Towards Timbre Recognition of Percussive SoundsSkip other details (including permanent urls, DOI, citation information)
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License. Please contact email@example.com to use this work in a way not covered by the license. :
For more information, read Michigan Publishing's access and usage policy.
Page 00000001 Towards Timbre Recognition of Percussive Sounds Adam Tindale, Ajay Kapur, and Ichiro Fujinaga Music Technology Department, McGill University [ tindale, ich ]@music.mcgill.ca Electrical and Computer Engineering, University of Victoria ajay @ece.uvic.ca Abstract The development of computer algorithms for music instrument identification and parameter extraction in digital audio signals is an active research field. A musician can listen to music and instantly identify different instruments and the timbres produced by various playing techniques. Creating software to allow computers to do the same is much more challenging. This project will use digital signal processing and machine learning techniques to differentiate snare drum timbres produced by different strike positions and strike techniques. 1 Introduction There have been many studies into instrument classification (Herrera, Peeters, and Dubnov 2003) and percussion instrument classification (Gouyon and Herrera 2001; FitzGerald, Coyle, and Lawlor 2002; Silpanpdi 2000). This study is an attempt at providing additional tools that classify the subtle differences in the timbre that a snare drum can produce. In particular, this paper will investigate differences in snare drum timbres produced by different strike positions and strike techniques. Instrument classification systems try to classify the signal into a particular timbre space that represents a family of instruments or particular instruments. These systems do not provide provisions for the different timbres produced by a single instrument. However, these systems are able to place a timbre produced by an instrument into a feature subspace. Therefore it must be possible to examine the subspace and learn to identify the different timbres represented within the subspace. By identifying these different timbres it will be possible to use this information to allow other systems to provide more meaningful information to the user. An interesting possibility is to find ways to use this data in a realtime context. In order to allow for the possibility of realtime recognition of these timbres there must be careful consideration of the fea ture extraction algorithms. The eventual goal is to recognize the recognition data as control parameters in order to eliminate the need for sensors to capture snare drum strike position. 2 The Snare Drum The feature that distinguishes the snare drum from other drums is the snares, which are wires stretched across the bottom head of the drum that vibrate in sympathy when the drum is struck. The snare drum can be struck at different points along the radius of the batter head to produce different timbres because different modes of the membrane are excited and dampened (Rossing 2000). The snare drum may also be struck with a wire brush, which creates a unique timbre, or it can be struck so that the stick makes contact with the snare head and the rim at the same time producing a rimshot. While there has been some research on snare drums, very little of it deals directly with timbre recognition. The first major published study on the snare drum was mostly concerned with amplitude and durations (Henzie 1960). The study scientifically introduced the idea of a stroke height, the height that the stick starts its strike from, as being the major factor in the resulting amplitude of the strike. Complex interactions between the modes have been observed and discussed (Zhao 1990), which is useful evaluating what type of features to look for when trying to classify timbre. An empirical study (Lewis and Beckford 2000) showed that different types of snare drum heads on the same drum can produce varying timbres. Another study of this nature showed spectra from a snare drum with its snares engaged and not engaged (Wheeler 1989) that demonstrated a large difference in timbre. See Figure 1 for representations of snare drum signals. Proceedings ICMC 2004
Page 00000002 Brush Strike 2305 6............38..260..80 68 aA I E 0 -- 34 010 23 45 68 Halfwaye5120 3840 2560 1280 68 1ERdmshot Strike as s < E 034 0 23 45 68 12 840 2560 12.1 Center Strike 0 0 23 45 68 <5120 3840 2560 1280 6 1- Halfway Strike ferences in timbre of snar EdgeSrikeuds n f h airA 0.54 E0.5 - Tima __Ie) (ins t a3840 2560 -1280 ---e68 i m ( Figrder 1: timte Domain mand Frequaenc Domaiemntreporesnta3 2softw re resign 3.1 Feature t eracetnt Algoritm STanar features (Hperrerad Yetarianud and Gouyo 2002)aed werte exaine-dforan thei uFTsefulessi classifying thsubytlem df feenesyintimbrses of snuare widrum sounds One ofthec macion cri theri fornd feature is thatithe arte posintbletowcompthe in shot order sod thatte systemam ayit be. ltripeeneorni raTime. Tefeaturswres are intended ton texatract geeralocharaceitc fthe snare drum inndThfotrderst bnlue: zrobustroshandle different drumst