1. Beehive Audio Source Separation
Dakota Murray

2. What is Audio? Probably zoom in on the frequencies less than 2500 Hz.
Just a bunch of numbers!
Time domain
Frequency Domain
Spectrogram
Probably zoom in on the frequencies less than 2500 Hz.

3. The Spectrogram
Each column is a frequency spectrum at a point in time
Discretized. Tradeoff between number of time frames, and number of frequency bins
Discretized? Tradeoff between time resolution and frequency resolution.

4. Signal Separation
Separate sources from a mixture
Paris Smaragdis
Signal Processing researcher
Approached signal separation probabilistically
He models each source as a mixture of distributions over the frequency spectrum. The resulting spectrogram is a mixture of these sources and noise.

5. Smaragdis’s Approach 1
A source can be represented as a stochastic “picker” randomly selecting from a set of jars
Each jar is a component of the source and each has different distributions of marbles inside
A frequency is written on each marble
Given a component z, what is the probability of the picker pulling frequency f

6. Smaragdis’s Approach 2
Mixtures of sources add another variable, s, to the model.
If there are two sources, our picker is now selecting from two sets of jars, each set with their own distributions
Given a source s, and component z, what is the probability of the picker pulling frequency f

7. Original Signals
Voice
Chimes
Separated Signals

8. Training Target x

9. Training Interferencex

10. Step 2: Separation
Decompose the mixture of our known sources in terms of the contributions of the basis vectors for each source
In other words, find out how each source contributes to the mixture.

11. Step 3: Reconstruction * = * =

12. Variation on the Process
Supervised
We have an example of target and interference
Semi-Supervised
We only know the target. Interference could be anything
Train the target as normal. Learn Interference from our mixture. Interference represents everything that isnt the target.
Fixed Target: Target Vectors can not change
Target As Prior: Target Vectors are allowed to change

13. The Importance of Honey Bees
Honeybees are dying from Colony Collapse Disorder
Honeybee pollination results in ⅓ of the food humans consume.
An agricultural, ecological, and economic issue.
Data is useful. Good data is scarce.
Often full of interference from traffic, weather, and electronics
Cite all images and equations from somewhere else.

14. Bee Hive Audio Considerations
Can’t be certain what type of interference to expect. Should use a Semi-Supervised method of training the interference basis vectors
Don’t want to assume we know exactly what the bees will sound like. Use a Target as Prior method of training the interference, thus allowing the target to change.
Might be getting a lot of data. Must optimize parameters for best performance to execution time ratio
Testing should involve more real world scenarios

15. Problems
To get good results from a complicated signal, the algorithm is time consuming
The training process only find local minimum, may not produce idea set of basis vectors.
Has difficulty separating sources with occupy similar frequency ranges.

16. Direction for Future
Testing, testing, testing, and more testing (optimizing parameters, trying out different testing scenarios)
Gathering more bee hive data. App State students currently working on Low Cost hive monitoring hardware
More real-world scenarios
Work with other types of data, not just bees.

17. Acknowledgements Paris Smaragdis for paving the way
Dr. Parry for his mentoring, assistance, and wealth of knowledge
Dr. T and the Rest of the App State Department of Computer Science for providing the encouragement and opportunity
NC State for hosting this event!

18. References
Bhiksha Raj and Paris Smaragdis, "Latent variable decomposition of spectrograms for single channel speaker separation," in Proc. of IEEE Workshop on Applications of Signal Proc. to Audio and Acoustics, New Paltz, USA, 2005.
P. Smaragdis , B. Raj and M. V. Shashanka "Supervised and semi-supervised separation of sounds from single-channel mixtures", Proc. 7th Int. Conf. Ind. Compon. Anal. Signal Separation (ICA\'07), pp
M. V. S. Shashanka , B. Raj and P. Smaragdis "Sparse overcomplete decomposition for single channel speaker separation", Proc. Int. Conf. Acoust., Speech, Signal Process. (ICASSP\'07), vol. 2, pp.II-641 -II