Audio Note Extractor

            Songs play a fundamental job in our everyday life. A tune contains fundamentally two things, vocal and ambient sound. Where the attributes of the voice rely upon the artist and if there should be an occurrence of ambient sound, it includes a combination of various instruments like piano, guitar, drum, and so forth To remove the trait of a tune turns out to be more significant for different goals like getting the hang of, educating, making. This undertaking accepts melody as information, removes the highlights, and distinguishes and recognizes the notes, each with a span. First, the melody is recorded and advanced sign handling calculations are used to recognize the qualities. The trial is finished with the few piano songs where the notes are now known, and recognized notes are contrasted and unique notes until the identification rate goes higher. And afterward, the trial is finished with piano songs with obscure notes with the proposed calculation.

Sound

            A sound can be portrayed by the accompanying three amounts: 

                 (a) Pitch. 

                 (b) Quality

                 (c) Loudness.

           Pitch is the frequency of a sound as seen by the human ear. A high frequency leads to a high pitch note and a low frequency delivers a low pitch note. An unadulterated tone is the sound of just a single frequency, for example, that given by a tuning fork or electronic signal generator. The principal note has the best amplitude and is heard prevalently on the grounds that it has a bigger intensity. Other frequencies, for example, 2fo, 3fo, 4fo.... are called overtones or harmonics and they decide the nature of the sound. It relies fundamentally upon sound strain yet in addition on the range of the harmonics and the actual length.

Musical Notes

          People can hear signal frequencies going from 20-20 kHz. From this wide reach, some part is related to the piano. Various pianos are having various reaches. Each tone of the piano is having one specific central frequency and is addressed by a note like C, D, ..., and so forth as displayed in fig 1. The later C is 12 half advances from the past one and has twofold the principal frequency. Thus this piece (from one C prompt next C) is called one octave. Various octaves are separated by C₁, C₂, and so on.

Digital Signal Processing for Music

• Sampling

               At the point when a sound wave is made by your voice (or an instrument), it's an analog wave of changing pneumatic stress. In any case, for a PC to store a sound wave, it needs to record discrete qualities at discrete time stretches. The method involved with recording discrete-time esteems is called sampling, and the most common way of recording discrete pressures is called quantizing. Recording studios utilize a standard sampling frequency of 48 kHz, while CDs utilize the pace of 44.1 kHz. Signs ought to be inspected at double the most noteworthy frequency present in the sign. People can hear frequencies from around 20-20,000 Hz, which clarifies why normal sampling frequencies are in the 40 kHz range. 

• Frequency and Fourier Transforms   

                A Fourier change gives the necessary resources to separate a convoluted sign, similar to a melodic tone, into its constituent sinusoids. This technique includes numerous integrals and a consistent sign. We need to play out a Fourier change on a tested (rather than nonstop) signal, so we have to the Discrete Fourier Transform all things considered. 

Fig. 1. FFT of a Musical Signal

From fig 2 we can see that the record compared to the greatest adequacy addresses the main frequency part, which can be tracked down utilizing the formula. 

                     

• Padding with Zeroes

          Padding with zeroes is standard practice and valuable in numerous applications, doesn't appear to essentially further develop our information in this application. However we have double the frequency resolution, it's not yielding any better information. Fast tests padding with a lot more occasions (1 section test, 9 sections zeroes) show that however, the pinnacles improve frequency resolution, the inconsistency between the most noteworthy point on the first example and the most elevated point on the cushioned example is I Hz probably, which implies that it is likely, not worth the work, even at fewer frequencies.

Fig. 2 Flow- Chart

               The second we press one note to the prompt other note, the amplitude is at first sufficiently high and diminishes with time. Assuming that we can recognize the term of each note from the time-space attributes, we can distinguish and recognize the frequency.

Steps:

1. Averaging

    As there are countless examples for pretty much nothing and numerous changes are additionally present, the initial step is averaging, where for every 100 examples normal is found and the worth is appointed to the principal test, again for the following 100 examples the normal worth is allocated to the second example. This won't just diminish the number of tests yet additionally eliminate the variances present.

Fig. 3 Effect of Averaging

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

2. Thresholding

Constant Thresholding: After averaging, we really want to distinguish the tops from the averaged signal. As the name recommends, in steady thresholding one ideal worth is chosen for which we can get the greatest number of pinnacles.

    

Adaptive Thresholding: There are the potential outcomes when we take some consistent edge esteem, for certain notes it very well might be higher than the most extreme worth of the note, and for certain notes, it is low enough with the end goal that two pinnacles of the note get combined and just one pinnacle (one note) could be distinguished. To beat this issue the idea of adaptive thresholding came into the picture.

Indeed, even in adaptive thresholding, one issue is the point at which there is a long quiet between two notes, the edge an incentive for the quietness will be exceptionally low that expected to to be disposed of.

2. Width Selection for Finding Instant

When the thresholding is done, our point is to recognize the event of the tops from the signal we get subsequent to thresholding.

The width of the window contains each of the zeros are chosen based on a worst-case situation. For our melody it is found as 57, so the width was picked as 50 for wellbeing. The worst-case relies upon the speed for a given sampling frequency, For low tunes it is metal and for quicker ones, it is less. So independent of the melody, the base width was chosen.

3. Identification

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

4. Padding Zeroes

For a given note length, on the off chance that crop the comparing moments of the first signal and observe the discrete Fourier change, the outcomes are too far off as required. So we really want to cushion the zeroes. We can cushion the zeroes with various lengths and various pieces of the trimmed signal and observe the DFT that will give the various outcomes. The diverse part implies it very well may be either just before the edited adaptation after it or both the sides with various lengths. From these various varieties, the nearest results are gotten in the case where the part of zeroes with a similar length is cushioned on the two sides of the edited form. So DFT of the resultant signal is found to decide the documentation.

4. 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. and actual note is assigned.

Fig. 4 Discrete Notes of song

Simulation Results 

The experiment was finished with various melodies like Happy Birthday, Jingle Bell, Twinkle-Twinkle,..etc

Fig. 5 Entire Songs

Fig. 6  Original Song and Notes

Conclusion

The Frequencies of a song is detected, corresponding notes are identified with duration. The method used here for note identification is more optimized than previously used methods. By varying the parameters like threshold values and width, we can get the desired results with the time duration of each note.