and Diferuen 20t3pl 00),yerimSa thepfeatuesimleentedi.Ainc thiszrocrstudy areld al calulate freueny eTimatio basd onthe qumenyofHzer) rssns in the attTime-Domain, n a FFTare qncmuteD aincths system. her Sofst armpes) as in the attack section of hel utnd wuhc is dened as the pitack betwen. th net temandarddf the peeaapte.nd reofgra The-fawere allnda i th coqutesd oenre attaso lsuection fdtersnae nrum siondhspeifcbatdurs. i nlue: ztercrefoshingr (Gouyon Pac00Het, and0Deleru 2z 000),00 atacztm, RMS-0,0 an tema foral entrsisd Attak ftieyzerpo-scritossipngytels eacrdhe freqenyes o htimtion baste a edonthripemnmernfter o rossing inte alti.Tacksecto.Atursack itienyieldstetheactugnealtimea(i numberisofcsamlsoffthethae dattacksrec tieon.uAtacktimeanRMS yifeeldstheramouantoifenerg nthpaeratcs scin.Atcktm Temporaleturoidimpeeldsthedlocathionofthedcenreaof grcuavtyd of the winedowmbing examinsaed. muedi hs ytm Tesub-bandseanalysisatechiniquesowerte atalsoemloedtiondetermieeandh ergyinmpecificu and. Curnlteeaefr bands: 0-200 Hz, 200-1000 Hz, 1000-3000acktHz, 300020,00 Hz. Wehesignck edvaious ftak iter fncionds tohepacteuachsig-(i nal into 4 different subbands. We designed various IIR filters of lengthh5 to separate the signals intosthe four different subbands. The energy in each band is measured during the attack phase, which gives a rough estimation as to what modes of the drum are being excited. 3.2 Classification Techniques Although there are many possible strategies available to classify data, this study uses a feed-forward backpropogation artificial neural network. This classifier was selected because it allows us to use an exemplar-based learning environment in order to draw analogy to the way in which a human learns to label a technique based its resulting timbre. The network is a three layer network with 6 hidden nodes, and 3 output nodes. 4 Results Snare drum sounds were recorded to a hard drive using a Mark of the Unicorn 896 set to 16-bit resolution and 44.1kHz sampling rate with a Neuman U-87 microphone placed near the edge and suspended perpendicularly over the drum. All strikes were played with an eight-inch stroke height by an expert player. A total of 100 recordings were made: 20 strikes at the center (actually off-center to avoid the "dead spot"), 20 strikes at the edge, and 20 strikes half-way between the edge and the center of the drum, 20 brushes strikes and 20 rimshots. The resulting files were segmented manually and their amplitudes were normalized so that only the timbral qualities were relevant. The files were also run through a gate function which truncated the sample from the previous zero-crossing above |0.1| and then kept only the first 10,000 samples. Leave-one-out cross-validation was used to select testing and training data for evaluating the system. 4.1 Stage 1 Three initial experiments were conducted that were expected to be progressively more difficult. Our first experiment was to distinguish between rimshot, brush strike, and center strike. The second experiment was to distinguish between edge strike, center strike and rimshot. The third experiment was to distinguish between edge strike, halfway strike, and center strike. Subband features were not used in this experiment. As shown in Table 1, the system was able to successfully classify rimshot and brush strokes with great accuracy. These two timbres are significantly different from standard playing techniques. The standard playing techniques proved to be more difficult to accurately classify but the system was Proceedings ICMC 2004
Page 00000003 successful enough that with further investigation into feature selection algorithms it can only increase. Another experiment was conducted to improve the accuracy of the identification of the standard playing techniques. The subband features were added to the system and experiment 3 was run again. This time center strikes were recognized 90%, middle strikes 95%, and edge strikes 87.5%. Exp. 1 Exp. 2 Exp. 3 Exp.4 Center 100% 85% 80% 90% Halfway - - 70% 95%~ Edge - 70% 70% 87.5% Rimshot 100% 100% - - Brush 100% - Table 1: Results 4.2 Stage 2 Five additional experiments were run with the introduction of some basic spectral domain features. The features used for testing were: spectral centroid, spectral rolloff, spectral flux, mel-frequency cepstrum coefficients, and linear predictive coding coefficients. Some of these features add an FFT component to the system. These features were calculated on a larger window of the sound. As in the previous stage, a gate function was used to determine the onset, which was then used as the beginning point of signal for analysis. The window was ended when the signal dropped below the initial threshold of |0.1|. Table 2 shows the results of several different experiments run with the selections of this feature set. The feature sets for these experiments is outlined below. 