Research And Implementation Of Key Techniques Of Piano Training Software

With the improvement of living standards,people’s pursuit of art has a higher standard,and the trend of piano learning has been set off for a while.For Piano Learners,it is impossible to monitor their learning achievements in real-time when they are practicing the repertoire.Therefore,with the rapid development of computer technology,a large number of piano training software has emerged as the Times require,The key technology is the identification of the music,including the length of each note and its pitch.At present,in the research field of music recognition,the recognition of note length,that is,the end point detection of musical notes,is mainly based on the threshold method,which is greatly affected by human factors.The recognition accuracy is low,and is susceptible to environmental noise.Although the pitch recognition has made breakthroughs in single-tone recognition,there is still no effective method for chord recognition.This thesis takes the audio generated during piano playing as the research object,and uses an improved Bi-directional long-term short-term memory network and a convolutional neural network to study musical tone endpoint detection and pitch recognition.The specific work is as follows:(1)On the basis of analyzing the characteristics of piano music and the pros and cons of the key technologies of the current piano training software,the technical ideas of this article have been established.The experimental environment,evaluation standards and The overall algorithm flow have been designed;(2)Aiming at the shortcomings of the traditional classic dual-threshold algorithm that relies too much on the threshold in the endpoint detection of music,an endpoint detection algorithm based on a Bi-directional long and short-term memory network and short-term energy difference is proposed,which is characterized by the short-term energy difference between two adjacent frames.The short-term energy difference of 11 consecutive frames is sent as an input,and whether the end point at the 6th frame is sent as a label.Take these data to the Bi-directional long-term memory network for training to complete the endpoint detection of the musical tone;(3)In pitch recognition,aiming at the shortage and incompleteness of chord data sets,firstly,the construction of data sets is completed by using the characteristics that MIDI files are easy to write.Then,the constant Q transform is used to preprocess the data to ensure the accuracy of frequency recognition.Then,the spectrogram of each note after constant Q change is drawn and input into convolutional neural network to complete training.The constructed convolution neural network and its training,for each fundamental frequency,use the way of binary classification to detect whether the chord contains the sound,so as to ensure the full coverage of 88 tones of the piano.Based on this improved algorithm,the method of energy normalization is adopted to avoid the decline of recognition accuracy in case of frequency doubling and half frequency;(4)Experiments with multiple extreme audio data are used to compare the dual-threshold method,the time-domain analysis method based on the short-term energy difference,and the algorithm implemented in this thesis.The endpoint detection algorithm based on the Bi-directional long-short-term memory network and the short-term energy difference can be used to identify the effect is better,and the generalization ability is stronger.In pitch recognition,experiments on the NSynth Dataset and self built dataset can achieve an average accuracy of 91.68%,which proves the effectiveness of the algorithm.