Within Malaysia, one of the oldest Malay’s art heritage is traditional music with accompanying instruments where knowledge is passed down from one generation to another. These days, they still appear as a medium of communication and entertainment by supplementing the performing arts in the local community (Kechot 2013). Henceforth, the existence of the musical instruments symbolised proof of the old local’s wisdom and understanding of social needs. The common ones are cak lempong, gambus, gendang, gong, kompang, seruling (flute), rebab and rebana. They produced vast acoustic diversities with varying tones, melody and rhythms. Alternatively, the sound can be converted and observed in the time series and frequency domain (Darji 2017). Any changes of physical characteristics concerning time-frequency components, such as intensity and exciting frequency, can be easily attained. Further, they can be described numerically with descriptive statistics, which primarily used statistical measures to quantify them as signal features (Chaudhary & Kakarwal 2019).
(Image: hmetro.com.my)
Typically, musical devices are distinguished by their physical appearance and sound made through human perception. Despite this, errors frequently happen due to human misinterpretation, weak verdicts, and interferences (Ferguson 2006). Therefore, the musical classification technique was explored to enhance and aid people by applying specific statistical classifiers. For example, Prabavathy et al. (2020) utilising SVM and k-NN algorithms to classify several musical instruments by extracting Mel-Frequency Cepstral Coefficients (MFCC) and Sonogram features as input data. They found out that SVM performs better. Hence, A musical instrument sound classification technique needed input data with the best feature scheme, which is crucial for an effective pattern recognition system in the music information retrieval (MIR) field. Concerning the Malay’s, past researchers conducted recognition studies with several traditional instruments by utilising rough and soft set theory design several years ago (Senan et al. 2012). They used MFCC and Perception-based measurement that brought out 37 features that later filtered to 17 for the best classification accuracy with a 1-NN classifier.
However, apart from that, there has been little to none of the latest contributions regarding the exploratory data analysis of Malay’s traditional musical instrument sound signals for scientific discovery and analytical purposes, especially in representing the data graphically. To cope with the redundancy, we have led new findings by subjecting the acoustic signal data of the instruments with statistical signal analysis and clustering techniques in the decision-making phase. The main idea is to find the difference between Malay’s instruments and visualise them graphically by benefiting from the acoustical sound signal they produced. The experimental scheme is represented as in Figure 1.
The design of the experiment was simple but enough to capture and store the individual’s musical note audio signal of the instrument in the .wav format files. The individual audio signal data were then extracted in the form of a time-domain using MATLAB programming. Figure 2 shows the example of time series produced for different instruments but with the same musical note, a B3.
Cak lempong and gambus yield transient signals with the former shows a better exponential shape, while the flute instrument produced a nearly constant amplitude across time before diminishing rapidly. Yet by observing only in the time-frequency domain, there was no other way to describe and quantify the instruments differences rather than the signal’s physical characteristics. These are the primary reason statistical measures are needed to describe the components in several detail levels and thus give a better understanding and meaning (Ahmad et al. 2017).
Therefore, this article will highlight the use of advanced statistical signal measurement of Integrated Kurtosis-based Algorithm for Z-filter (I-kazTM), Mesokurtosis Zonal Non-parametric (M-Z-N) and Z-freq to cypher the audio data into signal features. I-kaz decomposes a time-domain signal into three frequency ranges and measures the scattering data points from the centroid. Meanwhile, M-Z-N divides the signal into five segments and measure data’s scattering points across every segment with respect to the root mean square (RMS). Last but not least, Z-freq method shares similarity with I-kaz but with the major differences lies in the measurement of signal’s components in the frequency domain instead of time series. Each of them is quantified as coefficients using equations as represented in Figure 3 (Ngatiman et al. 2018; Bahari et. al 2013; Nuawi et. al 2008).
In the first phase, a dissimilarity study of a single but the same musical note over three distinctive instrument categories was developed by doing a repetitive recording. Each of the instruments’ notes was analysed using I-kaz and M-Z-N statistical signal analysis methods to extract their coefficients and mapped on a single graph for clustering purposes, as shown in Figure 4. Although using the same musical note, a different cluster with a different instrument was produced on specific zones.
