Study On The Separation Algorithm Of Singing Voice And Accompaniment In Single Channel Music Signal

With the continuous increasing of the number of music,the demand for music signal processing technologies such as music annotation,retrieval,and identification has also increased dramatically under the background of the information age.separating the music signals into background music and song and can be used in the instrument/singing recognition,audio post-processing,repeated background noise removal applications,which has important significance in the fields of singer and song retrieval,speech recognition in music background,and melody extraction.a good vocal accompaniment separation system can bring convenience and provide good performance guarantee for post-processing,so it has important research value.However,vocal accompaniment separation technology is different from noise reduction separation technology in audio signal,and the mutual interference between them brings many challenges to academic researchThis thesis mainly studies the separation of singing and accompaniment in music signal,including the following aspects(1)For the difficulty of separation of accompaniment from mono music,a kind of music accompaniment separation method based on two-dimensional Fourier transform(2DFT)was proposed,first of all,mono music was transformed by two-dimensional Fourier transform into a two-dimensional sonogram.Secondly,the position of periodic peak energy was determined by image filtering,and then masking matrix was constructed by rectangular window to extract the constituent of the music accompaniment.Finally,the accompaniment of the time-domain signal was restored by means of inverse transformation.The simulation experiments show that the method in this section has an advantage over other separation algorithm.The separation index SIR can be improved by about 0.5-4 d B,and SAR by more than 15 d B.(2)In order to solve the problem of separation between singing and accompaniment in the musical signals,a music separation method based on repeating structural model and Sound Source Separation is proposed.Firstly,the harmonic source and the percussive source of the music are separated by iterative method.Then the beat spectrum was introduced to analysis the energy spectrum matrix of the different sound sources.And the repeated periodic components of the harmonic source are retained to obtain background music while the repeated periodic components of the percussive source are removed to separate the singing voice.For the MIR-1K database,the separation experimental results of 1000 music fragments show that the proposed method has an advantage over other existing music separation methods.(3)For the difficulty of separating the accompaniment and singing from each other in musical signals and difficulty of efficient use of phase information during separation,a method of music separation based on the bidirectional neural network of discriminative training in complex domain is proposed.First,considering the temporal correlation of music signals,a deep stacked bidirectional neural network based on the traditional LSTM network is proposed to preserve the temporal information of music signals.Secondly,on the basis of time-frequency masking and spectrum mapping,a signal approximation algorithm for discriminative training in complex domain is proposed as the objective function of neural network,which makes full use of the phase information in music signal for separation.Finally,the corresponding time domain signal is obtained by the inverse Fourier transform.Experiments show that the new objective function can significantly improve the separation performance of the neural network.Compared with the existing music separation methods,the method presented in this section has excellent performance in both accompaniment and singing separation.