The Research Of The Piano Multi-note Recognition Based On The HMM

The piano music note recognition is the process that automatically converted aaudio music files to a digital music files (MIDI), which plays an important role inassisted teaching piano, music automatically records and so on. The main research ofthis issues is to identified the piano notes, that according to the recorded piano playingaudio files, with the help of the specific algorithm, determined which notes areincluded in the every time period. Depending on the application requirements and therestrictions of note recognition algorithm itself, the piano notes recognition tasks canbe divided into two parts: real-time and robust single note recognition; non-real-timemulti-note recognition. For the single note recognition, this thesis focuses on thereal-time and robustness aspects. So, we propose to use the local energy onsetdetection algorithm, Mutil-sample Dictionary, and its post-processing algorithms toimprove the results; For the multi-notes, the single time domain method cannotworked well, so this paper turned to the method of pattern recognition. Namely, usingthe multi-note as the basic unit of the model, to establish the hidden Markov modelsand then recompile the related modules of the HTK Tools to build the multi-noterecognition system.This thesis mainly concluded the following aspects:1) Extracting the feature of the multi-notes audio file. On the basis of analyzingthe parameters of the audio files, we used the feature of the multi-notes audio tooptimize existing feature extraction module in the HTK Tools;2) Determine the onset location of the audio notes. We firstly analysis thecommon onset detection, and then according to the different mission requirementsbetween the single note and the multi-notes, we design or rewrite the correspondingonset detection algorithm;3) Based on the principle of the HMM speech recognition, we find thedifferences and similarities between the voice recognition and the multi-notesrecognition, then we established the multi-notes HMM acoustic model and the modelof two multi-note; 4) According to the MIDI files and the corresponding audio files piano, we builtthe training, testing waveform data and the corresponding annotation files. On thisbasis, we used the recompiled HTK tools to initialization and training each of themulti-note’s HMM, and then build a complete multi-note recognition system. Finally,comparing the results of the experimental and the corresponding annotation files, wecan evaluate the performance of multi-note system.This thesis successfully achieved the real-time robust single notes recognition,modeling the HMM for multi-note and model recognition. For the recognition of thesingle note, compared with the linear model approach, the correct rate increased by3%to98%, and the average recognition frames increased four and improving therobustness aspects of two times; For the recognition of the multi-notes, with thecompared to the common methods, the recognition rate of the method that based onthe HMM increased by nearly5percent and it improved the practical possibilities ofpiano music transcription.