1 / 17

Audio Tempo Extraction

Audio Tempo Extraction. Presenter: Simon de Leon Date: February 9, 2006 Course: MUMT611. Agenda. Introduction Algorithm Onset extraction Periodicity detection Temporal estimation of beat locations Examples Conclusion Discussion. Introduction. Tempo extraction is useful for

PamelaLan
Download Presentation

Audio Tempo Extraction

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Audio Tempo Extraction Presenter: Simon de Leon Date: February 9, 2006 Course: MUMT611

  2. Agenda • Introduction • Algorithm • Onset extraction • Periodicity detection • Temporal estimation of beat locations • Examples • Conclusion • Discussion

  3. Introduction • Tempo extraction is useful for • Automatic rhythm alignment • Beat-driven effects • Cut & paste operations in audio editing • Tempo extraction in general is mature for straightforward, rhythmic music (rock, rap, reggae, etc.) • The challenge is to be accurate across the widest range of genres

  4. Introduction • We will focus on the winning algorithm for MIREX 2005 [1] • The top algorithms belong to the class that performs the following: • Time-freq. analysis to determine beat onset • Pitch detection and autocorrelation techniques for periodicity estimation • Evaluation of algorithms is difficult due to different perceptions of rhythm

  5. Algorithm • Divided into three sections • Onset extraction • Where are the exact locations of the musical salient features? • Periodicity estimation • What is the tempo of the beats found? • Temporal estimation of beat locations • We found the onset locations in the spectral domain, but they are not all necessarily the beats

  6. Algorithm – Onset extraction • Idea is that the beat onsets correspond with • Note changes • Harmonic changes • Percussive events • Define spectral energy flux • Time derivative of the frequency component magnitudes • Technique of [1] assumes onsets correspond to the fastest change of frequency component magnitudes

  7. Algorithm – Onset extraction • Step 1: Take STFT of signal • Step 2: Take time derivative of frequency components (spectral energy flux) • a) Low-pass filter STFT magnitude • b) Apply logarithmic compression [2] • c) Pass through FIR filter differentiator [3] • Step 3: Use dynamic threshold and remove the smallest onset spectral energy flux “spikes” from previous step

  8. Algorithm – Onset extraction • Top left: Piano signal. Bottom left: STFT • Top right: Spectral energy flux. Bottom right: Detection function

  9. Algorithm – Onset extraction • Top left: Violin signal. Bottom left: STFT • Top right: Spectral energy flux. Bottom right: Detection function

  10. Algorithm – Periodicity Detection • Two techniques studied in [1] • Spectral product • Autocorrelation function • Assume tempo T is between 60bpm and 200bpm • Spectral product • Step 1) Take FFT of detection function • Step 2) For each frequency, multiply it by all of it’s integer multiples • Step 3) Largest product corresponds to frequency of periodicity

  11. Algorithm – Periodicity Detection • Autocorrelation function • Classical periodicity estimation, slightly outperforms spectral product method • It is the cross-correlation of a signal with itself • Three largest peaks of cross-correlation are analyzed for a multiplicity relationship

  12. Algorithm – Beat location • Given the tempo extracted from previous steps, we need to align the beat in phase • Step 1) Create pulse train q(t) with period Tderived from periodicity algorithm • Step 2) Find phase by cross-correlating q(t) with detection function, evaluating only at indices corresponding to detection function maximas • Step 3) For successive beats in an analysis window, simply add T and search for peak in detection function in vicinity • Repeat (2) to re-align phase if peak not found

  13. Examples • Let’s listen to some demos of the algorithm in action • Jazz – very good to good • Rock – very good • Classical – very bad to good • Soul – very good • Latin – satisfactory to good

  14. Conclusion • This algorithm represents the state-of-the-art in tempo extraction, the majority of the work focusing on onset detection • Problem areas • Long fading attacks and decays produce false onsets • Many instruments playing continuously with no stable regions produces too many false onsets • Cannot keep up when tempo varies quickly

  15. Conclusion • Results from [1] indicate roughly • 80-90% accuracy for classical, jazz, rock • 90-100% for latin, pop, reggae, soul, rap, techno • Results from [1] using MIREX database • 95% of the time gave correct tempo

  16. Discussion • Can evaluation methods be improved? How can we avoid the subjective nature of tempo perception? • Any suggestions on how we might improve the onset detection algorithm? How about the periodicity algorithm?

  17. References [1] Alonso, Miguel, Bertrand David, and Gael Richard. 2004. Tempo and Beat Estimation of Musical Signals. Proceedings of the 5th International Conference on Music Information Retrieval. [2] Klapuri, Anssi. 1999. Sound Onset Detection by Applying Psychoacoustic Knowledge. Proceedings of the IEEE International Conference of Acoustics, Speech and Signal Processing: 3089-3092. [3] Proakis, John G., and Dimitris K. Manolakis. 1995. Digital Signal Processing: Principles, Algorithms and Applications. 3rd Ed. New York: Prentice Hall.

More Related