Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 57 (2015) 876 – 884

3rd International Conference on Recent Trends in Computing 2015 (ICRTC-2015)

Musical Notes Identification using Digital Signal Processing

Jay K. Patela, E.S.Gopia
a
National Institute of Technology, Trichy-620015, India.

Abstract

Songs play a vital role in our day to day life. A song contains basically two things, vocal and
background music. Where the characteristics of the voice depend on the singer and in case of background
music, it involves mixture of different musical instruments like piano, guitar, drum, etc. To extract the
characteristic of a song becomes more important for various objectives like learning, teaching, composing. This
project takes song as an input, extracts the features and detects and identifies the notes, each with a duration.
First the song is recorded and digital signal processing algorithms used to identify the characteristics. The
experiment is done with the several piano songs where the notes are already known, and identified notes are
compared with original notes until the detection rate goes higher. And then the experiment is done with piano
songs with unknown notes with the proposed algorithm.
© 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
© 2013 T he Authors. Published by Elsevier B.V.
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Selection and/or
Peer-review peer-reviewof
under responsibility under responsibility
organizing committeeofofthe organizers
the 3rd of the
International 2013 International
Conference Conference
on Recent Trends on Computational
in Computing 2015
Science
(ICRTC-2015)

Keywords: Time-Frequency Analysis; Musical Note; Sampling Frequency; Recording; Extraction;

1. Introduction

The ability to derive the relevant musical information from a live or recorded performance is relatively
easy for a trained listener, but highly non-trivial for a learner and computer. For a number of practical
applications, it would be desirable to obtain this information in a quick, error-free, automated fashion. This
thesis discusses the design of a software system that accepts as input a digitized waveform representing an
acoustical music signal, and that attempts to derive the notes from the signal so that a musical score could be
produced. This signal processing algorithm involved include event detection, or precisely where within the
signal the various notes actually begin and end, and pitch extraction , or the identification of the pitches being
played in each interval. The event detection is carried out using the time domain analysis of the signal, where
the problem arise with different speed. The pitch detection is (nothing but frequency identification) is more
complicated because of a situation we call harmonic ambiguity; this occurs when one pitch whose fundamental

1877-0509 © 2015 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of organizing committee of the 3rd International Conference on Recent Trends in Computing 2015 (ICRTC-2015)
doi:10.1016/j.procs.2015.07.499
 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884 877

frequency is an integer multiple of another pitch. The problem is solved by the careful signal processing in both
the time domain signals and frequency domain signals.

The main objective of this project is to create an aid tool for learning for Musicians, Producers,
Composers, DJs, Remixer, Teachers and Music Students. This project can be treated as a box, where you give
any song as input and get the features of the song out. The aim of this project is to propose methods to analyze
and describe a signal, from where the musical parameters can be easily and objectively obtained, in a sensible
manner. A common limitation find in the musical literature is that the way in which such paramete rs are
obtained is intuitively satisfactory but, to our view, not very sound from a signal processing perspective.

2. Literature Survey

2.1 Sound

A sound can be characterized by the following three quantities: (i) Pitch.(ii) Quality.(iii) Loudness.

Pitch is the frequency of a sound as perceived by human ear. A high frequency gives rise to a high
pitch note and a low frequency produces a low pitch note. A pure tone is the sound of only one frequency, such
as that given by a tuning fork or electronic signal generator. The fundamental note has the greatest amplitude
and is heard predominantly because it has a larger intensity. The other frequencies such as 2fo, 3fo, 4fo,... are
called overtones or harmonics and they determine the quality of the sound. Loudness is a physiological
sensation. It depends mainly on sound pressure but also on the spectrum of the harmonics and the physical
duration.

2.2 Musical Notes

Human can hear signal frequency ranging from 20-20 kHz. From this wide range some part is
associated with piano. Different pianos are having different ranges. Each tone of piano is having one particular
fundamental frequency and represented by a note like C, D, ...etc. as shown in fig 1 .The later C is 12 half steps
away the previous one and having double the fundamental frequency. Hence this portion (from one C
immediate next C) is called one octave. Different octaves are differentiated by C1,C2, etc.

