Automatic Transcription of Piano Music

The goal of this project is to automate the music transcription process from a musical recording. In this project, we will consider piano recordings consisting of only monophonic notes i.e. only a single note will be played at a time. There are many algorithms to generate musical sheets from an audio such as autocorrelation, CQT transform and cepstral based. This project will focus on YIN, which is a modified autocorrelation algorithm, and cepstral algorithm.

Running the Code

Rscript pitch-detection.R sound-file-path

It accepts one command line argument which is the path of sound file whose notation is to be identified.

Results:

1. DG

   Actual notes:

    D G
   

   Cepstrum method:

    D G
   

   Accuracy: 100%

2. D chord Actual notes:

    D F# A D
   

   Cepstrum method:

    D F# A D#
   

   Accuracy: 75%

3. Ode to Joy

   Actual notes:

    E4 E4 F4 G4 G4 F4 E4 D4 C4 C4 D4 E4 E4 D4 
    E4 E4 F4 G4 G4 F4 E4 D4 C4 C4 D4 E4 D4 C4
   

   Cepstrum method:

    E4 E4 F4 G4 G4 F4 E4 D4 C4 C4 D4 E4 E4 D4
    E4    F4 E4 G4 G4    E4 D4 C4 C4 D4 E4 D4 C4
   

   Accuracy: 89% (2 missing notes and one additional note is detected)

4. Mary had a little lamb Actual notes:

    E4 D4 C4 D4 E4 E4 E4 D4 D4 D4 E4 E4 E4
   

   Cepstrum method:

    E4 D4 C4 D4 E4 E4 E4 D4 D4 D4 E4 E4 E4 E3
   

   Accuracy: 96% one extra low E (E3) detected