Evaluation of BIC-based algorithms for audio segmentation

BIC-Based Speaker Segmentation Using Divide-and-Conquer Strategies With Application to Speaker Diarization

“…The window repeatedly grows by N g samples until a change point is detected or its size exceeds a pre-defined upper bound N max . Here, the upper bound ensures the search efficiency [15], [13]. If a change point is detected during the window growing step, the detection process restarts at that change point with an analysis window of N ini samples.…”

“…whereμ,μ X , andμ Y are, respectively, the sample mean vectors of Z, X , and Y;Σ,Σ X , andΣ Y are, respectively, the sample covariance matrices of Z, X , and Y; and d is the dimension of the feature vector [13]. The larger the value of ∆BIC, the less similar the two segments will be; thus, the larger the distance between the two segments will be.…”

“…Zhou and Hansen [17] used Hotelling's T 2 -Statistic, which has the advantage of low computational cost, as the distance measure in the WinGrow detection process, and only used ∆BIC to verify the acoustic change candidates. In addition, [15] and [13] proposed more efficient implementations for the ∆BIC computation that do not affect the segmentation accuracy. 2) In FixSlid, a certain distance measure is used to evaluate the dissimilarity between two adjacent windows that slide along the audio stream to produce a distance curve; then, some heuristic thresholds are used to judge whether the locations of peaks are acoustic changes.…”

“…Existing audio segmentation approaches generally fall into two categories, namely, distance-based segmentation [9], [10], [11], [12], [13] and model-decoding-based segmentation [4], [11]. In distance-based segmentation, a distance measure of two audio segments is defined first, and then an acoustic change detection strategy is designed based on the distance Manuscript received January 10, 2009; revised May 06, 2009.…”

“…More precisely, it is assumed that the feature vectors in each of the two audio segments arise from some probability distribution (e.g., the multivariate Gaussian distribution); then, the distance between the two segments is represented by the dissimilarity between the two distributions. Several distance measures have been proposed, e.g., the Kullback-Leibler distance (KL or KL2) [10], the Generalized Likelihood Ratio (GLR) [9], [14], ∆BIC [11], [15], [16], [13], [17], [18], the Bhattacharyya distance [12], and the XBIC [19]. In addition, some high-level features have been used for audio segmentation; e.g., the spectrum flux and zero-crossing rate (ZCR) [20], [21], and the smoothed zerocrossing rate (SZCR) [22].…”