Fig 1. An octave of a piano

2.2.1 Equation For the Frequency Table

The basic formula for the frequency of the notes of the equal tempered scale is given by fn = f0 * (a) n
where f0= the frequency of one fixed note which must be defined. A common choice is setting the A above
middle C(A 4) at f0 =440 Hz, n = number of half steps away from the fixed note you are, fn= the frequency of
the notes n half steps away, a = (2)1/12

2.2.2 Frequencies for Equal Tempered Scale at A 4 = 440 Hz

878 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884

Table 1. Notation to Frequency Mapping [Middle C is C4]

n Note Fundamental Frequency(Hz)

-4 F3 174.61
-3 F 3# 185
-2 G3 196
-1 G3# 207.65
0 A4 220
1 A4# 233.08
2 B4 246.94
3 C4 261.63
4 C 4# 277.18
5 D4 293.66
6 D4# 311.13
7 E4 329.63

2.3 Digital Signal Processing for music

2.3.1 Sampling

When a sound wave is created by your voice (or a musical instrument), it's an analog wave of
changing air pressure. However, in order for a computer to store a sound wave, it needs to record discrete
values at discrete time intervals. The process of recording discrete time values is called sampling, and the
process of recording discrete pressures is called quantizing. Recording studios use a standard sampling
frequency of 48 kHz, while CDs use the rate of 44.1 kHz. Signals should be sampled at twice the highest
frequency present in the signal. Humans can hear frequencies from approximately 20-20,000 Hz, which
explains why common sampling frequencies are in the 40 kHz range.

2.3.2 Frequency and Fourier Transforms

A Fourier transform provides the means to break up a complicated signal, like a musical tone, into its
constituent sinusoids. This method involves many integrals and a continuous signal. We want to perform a
Fourier transform on a sampled (rather than continuous) signal, so we have to use
the Discrete Fourier Transform instead.

Fig 2. FFT of a Musical Signal

From the fig 2 we can see that the index corresponding to the maximum amplitude represents the most
significant frequency component, that can be found using the formula
 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884 879

࢏
f = *fs (1)
ࢀ

Where i = index at which maximum amplitude exists , T = Total samples in the fft at a time.

2.3.3 Padding with Zeroes

Padding with zeroes, though a standard practice and useful in many applications, does not appear to
significantly improve our data in this application. Though we have twice the frequency resolution, it's not
yielding any better data. Quick tests padding with many more zeroes (1 part sample, 9 parts zeroes) show that
though the peaks get rounder due to better frequency resolution, the discrepancy bet ween the highest point on
the original sample and the highest point on the padded sample is 1 Hz at most, which means that it is probably
not worth the effort, even at low frequencies.

3. Methodology

Detection Identification

Averaging
Padding Zeroes
Thresholding
Input Piano Song DFT Notes
Width Selection
Assignment
Finding Instant

Fig 3. Flow-Chart

The experiments are done on different songs, some downloaded and some recorded from the virtual
piano. Here each note follows a kind of similar pattern as shown in the fig 3.

Fig 4. Piano note time domain analysis

The moment we press one note to the immediate other note, the amplitude is initially high enough and
decreases with time. If we are able to detect the duration of each note from the time domain characteristics, we
can detect and identify the frequency.

3.1 Detection
880 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884

3.1.1 Averaging
As there is large number of sample for a song and many fluctuations are also present, first step is
averaging, where for every 100 samples average is found and the value is assigned to first sample, again for
next 100 sample the average value is assigned to 2nd sample. This will not only reduce the number of samples
but also remove the fluctuations present.

Fig 5. Effect of Averaging

When the decay in the signal is slow, the averaged signal is more densed. But in case of fast decaying
the averaged signal represents the envelope of the original signal.

3.1.2 Thresholding

Constant Thresholding : After averaging we need to detect the peaks from the averaged signal. As the
name suggests, in constant thresholding one optimum value is decided for which we are able to get maximum
number of peaks.

Fig 6. Signal after thresholding

Adaptive Thresholding : There are the possibilities when we take some constant threshold value, for
some notes it may be higher than the maximum value of the note and for some notes it is low enough such that
two peaks of the note get merged and only one peak (one note) could be detected. To overcome this problem
the concept of adaptive thresholding came into picture.
 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884 881

Fig 7. Values of adaptive thresholding

Even in adaptive thresholding one problem is, when there is long silence between two notes, the
threshold value for the silence will be very low that supposed to be discarded.

3.1.3 Width Selection for Finding instant

Once the thresholding is done, our aim is to detect the occurences of the peaks from the signal we get
after thresholding.

Fig 8. Width selection (a) Signal after thresholding (b) Minimum Width

The width of the window contains all zeroes is selected on the basis of worst case scenario. For our
song it is found as 57, so the width was chosen as 50 for safety. The worst case depends on the speed for a
given sampling frequency. For slow songs it is more and for faster one it is less. So irrespective of the song, the
௙௦
minimum width was decided as per following equation, Wmin= where, fs = sampling frequency, A =
ሺ஺‫כ‬ଶ଴ሻ
Total number of samples in the original signal that corresponds to one sample in the averaged signal

3.2 Identification
882 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884

So far we have decided the instants at which the notes are played (key is pressed). The time duration
from the first instant to the second instant is the duration of the first note being played, from second to third
instant, it’s second note being played and so on. Now our aim is to identify these notes based.

