Boosting the Performance of I-Vector Based Speaker Verification via Utterance Partitioning

Abstract

Abstract-The success of the recent i-vector approach to speaker verification relies on the capability of i-vectors to capture speaker characteristics and the subsequent channel compensation methods to suppress channel variability. Typically, given an utterance, an i-vector is determined from the utterance regardless of its length. This paper investigates how the utterance length affects the discriminative power of i-vectors and demonstrates that the discriminative power of i-vectors reaches a plateau quickly when the utterance length increases. This observation suggests that it is possible to make the best use of a long conversation by partitioning it into a number of sub-utterances so that more ivectors can be produced for each conversation. To increase the number of sub-utterances without scarifying the representation power of the corresponding i-vectors, repeated applications of frame-index randomization and utterance partitioning are applied. Results on NIST 2010 speaker recognition evaluation (SRE) suggest that (1) using more i-vectors per conversation can help to find more robust linear discriminant analysis (LDA) and withinclass covariance normalization (WCCN) transformation matrices, especially when the number of conversations per training speaker is limited; and (2) increasing the number of i-vectors per target speaker helps the i-vector based support vector machines (SVM) to find better decision boundaries, thus making SVM scoring outperforms cosine distance scoring by 19% and 9% in terms of minimum normalized DCF and EER.

Robust scream sound detection via sound event partitioning

“…This is because for medium and long sound events, there must be some spectral variations within the events but the means and standard deviations fail to capture these sub-event variations. To address this deficiency, we extend our recently proposed utterance partitioning technique [23,31] to sound-event detection. The partitioning procedure is as follows:…”

“…This step follows the argument in [23,31] that the mean and standard deviation will not be affected by rearranging the indexes.…”

“…Because the amount of scream sound data is much smaller than that of non-scream sounds, there is a severe imbalance between the two classes of data. Recently, a technique called utterance partitioning [23,31] has been developed for speaker verification to address this issue. Here, we extend the technique to sound-event detection and refer to it as sound event partitioning (SEP).…”

Likelihood-ratio empirical kernels for i-vector based PLDA-SVM scoring

“…This difficulty, however, can be overcome by a technique called utterance partitioning with acoustic vector resampling (UP-AVR) [11][12][13]. This technique has been successfully applied to both GMM-SVM [14][15][16] and i-vector based systems [17].…”

“…This frame-index randomization and partitioning process can be repeated several times to produce a desirable number of i-vectors for each conversation. It has been demonstrated in [17] that increasing the number of target-speaker i-vectors can help the SVM training algorithm to find better decision boundaries, thus making SVM scoring outperforms cosine-distance scoring. In this paper, we further demonstrate that UP-AVR is indispensable for training the speaker-dependent empirical LR SVMs.…”