Showing 1–1 of 1 results for author: C, K K
-
openFEAT: Improving Speaker Identification by Open-set Few-shot Embedding Adaptation with Transformer
Authors:
Kishan K C,
Zhenning Tan,
Long Chen,
Minho **,
Eunjung Han,
Andreas Stolcke,
Chul Lee
Abstract:
Household speaker identification with few enrollment utterances is an important yet challenging problem, especially when household members share similar voice characteristics and room acoustics. A common embedding space learned from a large number of speakers is not universally applicable for the optimal identification of every speaker in a household. In this work, we first formulate household spe…
▽ More
Household speaker identification with few enrollment utterances is an important yet challenging problem, especially when household members share similar voice characteristics and room acoustics. A common embedding space learned from a large number of speakers is not universally applicable for the optimal identification of every speaker in a household. In this work, we first formulate household speaker identification as a few-shot open-set recognition task and then propose a novel embedding adaptation framework to adapt speaker representations from the given universal embedding space to a household-specific embedding space using a set-to-set function, yielding better household speaker identification performance. With our algorithm, Open-set Few-shot Embedding Adaptation with Transformer (openFEAT), we observe that the speaker identification equal error rate (IEER) on simulated households with 2 to 7 hard-to-discriminate speakers is reduced by 23% to 31% relative.
△ Less
Submitted 24 February, 2022;
originally announced February 2022.
