Multipile Fundamental Frequency Estimation Of Piano Music Based On Rtfi

Music is an important carrier of human culture history, always playing an essential part in human’s life. At the same time, multi-pitch estimation also plays an important role in the area of music analysis. It can essentially used for the detection of musically relevant features such as melody and harmony. It is applied widely in music separation, automatic music transcription, music retrieval etc.Musical signal has strong spectral levels and is time-variable most of the time, which makes it inherently different with other audio signals. When playing multi-tone instruments, like piano, melody and accompaniment vary differently. So notes with different frequencies overlap together, this makes it hard to detect original pitch. What’s more, in case of notes with lower fundamental frequency and disappearing fundamental frequency also cause errors in multi-pitch detection. At present, schemes to resolving such problems in time and frequency domain have shortcomings more or less. In spite of this, relevant scholars have been researched on multi-pitch estimation from different angles, and provided many ideas to us, method like use of human hearing characteristics and identifying lower frequency notes with high frequency resolutions.Based on such excellent research results describe above, the thesis is aim to detect multi-pitch in real piano music. The main work of this paper will be shown as follows:1. The thesis first analyzes some time-frequency technologies like STFT and Wavelet etc, and finally choose constant-Q transform based on Resonator Time-Frequency Image (RTFI). The corresponding multi-resolution fast implementation has also been developed.2. To segment notes in a piece of music, we should first do onset detection. It is piano music that the thesis worked on. And the piano music notes are regarded as hard notes. So the paper researched on the endpoint detection functions and finally selected the function based on spectral energy difference function. And a moving average filter was applied to complete peak-picking procedure.3. Design filter bank and separate it into different octave frequency bands. When came to estimate fundamental frequency, the paper introduced and analyzed two kinds of algorithm, iterative and non-iterative estimation algorithm. Finally considering harmonic structure and relationship between harmonic partials, the non-iterative algorithm which based on RTFI was selected. It firstly took a preliminary estimation base on the implementation of endpoint detection. Later, it screened to get the candidates for fundamental frequencies base on spectral irregularity and knowledge of the harmonic structures of the music notes played on commonly used music instruments like piano.The scheme in this paper greatly reduces the computation and improves the efficiency of multi-pitch estimation. Compared with iterative algorithm proposed by Klapuri, the experimental results show clear improvement versus the state of the art methods.