Research And Implementation Of Android Based Music Recognition And MIDI Output

As all know, people can talk with iPhone via Siri. It shows speech recognition technology has made big progress. As a related technology, music recognition has also achieved some amazing applications, the latest example is WeiXin, developed by Tencent Co., it can track a music which is on playing around, and recognize song name, then shows the lyrics synchronously. Technology has always made differences to our lifes.This thesis aims at developing such a APP (based on Android for now) for music fans:it records the music played by some musical instrument, identify the musical notes, and tempos, write the music information into a MID file, and user can choose a musical instrument for the MID file in advance. The interesting part is, user can play with one kind of musical instrument, but the output MID could be another musical instrument.This is not a big project but it’s interdisciplinary. It requires some music theory background as well as computer knowledge. From technical perspective, this APP requirement needs supports at least on2key points:One is Endpoint detection, the other is Pitch recognition. In this thesis, the author begins with study on current audio recognition technology, such as short-term energy for endpoint detection, ACF for pitch recognition, FFT and wavelet transform for frequency detection, and study some classifier popular in speech recognition, e.g. ANN, HMM. Based on all these and some algorithm experiments has been carried out on MATLAB, this thesis proposes2improvements in endpoint detection:First,"Detail" signal reconstructed by wavelet transform is good for short-term energy detection, as it’s relative smooth in silent period, even with "noisy", it’s smooth "noisy" from short-term energy perspective. This is good for endpoint detection.Second, an improved energy inverse signal detection method has been proposed. It cuts4-state of traditional "double threshold energy endpoint detection method", to2-state, with focus on capturing the inverse signals on energy, and the same, with dynamic threshold. As for pitch recognition, based on a truth that it is easier to get the frequency from musical instrument playing than human being’s voice. And experiments show AMDF(Average Magnitude Difference Function) can do good job on pitch recognition, with both Piano playing and Harmonica playing, so AMDF will be used in the APP.Finally a easy featured demo APP has been designed and implemented on Android device according to this thesis, user can input a MID file without a MIDI keyboard on an Android device. As none of such APP is found in today’s Android market, this demo APP is planned to be shared with all music fans after some improvement, for free.