Modeling Of Singer Voice Characteristics And Its Applications

In recent years,under the double stimulation of both the rapid development of mobile Internet applications and various large-scale reality shows for music,singing APPs flourish,and bring new challenge for music recommendation system.The traditional music recommendation system focuses on recommending favorite songs to the user.In the application scene of singing songs,it is not comprehensive to just recommend the song according to the user's preference,but also need to consider whether the singing ability of the user matches the deductive requirement of the recommended song.However,the music recommendation method does not migrate synchronously with the application scenario.At present,most singing APPs still remain in the status of recommeding songs suitable for users to sing by music classification,heat or user-selected recordings without according to the user singing ability.To this end,this paper describes the portrait of non-professional singers' singing performance as the main research objectives,takes the singer's unaccompanied singing audio signal as reseach object,selects the singer's vocal range and timbre as its characterization,and then builds the singer's voice characteristic model as the portrait of the user's singing ability,and ultimately applies to personalized music recommendation system.The personalized system recommends suitable singing songs by the portrait of the user's singing ability.Firstly,this paper presents a method to extract the singer's vocal range based on Wilson confidence statistics.Specifically,compare the singer's MIDI pitch sequence with the song's stadard MIDI pitch model at first,and then evaluate the completion of the non-professional singer at each basic pitch level based on the Wilson confidence interval.Furthermore,define the singer's basic singing ability and finally determine the singing vocal range of the singer.The experimental results show that the singing range extracted by this method has a high degree of consistency with that of the experimental object.Secondly,this paper explores the representations of the voice timbre and similarity measures.In detail,embed the high-dimensional and time-varying characteristic of voice spectrum into three-dimensional timbre embedding space with the help of powerful dimensionality reduction and feature learning of the deep learning.So the measurability of timbre similarity can be achieved in the three-dimensional timbre embedding space.Experiments show that the classification accuracy achieves 73.12%in timbre embedding space including 15 singers,which can effectively ensure the accuracy of the similarity metric.Finally,this paper builds singer's voice characteristic model with the combination of the singer's vocal range and timbre characterization,and uses it to describe user's singing ability subsequently.At the same time,according to the song's numbered musical notation,accompaniment,original singer's voice timbre and other information,a song reference model is established.By matching the range and timbre between the singer's voice characteristic model and songs,a personalized recommended music list suitable for singing is provided.