Baselines and Protocols for Household Speaker Recognition

View on arXiv ← Back to list

Authors: Alexey Sholokhov, Xuechen Liu, Md Sahidullah, Tomi Kinnunen

Published: 2022-04-30 15:04:56+00:00

AI Summary

This research paper introduces an evaluation benchmark and open-source baselines for household speaker recognition, addressing challenges like domain robustness, short utterances, and passive enrollment. It provides several algorithms for both active and passive enrollment scenarios.

Abstract

Speaker recognition on household devices, such as smart speakers, features several challenges: (i) robustness across a vast number of heterogeneous domains (households), (ii) short utterances, (iii) possibly absent speaker labels of the enrollment data (passive enrollment), and (iv) presence of unknown persons (guests). While many commercial products exist, there is less published research and no publicly-available evaluation protocols or open-source baselines. Our work serves to bridge this gap by providing an accessible evaluation benchmark derived from public resources (VoxCeleb and ASVspoof 2019 data) along with a preliminary pool of open-source baselines. This includes four algorithms for active enrollment (speaker labels available) and one algorithm for passive enrollment.

Key findings

A simple online update rule using weighted averaging of embeddings achieved competitive performance compared to offline algorithms. The proposed spherical covariance PLDA outperformed cosine scoring methods. Passive enrollment using SpeechBrain embeddings showed the best performance on the ASVspoof dataset.

Approach

The authors address household speaker recognition by developing several algorithms for active and passive enrollment. These algorithms use deep neural network embeddings (E-TDNN, ECAPA-TDNN, ResNet34) and various back-ends (cosine similarity, PLDA) to compare test utterances with speaker models. They also introduce modifications to existing clustering algorithms for passive enrollment.

Datasets

VoxCeleb (VoxCeleb1, VoxCeleb2), ASVspoof 2019

Model(s)

E-TDNN, ECAPA-TDNN, ResNet34; PLDA (with spherical covariances), cosine similarity

Author countries

Finland, France

← Previous