1. Spectral Centroid, Spectral Rolloff, RMS 2. RMS, Ramptime, Spectral Centroid 3. Ramptime, Spectral Centroid, Spectral Rolloff, Spectral Flux 4. Linear Predictive Coding Coefficients, RMS, Spectral Centroid 5. Mel-Frequency Cepstrum Coefficients Although none of these experiments provide outstanding results they use minimal features to achieve their task. Interestingly, the Mel-Frequency Cepstrum Coefficients provide satisfactory results on their own. The lower results of this phase of experimentation suggests that the initial window size is more appropriate for discriminating between the subtle timbres produced by the snare drum. Exp. 1 Exp. 2 Exp. 3 Exp. 4 Exp. 5 Center 85% 90% 90% 60% 80% Halfway 95% 75% 70% 80% 55% Edge 85% 85% 75% 85% 70% Rimshot 85% 90% 90% 65% 85% Brush 65% 75% 50% 70% 80% Table 2: Spectral Features Included 5 Future Research Given the very encourging results ovbserved in this research, further experiments are planned to investigate deeper into the subtle percussion timbre space A larger data set of snare drum recordings is being planned that will include multiple players on multiple drums. The problem of recognising the timbres produced at different strike points along the snare head seems to be more complex as the step sizes gets smaller. Along with increasing the recognition rate of the timbres investigated in this study, smaller step sizes will examined. The study will continue to shrink the step sizes until they become so similar that the system can no longer distinguish between them. The results of this research will then be applied towards other percussion instruments, beginning with tabla, to begin investigating strategies for timbre recognition systems that can differentiate multiple instruments as well as classify their timbre. As the system improves it is hoped that it can be included into broader recognition systems (Fujinaga and MacMillan 2000; Martin and Kim 1998) thus creating a new type of transcription system that is not only able to label the instrument but also provide timbral information about the musical signals' components. The performance of different windowing techniques will be evaluated more thoroughly throughout the course of this study. Along with the variable window size demonstrated in this paper, fixed width and multiple windowing techniques will be investigated. More spectral and time domain features will be included and tested for their effectiveness in classification. 6 Conclusion Our initial experiments into timbre classification of snare drums provides promising results which suggest that with further development and investigation better results can be achieved. As it becomes more apparent which features are major contributors, the system will be modified so that it will be able simply and accurately classify subtle differences in Proceedings ICMC 2004
Page 00000004 snare drum timbres. The system provided in this paper offers a classification system which can classify in realtime since the features are calculated in a a very small window. Hopefully this research will demonstrate to larger projects that not only is it possible to identify instrument types, but that it is possible to classify the different timbres produced by an instrument. References FitzGerald, D., E. Coyle, and B. Lawlor (2002). Sub-band independent subspace analysis for drum transcription. Proceedings of Workshop on Digital Audio Effects, 65-9. Fujinaga, I. and K. MacMillan (2000). Realtime recognition of orchestral instruments. Proceedings of the International Computer Music Conference, 141-3. Gouyon, E and P. Herrera (2001). Exploration of techniques for automatic labeling of audio drum tracks' instruments. Proceedings of MOSART: Workshop on Current Directions in Computer Music. Gouyon, E, E. Pachet, and O. Delerue (2000). On the use of zerocrossing rate for an application of classification of percussive sounds. Proceedings of Workshop on Digital Audio Effects. Henzie, C. (1960). Amplitude and duration characteristics of snare drum tones. Ed.D. Dissertation. Ph. D. thesis, Indiana University. Herrera, P., G. Peeters, and S. Dubnov (2003). Automatic classification of musical instrument sounds. Journal of New Music Research 32(1), 3-21. Herrera, P., A. Yetarian, and E Gouyon (2002). Automatic classification of drum sounds: A comparison of feature selection and classification techniques. Proceedings of Second International Conference on Music and Artificial Intelligence, 79-91. Lewis, R. and J. Beckford (2000). Measuring tonal characteristics of snare drum batter heads. Percussive Notes 38(3), 69-71. Martin, K. and Y. Kim (1998). Musical instrument identification: A pattern-recognition approach. Presented at the 136th meeting of the Acoustical Society of America. Rossing, T. (2000). The science of percussion instruments. River Edge, NJ: World Scientific. Schloss, A. (1985). On the Automatic transcription of percussive music: From acoustic signal to high-level analysis. Ph. D. thesis, CCRMA, Stanford University. Silpanpdi, J. (2000). Drum stroke recognition. Technical report, Tampere University of Technology. Wheeler, D. (1989). Focus on research: Some experiments concerning the effect of snares on the snare drum sound. Percussive Notes 27(3), 48-52. Zhao, H. (1990). Acoustics of snare drums: An experimental study of the modes of vibration, mode coupling and sound radiation patterns. Master's thesis, Northern Illinois University. Proceedings ICMC 2004