Short Test Utterances Research Articles

Exploring kernel discriminant analysis for speaker verification with limited test data

Speaker verification (SV) with limited test data condition is desirable for practical application oriented systems. The i-vector based speaker modeling has shown its significance for SV tasks, but its performance degrades as the utterance becomes shorter. The i-vectors apart from being compact and dominant speaker representations, bear channel and session information, which has to be compensated for robust speaker modeling. The conventional techniques for channel/session compensation include linear discriminant analysis (LDA) followed by within class covariance normalization (WCCN) and Gaussian probabilistic linear discriminant analysis (GPLDA) that eliminate the channel/session variation across the i-vectors by assuming these are linearly separable. In this work, a novel method for channel/session compensation is proposed using kernel discriminant analysis (KDA) that projects the i-vectors into a higher dimensional space and performs discriminant analysis to remove the unwanted information for speaker modeling. The SV studies are performed on standard NIST speaker recognition evaluation (SRE) 2003 and 2008 databases that convey the significance of the proposed compensation over the conventional methods, which is greater on using short test utterances. The achieved improvements are hypothesized due to the non-linearities of channel/session information in the i-vector domain.

짧은 음성을 대상으로 하는 화자 확인을 위한 심층 신경망

본 논문에서는 짧은 테스트 발성에 대한 화자 확인 성능을 개선하는 방법을 제안한다. 테스트 발성의 길이가 짧을 경우 i-벡터/확률적 선형판별분석 기반 화자 확인 시스템의 성능이 하락한다. 제안한 방법은 짧은 발성으로부터 추출한 특징 벡터를 심층 신경망으로 변환하여 발성 길이에 따른 변이를 보상한다. 이 때, 학습시의 출력 레이블에 따라 세 종류의 심층 신경망 이용 방법을 제안한다. 각 신경망은 입력 받은 짧은 발성 특징에 대한 출력 결과와 원래의 긴 발성으로부터 추출한 특징과의 차이를 줄이도록 학습한다. NIST (National Institute of Standards Technology, 미국) 2008 SRE(Speaker Recognition Evaluation) 코퍼스의 short 2-10 s 조건 하에서 제안한 방법의 성능을 평가한다. 실험 결과 부류 내 분산 정규화 및 선형 판별 분석을 이용하는 기존 방법에 비해 최소 검출 비용이 감소하는 것을 확인하였다. 또한 짧은 발성 분산 정규화 기반 방법과도 성능을 비교하였다. We propose a method to improve the robustness of speaker verification on short test utterances. The accuracy of the state-of-the-art i-vector/probabilistic linear discriminant analysis systems can be degraded when testing utterance durations are short. The proposed method compensates for utterance variations of short test feature vectors using deep neural networks. We design three different types of DNN (Deep Neural Network) structures which are trained with different target output vectors. Each DNN is trained to minimize the discrepancy between the feed-forwarded output of a given short utterance feature and its original long utterance feature. We use short 2-10 s condition of the NIST (National Institute of Standards Technology, U.S.) 2008 SRE (Speaker Recognition Evaluation) corpus to evaluate the method. The experimental results show that the proposed method reduces the minimum detection cost relative to the baseline system.

Frame-by-frame language identification in short utterances using deep neural networks

This work addresses the use of deep neural networks (DNNs) in automatic language identification (LID) focused on short test utterances. Motivated by their recent success in acoustic modelling for speech recognition, we adapt DNNs to the problem of identifying the language in a given utterance from the short-term acoustic features. We show how DNNs are particularly suitable to perform LID in real-time applications, due to their capacity to emit a language identification posterior at each new frame of the test utterance. We then analyse different aspects of the system, such as the amount of required training data, the number of hidden layers, the relevance of contextual information and the effect of the test utterance duration. Finally, we propose several methods to combine frame-by-frame posteriors. Experiments are conducted on two different datasets: the public NIST Language Recognition Evaluation 2009 (3 s task) and a much larger corpus (of 5 million utterances) known as Google 5M LID, obtained from different Google Services. Reported results show relative improvements of DNNs versus the i-vector system of 40% in LRE09 3 second task and 76% in Google 5M LID.