English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Tools annual

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Speaker embedding is an important front-end module to explore discriminativespeaker features for many speech applications where speaker information isneeded. Current SOTA backbone networks for speaker embedding are designed toaggregate multi-scale features from an utterance with multi-branch networkarchitectures for speaker representation. However, naively adding many branchesof multi-scale features with the simple fully convolutional operation could notefficiently improve the performance due to the rapid increase of modelparameters and computational complexity. Therefore, in the most currentstate-of-the-art network architectures, only a few branches corresponding to alimited number of temporal scales could be designed for speaker embeddings. Toaddress this problem, in this paper, we propose an effective temporalmulti-scale (TMS) model where multi-scale branches could be efficientlydesigned in a speaker embedding network almost without increasing computationalcosts. The new model is based on the conventional TDNN, where the networkarchitecture is smartly separated into two modeling operators: achannel-modeling operator and a temporal multi-branch modeling operator. Addingtemporal multi-scale in the temporal multi-branch operator needs only a littlebit increase of the number of parameters, and thus save more computationalbudget for adding more branches with large temporal scales. Moreover, in theinference stage, we further developed a systemic re-parameterization method toconvert the TMS-based model into a single-path-based topology in order toincrease inference speed. We investigated the performance of the new TMS methodfor automatic speaker verification (ASV) on in-domain and out-of-domainconditions. Results show that the TMS-based model obtained a significantincrease in the performance over the SOTA ASV models, meanwhile, had a fasterinference speed.

TMS: A Temporal Multi-scale Backbone Design for Speaker Embedding