i-Vector based text-independent speaker verification (SV) systems often
have poor performance with short utterances, because the biased phonetic
distribution in a short utterance makes the extracted i-vector unreliable.
This paper proposes an i-vector compensation method using a generative
adversarial network (GAN), where its generator network is trained to
transform an unreliable i-vector from a short utterance into a reliable one
which can only be extrated from a long utterance and its discriminator
network is trained to determine whether an i-vector is from the generator or
from a long utterance. Additionally, we assign two other learning tasks to
the GAN to stabilize its training and to make the generated i-vector more
speaker-specific. Speaker verification experiments conducted on the NIST
SRE 2008 “short2-10sec” and “10sec-10sec” conditions show that our
method can help reduce the average equal error rate of the conventional
i-vector and PLDA system.
各種検索
サポート
ヘルプ