Font Size: a A A

Research And Implementation Of Humming Melody Automatic Extraction Technology

Posted on:2018-03-21Degree:MasterType:Thesis
Country:ChinaCandidate:W X XuFull Text:PDF
GTID:2348330515960118Subject:Computer technology
Abstract/Summary:PDF Full Text Request
With the rapid development of computer automation technology5the automatic singing transcription is to replace the artificial notation is becoming a hot spot of present research of computer music field.Not only liberated the musician notation work,but also carry out the immediate notation.It can also be used in the areas such as music retrieval in real life,and it has certain practical meaning.In this paper,we make a discussion according to Automatic singing transcription which is explored by us:1.Put forward a kind of standard which is Based on subjective hearing people sing the notes of manual annotation.Being able to make a hum signal as an auditory composition sequence,then base on it to make a rule that the auditory component recognition of the notes block sequence.And according this rule to make a ATN network construction.To realize the Note segmentation and pitch estimation automatically,in order to obtain the notes block sequence which is continuous.For the problems of the different component in each note block and the base band signal frame of Notes block.We established a Model to estimation the pitch of each blocks.In the humming signal manually marked by the musician,if one of the humming signal areas contains2.Or more adjacent note block which have the same pitch,the boundary cutting technique of the adjacent sound block in the segment signal area is mainly used according to the energy variation characteristics of the signal frame in this region.The method used in the research group was to manually observe and collect the energy change of the signal frame in the note boundary which was marked by the musician as the energy change threshold of the note boundary.But this method may regularly make the design of the algorithm loss the accuracy because of the manual analysis of the data.In the long RS region(a humming signal region containing 6 or more frames and the auditory component is a pitch stabilization region),specifically,the segmentation procedures may lead to some situation include missing cut or wrong cut,if the area is determined to be a continuous sound block containing the same pitch.To solve this problem,this paper proposes a model based on the signal frame energy jump,through analyzing and comparing the data manually marked by the musician.Besides,a certain number of hand-marked songs are used as a training set.By adding a layer of "filter" program,we make the second cut in the note areas which have already cut and satisfied the required range,which can reduce the probability of missing cut or wrong cut,improving the accuracy of cut.This algorithm improve the 38 song's notes to estimate the average accuracy of test data from the original 45.39%to 51.72%.According to Emilio Molina assessment framework,The result is almost the method based on the HMM Ryynanen.But our approach does not use YIN on the fundamental frequency estimation algorithms,which havs the advantage of to real-time estimate note pitch.3.In hum fragments tonal level,through the analysis of the interval of adjacent to sing the notes,estimate the tonality of the matching hum fragments.Notes according to tonal step pitch to achieve expected to hum melodies estimate.Finally,establish new humming melody estimating evaluation method.New evaluation method involves describing humming melody line accuracy calculation interval time and reflect the ability of system to automatic correction calibration interval time.4.We put forward a kind of model to Evaluating our system performance,By comparing the analysis results,to analyze the reliability and accuracy of the segmentation algorithm.After these steps,we have implemented the Singing Tracker system.To some extent,it can replace the work which is by the notation experts,it has certain using value of the actual notation and computer music research.
Keywords/Search Tags:Auditory component, Energy jump, Melody recognition
PDF Full Text Request
Related items