rate of 7.2%, about 25% better than standard methods.

2. 2008

   The primary CRIM system

   In 2008, he coauthors the Interspeech paper on the primary CRIM system for the NIST 2008 speaker recognition evaluation, with Najim Dehak, Pierre Ouellet, Vishwa Gupta and Pierre Dumouchel. The paper describes how the CRIM factor analysis speaker verification system is modified for the new cross-channel conditions of that evaluation.

3. 2008

   A Johns Hopkins team report

   Also in 2008, he is one of the authors of the 81-page Johns Hopkins University report "Robust Speaker Recognition Over Varying Channels", written with Lukáš Burget, Niko Brümmer, Douglas Reynolds, Najim Dehak and others.

4. 2010

   Bayesian heavy-tailed priors

   In 2010, he is the sole author of "Bayesian Speaker Verification with Heavy-Tailed Priors" at the Odyssey workshop. The model represents each utterance as a low-dimensional vector and uses heavy-tailed distributions, and on NIST 2008 telephone speech its results are about 30% better than joint factor analysis.

5. 2011

   The total variability space

   In 2011, he coauthors "Front-End Factor Analysis for Speaker Verification" with Najim Dehak, Réda Dehak, Pierre Dumouchel and Pierre Ouellet, published in IEEE Transactions on Audio, Speech, and Language Processing. The paper defines a low-dimensional total variability space that models speaker and channel variability, and reports an equal error rate of 1.12% on the male English core trials of the NIST 2008 evaluation.

6. 2014

   Text-dependent joint factor analysis

   In 2014, he and four coauthors, including Themos Stafylakis, apply Joint Factor Analysis to text-dependent speaker verification at the Odyssey workshop. With a large UBM and the model decomposed into a feature extractor and a simple back end classifier, the system reaches an equal error rate of 0.5% on the RSR2015 evaluation set.

7. 2018–2019

   Deep speaker embeddings

   In 2018, he coauthors "Deeply Fused Speaker Embeddings for Text-Independent Speaker Verification" with Gautam Bhattacharya, Md Jahangir Alam and Vishwa Gupta; it combines neural speaker embeddings with i-vectors and lowers the error rate to 2.22% with a Joint Bayesian classifier. In 2019, he coauthors "Deep Speaker Recognition: Modular or Monolithic?" with Bhattacharya and Alam, whose end-to-end models reach state-of-the-art performance on VoxCeleb.