Magnitude-aware Probabilistic Speaker Embeddings

TL;DR

This paper introduces a probabilistic speaker embedding method that utilizes embedding magnitudes to improve speaker verification and diarization, demonstrating significant performance gains over traditional magnitude-agnostic approaches.

Contribution

It proposes a novel magnitude-aware probabilistic embedding extraction technique and quality-aware diarization methods that leverage embedding magnitudes for enhanced performance.

Findings

Significant improvements in speaker verification accuracy.

Enhanced diarization performance using magnitude-based quality measures.

Embedding magnitudes correlate with input quality and out-of-distribution detection.

Abstract

Recently, hyperspherical embeddings have established themselves as a dominant technique for face and voice recognition. Specifically, Euclidean space vector embeddings are learned to encode person-specific information in their direction while ignoring the magnitude. However, recent studies have shown that the magnitudes of the embeddings extracted by deep neural networks may indicate the quality of the corresponding inputs. This paper explores the properties of the magnitudes of the embeddings related to quality assessment and out-of-distribution detection. We propose a new probabilistic speaker embedding extractor using the information encoded in the embedding magnitude and leverage it in the speaker verification pipeline. We also propose several quality-aware diarization methods and incorporate the magnitudes in those. Our results indicate significant improvements over magnitude-agnostic baselines both in speaker verification and diarization tasks.