2016/6/41 Recent Improvement Over QBSH and AFP J.-S. Roger Jang (張智星) Multimedia Information Retrieval (MIR) Lab CSIE Dept, National Taiwan Univ.
-2- Outline zImprovement over QBSH (query by singing/huming) yWeights of rests and sorted error vectors yGPU optimization zImprovement over AFP (audio fingerprinting) yRe-ranking via learning to rank zNew tasks in MIREX yAFP ySinging/humming transcription
-3- Basic Method in QBSH zLinear scaling (LS)
-4- How To Deal with Rests in LS? zTo deal with rests (zero pitch) yReplace the rest with previous non-zero pitch zThis could go wrong for unstable trailing pitch due to yWrong endpoints yGlissando yVibrato
-5- Weights for Zero-pitch zAssign different weights for rests in the database and queries
-6- Example of Zero-pitch Weights
-7- Sorted Error Vector zCompute distance based on a partial set of growing errors, to deal with the problems of yDouble/half pitch error yMoving average
-8- QBSH Corpus Corpus 1Corpus 2Corpus 3 NameIndian (Indian)MIR-QBSHCHT (Chinese) Database formatwavemidi Database size Query set size 269 (chopped from 35 wave files) Query set format Pitch vector Query length10 sec8 sec sec
-9- Search Zero-pitch Weights zOptimize the weights for MIR-QBSH The best accuracy occurs at w1=0 and w2=2.
-10- Performance Evaluation Both SEV & zero-pitch weights improve top-10 accuracy!
-11- Efficiency Boost via GPU zWe can cut down QBSH response time via GPU (from 1.8 sec to 1.2 sec) by careful arrangement of blocks/threads and memory usage in order to ySpeed up memory access yAvoid bank conflicts zDetails of speedup via GPUsDetails of speedup via GPUs zDemo: toyshttp://mirlab.org/demo/miracletoys
-12- Improvement on AFP zRe-ranking of AFP by learning to rankRe-ranking of AFP by learning to rank zDemo:
-13- New Tasks in MIREX zWe’d like to propose two new tasks for MIREX yAudio fingerprintingAudio fingerprinting ySinging/humming transcriptionSinging/humming transcription
-14- Thank you for your attention! Questions & comments?