Musical instrument recognition and tone identificationLiterature survey by: Keerthi C. NagarajDate: 5thMarch 2003Objective:To define a musical instrument in terms of its spectro-temporal characteristics, and use the information to track the notes played by the instrument in a given recording.Applications:•Content based audio retrieval•Automated Music Transcription•Karaoke track(Music Minus One) production, •Computer participation in live performance, etc.Background:The identification of music is done among human beings based on:•Recognition of the instrument – timbre, pitch, spectral brightness, and other spectro-temporal features •Recognition of the tone – melody line, bass line, pitch to pitch transition, rhythmTo make a computer do the same thing, we model a musically trained human-like auditory system.Issues involved:•Timbre – mathematically ill defined quantity, different results by musicians and non musicians•Tone– multiple overlapping frequency components, noisy environment etcApproach to the solution in 3 stages:•Cochlear Modeling•Feature extraction •Pitch detection and trackingSTAGE 1: Modeling the CochleaLog lag frequency model of the human ear with autocorrelogramextractionMeddis, R. and Hewitt, M.J. (1991) 'Virtual pitch and phase-sensitivity studied using a computer model of the auditory periphery: I pitch identification', Journal of the Acoustical Society of America, 89, 2866-2882Stage 2: Feature extractionSpectral features: spectral centroid, Average relative spectrum, High frequency roll off rate, intensity, spectral envelope (becomes important later)Pitch, vibrato & tremolo features: pitch range, absolute strength and relative strength and phase (in comparison to vibrato)Attack features: Onset asynchrony, inharmonicity, etcBrown, J.C. (1999). ``Computer identification of musical instruments using pattern recognition with cepstral coefficients as features'' J. Acoust. Soc. Am. {\bf 105}, 1933-1941Stage 3: Pitch tracking and Note identificationThe method proposed in this paper provides a tradeoff against the computational rigor involved in the Meddis and Hewitt model. The key differences are:•warped linear prediction for pre-whitening filters,•Uses critical band frequency resolution rather than uniform frequency resolution•2 channels for computational efficiency and almost same performance as that of Meddis and Hewitt model (40-128) channels.Demerit: It does not use inputs from features extracted!Another interesting approach:Goto. M, “A Robust Predominant-F0 Estimation Method For Real-Time Detection Of Melody And Bass Lines In Cd Recordings” ICASSP 2000 (2000 IEEE International Conference on Acoustics, Speech, and Signal Processing) Proceedings, pp. II-757-760, June 2000.Tolonen, T., and Karjalainen, M., ''A Computationally Efficient Multipitch Analysis Model,'' IEEE Transactions on Speech and Audio Processing, vol. 8, no. 6, pp. 708-716, November 2000What I propose to do:•A “selective attention mechanism” can be devised to track multiple sources simultaneously using sound source discrimination based on the features extracted. •Comparative analysis of the pitch tracking mechanisms with the main aim of application to music transcription will be undertaken. • A working re-synthesis model will be designed to compare the performances of different models. •Possibility of enhancements in the pitch tracking method with the addition of prior knowledge about the musical instruments being played will be explored.•Analyze the performance improvement in comparison to the traditional pitch
View Full Document
Unlocking...