160 likes | 245 Views
Feature Vector Selection and Use With Hidden Markov Models to Identify Frequency-Modulated Bioacoustic Signals Amidst Noise. T. Scott Brandes. IEEE Transactions on Audio, Speech and Language Processing,2008. Outline. INTRODUCTION METHODS EXPERIMENTAL RESULTS AND DISCUSSION CONCLUSION.
E N D
Feature Vector Selection and Use With HiddenMarkov Models to Identify Frequency-ModulatedBioacoustic Signals Amidst Noise T. Scott Brandes IEEE Transactions on Audio, Speech and Language Processing,2008
Outline • INTRODUCTION • METHODS • EXPERIMENTAL RESULTS AND DISCUSSION • CONCLUSION
Introduction • A great need for automatic detection and classification of nonhuman natural sounds • Reduce bird-strikes by aircraft • Avoid bird-strikes of wind turbine generators • With the surge of interest in monitoring the effect of climate change • Monitor elusive species that can be indicators of habitat change • A range of techniques have been employed to detect sounds • Dynamic time warping • Hidden Markov models • Gaussian mixture models
Introduction • Improve bioacoustic signal detection in the presence of noise • Measurements of the peak frequencies directly • Pitch determination algorithms • Spectral subbandcentroid and their histograms are used to extract peak frequency • Extract first three formants with Linear predictive coding coefficients
Introduction Basic shape variety and type of calls
Methods HMM Use With Automatic Call Recognition (ACR) • To find the call that maximizes the probability • In the model testing stage, the equation is maximized with a Viterbi search The conditional probability p is calculated for each state transition The conditional probability is calculated for each feature vector observed during that state transition
Methods Creating Frequency Bands
Methods Applying the ThresholdingFilter • A value greater than average value in that band are kept, and the others are set to zero Extracting Features for Each Event and Detecting Patterns With HMMs • Peak frequency • Short-time frequency bandwidth
Methods Using a Composite HMM to Detect Higher Level Patterns
Conclusion • The performance of this process is most sensitive to the threshold-band filtering step • The contour feature vector used with the initial stage HMM is most effective • The sequence feature vector used with the second layer in the composite HMM is very effective at classifying sequences of calls