Environmental domain embedding for far-field speaker verification

Ze-xi Zhu; Ye Jiang

doi:10.1117/12.2628416

21 April 2022 Environmental domain embedding for far-field speaker verification

Ze-xi Zhu, Ye Jiang

Proceedings Volume 12175, International Conference on Network Communication and Information Security (ICNCIS 2021); 121750B (2022) https://doi.org/10.1117/12.2628416
Event: International Conference on Network Communication and Information Security (ICNCIS 2021), 2021, Beijing, China

Abstract

Speaker embedding based on neural network has achieved significant performance in speaker verification. However, this method results maladaptation of far-field speech. To fix this problem, this paper proposes applying environmental domain embedding to far-field speaker verification. By mapping speech environment to domain embedding and combining it with speaker features, the model is guided to detect the information of environment and speaker identity, thus improves the performance of far-field recognition. Comparing to traditional single domain and single embedding depth method, we explore the effects of domain type and embedding depth from gender, speech recording distance and noise. Furthermore, embeddings are applied to convolution through reshaping rather than vanilla fully-connected layers. The improved embedding method allows better combination. Experiments show the EER (Equal Error Rate) of gender-based, distance-based and noise-based model, comparing to baseline (ResNet-34, EER=1.28%), respectively decrease to 1.12%, 0.92% and 0.88%. And further comparison tests discover that carrying out combination in the second or third model stage achieves better performance.

Citation Download Citation

Ze-xi Zhu and Ye Jiang "Environmental domain embedding for far-field speaker verification", Proc. SPIE 12175, International Conference on Network Communication and Information Security (ICNCIS 2021), 121750B (21 April 2022); https://doi.org/10.1117/12.2628416

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available