3.2.1 Padding Zeroes

For a given note duration, if crop the corresponding instants of the original signal and find the discrete
fourier transform, the results are not close enough as required. So we need to pad the zeroes. We can pad the
zeroes with different length and different part of the cropped signal and find the DFT that will give the different
results. Different part means it can be either only before the cropped version, after it or both the sides with
different length. From these different variations the closest results are obtained in the case where the section of
zeroes with the same length is padded both the sides of the cropped version. So DFT of the resultant signal is
found to determine the notation.

3.2.3 Finding Frequency and Assigning Notation

After padding zeroes, the DFT of the resultant signal is found. Then the corresponding frequency of a
particular note is found using equation (1) and the actual note is assigned using table 1.

4. Simulation Results

The experiment was done with different songs like Happy Birthday, Jingle Bell, Twinkle-Twinkle, etc.
The results of Happy Birthday are shown below.

Happy Birthday
Total Length (L) = 830063 samples; Time (t) = 26 sec; Sampling Frequency (fs) = 32000 Hz;
Sampling Period (Ts) = 31.25 μs; Total Present Notes (P) =25; Total identifies Notes (D) = 25

Fig 9. (a) Results Amplitude v/s Time (samples) (b) Frequency (Hz) v/s Time (sec)
 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884 883

Table 2. Results obtained

Sr. No. Frequency Detected(Hz) Frequency Assigned(Hz) Output Notes time (sec)

1 393.9440624 391.995436 G4 3.0875

2 394.937165 391.995436 G4 3.8375
3 441.6215282 440 A4 3.99375
4 393.7113194 391.995436 G4 5.0875
5 525.7324486 523.2511306 C5 5.9375
6 496.1716113 493.8833013 B4 7.14375
7 393.9087979 391.995436 G4 9.0875
8 395.5372436 391.995436 G4 9.828125
9 441.7132749 440 A4 9.9875
10 393.686712 391.995436 G4 11.09375
11 590.350206 587.3295358 C5 12.00625
12 525.4171797 523.2511306 B4 12.9375
13 393.8059943 391.995436 G4 15.08125
14 394.937165 391.995436 G4 15.84063
15 788.3976356 783.990872 G5 16.17188
16 665.6706126 659.2551138 E5 16.97188
17 662.3520922 659.2551138 E5 17.075
18 525.7559079 523.2511306 C5 17.9375
19 496.0954288 493.8833013 B4 19.14688
20 441.5441926 440 A4 19.94688
21 701.768031 698.4564629 F5 22.07188
22 702.6373901 698.4564629 F5 22.82188
23 662.2568821 659.2551138 E5 23.07188
24 525.6850915 523.2511306 C5 23.9375
25 590.3044482 587.3295358 D5 25.0125

5. Conclusion

In this project, the frequencies of a piano song is detected, corresponding notes are identified with
duration. The method used here for note identification is more optimised than previously used methods.
By varying the parameters like threshold values and width, we can get the desired results with time
duration of each note. Thus project can be treated as an aid tool for learning.
884 Jay K. Patel and E.S. Gopi / Procedia Computer Science 57 (2015) 876 – 884

References

1. Prashanth, T. R., & Venugopalan, R. (2011). Note identification in Carnatic Music from Frequency Spectrum. In
Communications and Signal Processing (ICCSP), 2011 International Conference on (pp. 87-91). IEEE.
2. Kirthika, P., & Chattamvelli, R. (2012,). A review of raga based music classification and music information retrieval
(MIR). In Engineering Education: Innovative Practices and Future Trends (AICERA), 2012 IEEE International
Conference on (pp. 1-5). IEEE.
3. Sridhar, R., & Geetha, T. V. (2009). Raga identification of carnatic music for music inform ation retrieval. International
Journal of recent trends in Engineering, 1(1), 571-574.
4. Flexer, A., & Schnitzer, D. (2010). Effects of album and artist filters in audio similarity computed for very large music
databases. Computer Music Journal, 34(3), 20-28.
5. Huang, P. S., Chen, S. D., Smaragdis, P., & Hasegawa-Johnson, M. (2012). Singing-voice separation from monaural
recordings using robust principal component analysis. In Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE
International Conference on (pp. 57-60). IEEE.
6. Levy, M. A. (2005). Ringermute: An audio data mining toolkit (Doctoral dissertation, University of Nevada Reno).
7. Chordia, P. (2006). Automatic raag classification of pitchtracked performances using pitch-class and pitch-class dyad
distributions. In Proceedings of International Computer Music Conference.
8. Shetty, S., & Achary, K. K. (2009). Raga mining of Indian music by extracting arohana-avarohana pattern. International
Journal of Recent Trends in Engineering, 1(1), 362-366.