Later on a second phase, a whole set of instruments’ musical note signals were processed, in which the final objective was to find the real or actual cluster of the individual devices. A total of 84 musical notes from 4 instruments were used. Statistical signal features of I-kaz and standard deviation were extracted from every note and mapped on a two-dimensional graph. Table 1 shows every musical notes used for each instruments, while Figure 5 displays the graphical solution of the clustering approach. The attempt was successful, with the result showing that each instrument could be grouped and clustered in its categories.
Moving forward from there, an analytical approach on the instruments’ pattern recognition was pursued by utilising and comparing two unsupervised learning of machine learning algorithms; k-means and spectral clustering. Those two were considered based on popularity and general efficiency. Advanced statistical features of I-kaz and Z-freq were extracted from all instruments’ musical note signal data. Construction of external reference was made by mapping all the musical notes’ acoustic signal features and marked according to their instrument to form four clusters or groups. on a single graph and denoted as the true cluster for performance validation. Figure 6 depicts the feature scheme’s real cluster, while Figures 7 and 8 portray the clusters recognised automatically by k-means and spectral clustering algorithms. This time, aside from utilising machine learning, they tried to cluster those four instruments into their own individual’s groups.
By comparing the k-means and spectral clustering graphical result with the external reference, the latter showed excellent pattern recognition, giving clusters almost identical to the actual one, with only a tiny amount of mistakenly identified data belonging to other clusters or groups. The main reason because spectral tends to cluster together with the points where there was the edge of connectivity that form nodes, while k-means labels the points where the centroid was the closest. In that sense, spectral provides a much more flexible approach for uneven data allocation such as these.
This study could be extended further by investigating various signal features and classifiers on various musical instruments so that the wide gaps and similarities between them could be discovered and later establish the best feature scheme and statistical model for automatic pattern recognition.
To end, the authors want to thank the National University of Malaysia (UKM) for providing the necessary sources and funding to complete the project.
Contributors: Mohd Zaki Nuawi, Muhammad Arif Fadli Ahmad, and Suziana Mat Saad
Universiti Kebangsaan Malaysia (UKM)
mzn@ukm.edu.my
References
Kechot, A. (2013). Kajian berkaitan alat muzik tradisional Melayu: satu penerokaan awal/Ab. Samat Kechot.
Darji, M. C. (2017). Audio signal processing: A review of audio signal classification features. International Journal of Scientific Research in Computer Science, Engineering and Information Technology, 2, 227-230.
Chaudhary, S. R., & Kakarwal, S. N. (2019). Various Approaches in Musical Instrument Identification: A Review. International Journal of Applied Evolutionary Computation (IJAEC), 10(2), 1-7.
Ferguson, S. (2006, June). Learning musical instrument skills through interactive sonification. In Proceedings of the 2006 conference on New interfaces for musical expression (pp. 384-389).
Prabavathy, S., Rathikarani, V., & Dhanalakshmi, P. (2020). Classification of Musical Instruments using SVM and KNN. International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN, 2278-3075.
Senan, N., Ibrahim, R., Nawi, N. M., Yanto, I. T. R., & Herawan, T. (2012). Rough and Soft Set Approaches for Attributes Selection of Traditional Malay Musical Instrument Sounds Classification. International Journal of Software Science and Computational Intelligence (IJSSCI), 4(2), 14-40.
Ngatiman, N. A., Nuawi, M. Z., & Abdullah, S. (2018). Z-Freq: Signal Analysis-Based Gasoline Engine Monitoring Technique Using Piezo-Film Sensor. International Journal of Mechanical Engineering and Technology, 9(5), 897-910.
Nuawi, M. Z., Nor, M. J. M., Jamaludin, N., Abdullah, S., Lamin, F., & Nizwan, C. K. E. (2008). Development of integrated kurtosis-based algorithm for z-filter technique. Journal of applied sciences, 8(8), 1541-1547.
Bahari, A. R., Nuawi, M. Z., Ab Samad Kechot, F. M. H., & Saad, S. M. (2013). Pemetaan dan pengelompokan isyarat akustik alat muzik tradisional melayu menggunakan kaedah analisis isyarat statistik. Jurnal Kejuruteraan, 25, 33-38.
Ahmad, M. A. F., Nuawi, M. Z., Bahari, A. R., Kechot, A. S., & Saad, S. M. (2017). Correlation and clusterisation of traditional Malay musical instrument sound using the I-KAZTM statistical signal analysis. Journal of Mechanical Engineering and Sciences, 11(1), 2552-2566.
