Normalising Flows for Speaker and Language Recognition Backend

Normalising Flows for Speaker and Language Recognition Backend

Stať ve sborníku mimo WoS a Scopus

In this paper, we address the Gaussian distribution assumption made in PLDA, a popular back-end classifier used in Speaker and Language recognition tasks. We study normalizing flows, which allow using non-linear transformations and still obtain a model that can explicitly represent a probability density. The model makes no assumption about the distribution of the ob- servations. This alleviates the need for length normalization, a well known data preprocessing step used to boost PLDA performance. We demonstrate the effectiveness of this flow model on NIST SRE16, LRE17 and LRE22 datasets. We ob- serve that when applying length normalization, both the flow model and PLDA achieve similar EERs for SRE16 (11.5% vs 11.8%). However, when length normalization is not applied, the flow shows more robustness and offers better EERs (13.1% vs 17.1%). For LRE17 and LRE22, the best classification accu- racies (84.2%, 75.5%) are obtained by the flow model without any need for length normalization.

In this paper, we address the Gaussian distribution assumption made in PLDA, a popular back-end classifier used in Speaker and Language recognition tasks. We study normalizing flows, which allow using non-linear transformations and still obtain a model that can explicitly represent a probability density. The model makes no assumption about the distribution of the ob- servations. This alleviates the need for length normalization, a well known data preprocessing step used to boost PLDA performance. We demonstrate the effectiveness of this flow model on NIST SRE16, LRE17 and LRE22 datasets. We ob- serve that when applying length normalization, both the flow model and PLDA achieve similar EERs for SRE16 (11.5% vs 11.8%). However, when length normalization is not applied, the flow shows more robustness and offers better EERs (13.1% vs 17.1%). For LRE17 and LRE22, the best classification accu- racies (84.2%, 75.5%) are obtained by the flow model without any need for length normalization.

Speaker recognition, Language Recognition

Speaker recognition, Language Recognition

ESPUNA, A.; PRASAD, A.; MOTLÍČEK, P.; MADIKERI, S.; SCHUEPBACH, C.

18.06.2024

International Speech Communication Association

Quebec

Proceedings of Odyssey 2024: The Speaker and Language Recognition Workshop

74

80

7

https://www.isca-archive.org/odyssey_2024/espuna24_odyssey.pdf

espuna2024_odyssey