Research On Query By Humming Based On Deep Learning

As the basic core task in the field of audio retrieval research,Query By Humming is the key task of current popular audio research.For the construction of the Query By Humming system,there are some essential difficulties that need to be solved,that is the instability of the humming feature.Therefore,this thesis starts from the singer's humming feature instability,and uses the current deeping research learning to dig out the deeper nonlinear features of the melody feature itself,thus improving the performance of the QBH.The main research contents and achievemenst of this paper are as follows:1.Established a feature learning networks based on triplet loss learningThe melody is segmented first,the error caused by the inconsistent humming rate is weakened,and the melody has no labeling problem after segmentation.The learning strategy of the ternary group is utilized,and the advantages and disadvantages of the original pitch input and statistical feature input are explored.The similarity/dissimilarity between the triples are used as the label of the network learning to make the network learn the melody deeper.In addition,the local feature enhancement properties of the pitch sequence itself are analyzed.We use the convolutional neural network and the original.The network compares and finally tests the validity and feasibility of triplet network learning.2.Research two Query By Humming methods by deep hashing networksOn the baseline network framework based on triplet learning networks,the idea of deep hashing is researched by this paper.The network feature learning and hash coding are combined,and two algorithm frameworks for learning hash code are proposed.The first is to learn the melody features of the triple convolutional neural network through a Sparse Autoencoder,referred to as the SAH hash algorithm.Constraints of sparsity and quantization error are added to the loss function to obtain a more accurate binary hash code;the second is based on an end-to-end hash learning strategy.Introducing a hash layer in the triplet-based convolutional neural network,where in the hash layer includes a fully connected layer,a slice layer,an active layer,and a merge layer,and the feature values can be mapped into a continuous code for generating independent ha The Greek function,and the quantization error generated when the value is continuously converted into a binary hash code is added to the optimization objective function,and finally a hash code with stronger expression ability is obtained.Finally,the experiments of the two algorithms are better than the traditional Query By Humming based on the LSH system.