This paper provides a rough description of zero resource Query-by-Example retrieving system for the MediaEval 2013 spoken web search task. The proposed solution rstly implements the voice activity detection (VAD) utilizing variance of acceleration MFCC (VAMFCC) rule-based approach. A PCA-based segmentation, K-means clustering and GMM training are then used in order to built the posteriorgrams. Finally, two searching architectures based on posteriorgram matching (SDTW) and GMM modeling (GMM-FST) are evaluated. Results show that none of our systems is able to achieve the positive Actual Term Weighted Value, because of high number of insertions. We suppose that chosen clustering scheme caused generation of too many false alarms. Only provided data were used and no other resources were examined in any system component during the development.
The main purpose of our experiments was to check the
proposed approaches for the language independent audio
query detection and new speech feature analysis
components. The mentioned approach is used in the MediaEval
activity [
Proposed solution for SWS task uses posteriorgram term matching and audio segment GMM modeling. The overall architecture of proposed system is depicted on Fig.1.
At rst, a training phase is carried out using available development utterances. The VAMFCC-based silence detector performs the initial discrimination of silent parts in audio stream. The block of feature extraction is implemented after VAD utilizing 13 MFCCs. The phase of segmentation and clustering creates the audio segment units (ASU). ASU is represented as a small audio part (phoneme for example) with some spectral and temporal characteristics, di erent for each ASU. Then the training of acoustic models is performed using these ASU, where each ASU represents one class. Labels for these classes (ASUs) are assigned according to the number of GMM. Only the process of voice activity detection and feature extraction is then performed in preprocessing stage within the retrieving phase. Each utterance and 2.1
In order to identify and to distinguish the speech segments in the i-th utterance, PCA (Principal Component Analysis) was applied as follows. Each 13-dimensional MFCC vector xj was reshaped to matrix Xj with row dimension nr, where j 2 h1; nii is the number of vectors in i-th recording. In the next step, the covariance matrix C1 was computed from the rst matrix X1 and its eigenvectors and eigenvalues were computed. The eigenvalue spectrum 1 = f 1j gjn=r 1 was used to determine the signi cance ( 1max ) of the dominant eigenvalue of C1 as ( 1max ) = Pjn1=rm1ax1j , where 1j are the eigenvalues of C1. Then the matrix X1 was spliced together with X2 and the covariance matrix C12 and ( 12max ) were computed again. If ( 12max ) compared to ( 1max ) changed signi cantly, a new speech segment was created and PCA started from the current frame. In the other way, if ( 12max ) did not change signi cantly, the current matrix X12 was spliced together with X3 and the process was repeated automatically until a new segment was indicated. The created segments corresponded to ASUs.
In the next phase, the segments with similar acoustic and statistical properties were grouped together into several speech clusters using k-means clustering with k = 50 clusters and squared Euclidean distance metrics. As the input data for clustering the means of the segments were used. Each mean vector obtained an index (label) of the speci c cluster. This label was assigned to the original feature vectors corresponding to the speci c mean vector. 2.2
2.2.1
A retrieving process uses Weighted Finite State
Transducers (WFST) that allow us to nd the most probable path
(state sequence) in search network [
A simpli ed SDTW searching algorithm was utilized in
our system, similar to that used in [
P1n warpDist ;where n+1 n is the number of steps in local alignment, N is the length of utterance and M the length of query, P1n warpDist represents a summation of components in each warping path, where components are computed from Bhattacharyya distance matrix. GMM-FST
SDTW
The o cial results for SWS task are listed in Tab. 1. Two metrics were used to asses the overall performance of GMM-FST and SDTW on dev and eval queries: the actual AT W V , normalized Cnxe and minimal cross-entropy Cnmxien. The score normalization for both systems was performed only on development data. A minimum-cost alignment (MCA) for each segment was used as detection score at rst level of search in case of SDTW. Final detection of retrieved query was then carried out utilizing score parameter de ned in (2.2.2). A threshold for this parameter was set to 0:0819. A decision threshold for score parameter was set to 2:8 in case of GMM-FST based system, while = 3 (2.2.1). Both systems produced a huge amount of false alarms (FA) during the evaluation. Regarding the evaluation results, the GMM-FST system is more appropriate solution for SWS task, because of its lower tendency to detect spurious terms.
All the experiments were mainly done using 2x IBM System x3650 servers, 2x Intelr Xeonr QuadCore E5530 CPU @ 2.4 GHz Hyper-threading enabled (16 threads), 28 GB RAM, 1TB SAS HDD (RAID5), running Debian OS.
The Speed Factors (the ratio of the total time employed in searchingfindexingg the set of queries infandg the set of audio documents to the productfsumg of their total durations) and Peak Memory Usage during SearchingfIndexingg tasks are presented in Tab. 2. The Performance Load equals 0:9 SSF P M US + 0:1 ISF P M UI is derived from them.
In the future, an improved clustering and segmentation algorithm will be investigated in order to decrease the overlapping between individual ASUs. A minimal length of warping path algorithm will be integrated in SDTW approach, too. 4.
This research was supported by the ERDF funded projects ITMS-26220220155 (50%) & ITMS-26220220182 (50%).