The Study Of Melody Extraction Based On Multi-Pitch Extraction

Melody is the fundamental frequency of human voice. Melody extraction is defined as transcription melody of a piece of music to text. Melody extraction is valuable in many music research fields. This paper studied a melody extraction method based on multi-pitch extraction. The study is divided into two parts:(1). Studied a variety of multi-pitch extraction algorithms based on sub-harmonic summation.In the part of multi-pitch extraction, HPSS is used to improve the voice fundamental frequency energy scale. Then we analysis the disadvantage of STFT and introduce MRFFT, which is up for the lack of STFT. Finally, we introduce sub-harmonic summation and make a series of improvements to it. Experiments show that our method is effective.(2). Studied the melody judgment algorithms based on Viterbi decode.In the part of melody judgment, the melody generation process is treated as Markov process. One frequency bin stand for a state. The result of multi-pitch extraction stand for the probability of a T-F unit to be chosen as predominant frequency. Generating state-transition matrix utilize the knowledge that melody cannot change very quickly. Finally, we use Viterbi decode to get melody. Furthermore, we introduce trend-estimation. It use vibrato, tremolo and harmonic theory to get the voice fundamental frequency range. Trend-estimation can enhance melody judgment robustness and performance by decrease the search space.