Theresearch Of Robust Humming Feature

Query by humming system,a multimedia retrieval system based on content,as the current research hotspot,there has been humming features instability problems,the main cause of this problem is vocal personality,humming range of different people,different humming rhythms,individual pitch inaccuracies and other issues,will lead to humming features instability.In this paper,against humming feature instability problems inquery by humming system,carry out the following research,designed to extract more robust humming features:1,improved algorithm for extracting humming featureFor humming segment and after pitch extraction characteristics,in order to improve the matching degree of humming segment and MIDI,improved the humming segment feature extraction algorithm.Through analyzing the frequency distribution of humming,aligning of humming with MIDI,normlizing the pitches;combined music theory,note segmentation operations performed;according to individualvoice range for semi-note conversion,with people voice frequency instead of the original reference frequency 440Hz,making hummng note and midi note to reach a better match.Proved the effectiveness of the algorithm by experiments,to lay a good foundation for the feature matching.2,proposed a humming feature extraction algorithm based on local statisticThrough the humming sequence of notes in the horizontal and vertical distribution range of local statistical,obtaining the histogram statistical features of the humming melody.This algorithm in the longitudinal progressing projection statistical distribution of notes range interval;in horizontal progressing histogram statistical distribution of notes time patterns and,ultimately,obtaining the joint histogram characteristics of longitudinal and transverse,and join the mean,range,and variance characteristics.Finally,the distribution of the notes is described by the combination of the 4 segments of continuous sub sequence and the whole segment.This algorithm is different from the traditional representation to pitch or note features directly as humming feature,but the note is converted to the form of statistical features to ensure a relatively stable humming features,humming for different users in the different performance of speed,range,rhythm,etc.have a good fault tolerance.Finally,the validity of the above algorithms is tested by experiments,the experimental data include the 5000 MIDI,104 humming query,and local sensitive hash algorithm(Local Sensitive Hash)as the similar feature matching algorithm,TOP1 accuracy rate 86%,TOP5 accuracy was 92%,with the original humming recognition system results in the note as a features were compared,the results of the experimental results are better than the original.