This paper describes two query-by-example systems developed by Speech Lab, Queens College (CUNY). Our systems aimed to respond with quick search results from the selected reference les. Three phonetic recognizers (Czech, Hungarian and Russian) were utilized to get phoneme sequences of both query and reference speech les. Each query sequence were compared with all the reference sequences using both global and local aligners. In the rst system, we predicted the most probable reference les based on the sequence alignment results; In the second system, we pruned out the subsequences from the reference sequences that yielded best local symbolic alignments, then 39-dimension MFCC features were extracted for both query and the subsequences. Both the two systems employed an optimized DTW, and obtained Cnxe of 0.9989 and 1.0674 on the test data respectively.
[1]
S. Sathiya Keerthi,et al.
Improvements to Platt's SMO Algorithm for SVM Classifier Design
,
2001,
Neural Computation.
[2]
Florian Metze,et al.
Query by Example Search on Speech at Mediaeval 2015
,
2014,
MediaEval.
[3]
Eamonn J. Keogh,et al.
iSAX: indexing and mining terabyte sized time series
,
2008,
KDD.
[4]
Eamonn J. Keogh,et al.
Searching and Mining Trillions of Time Series Subsequences under Dynamic Time Warping
,
2012,
KDD.
[5]
M. Mostafizur Rahman,et al.
Cluster Based Under-Sampling for Unbalanced Cardiovascular Data
,
2013
.
[6]
Björn W. Schuller,et al.
Recent developments in openSMILE, the munich open-source multimedia feature extractor
,
2013,
ACM Multimedia.