Zuniga_Classification_of_bird_sounds

Introduction

There's lots of research for species recognition and classification. But what about categorizing the sound by our perception or types of sounds. For the task of bird sound synthesis, finding suitable species for a particular synthesis model is a problem. Classifying birds by features of their sounds provides way to narrow down which birds work best.

Literature Review

Bird song and bird communication is one of those subjects that is interesting to biology and audio scientists. Sounds happens to carry an abundance of information regarding these animals. For example, Stowell (2016) use recordings of zebra finch birds to build a model of call timings which furthers the understanding of animal communication and can subsequently be used as the basis of a synthesis model of calls within a group of birds.

Birds are also a great source of biological indicators for ecological research. They respond quickly to environmental changes. The advances in portable recording technology in the last couple of decades opens the doors to an easy way to gather massive amounts of audio data just by leaving a recorder outside. This data can give us information such as population flux and migratory patterns which may lead to clues on other issues such as global warming or even the state of other animals. However, field recordings are noisy so Briggs (2013) tries to tackle the species classification problem and make it work better on noisy environments.

Methodology

• • Dataset

    • 36 recordings of different bird species, commonly found in the UK.

    • manually cut , normalized and the noise reduction effect on audacity applied on some very noisy recordings. the noise reduction algorithm does spectral noise gating, which works by creating a mask from a manually selected noise section of audio and using it to reduce noise.

  • Feature set

    • Spectral Bandwidth: gives a measure of the amount of harmonic content in the sound.

    • RMS Variance: Commonly used as a measure for dynamic range.

    • Spectral Flux: How quickly the power spectrum changes.

Analysis

Feature space and classification analysis was inspired by McAdams (1995) who use a three dimensional feature space to find classes that describe differences in timbre of musical instruments.

In this case we will be classifying bird sounds so the extracted features are different.

How the features describe the sound:

Spectral Bandwidth:

Birds with a high spectral bandwidth create sounds with lots of harmonics. The harmonics may be present in different ways such as:

The Black-headed_Gull which has a harsh sound with a noise-like quality to it:

As opposed to the Black Woodpecker which has a pleasant quality to it coming from its evenly spaced harmonics.

Both birds have high spectral bandwidth values but have different harmonic qualities.

On the other end,  the Eurasian Golden Oriole presents a very narrow bandwidth.

RMS Variance:

In the case of bird sounds RMS variance describes fast amplitude changes in the sound’s amplitude envelope. It will not tell us about fast changes, only the sound envelope as a whole.

Bringing the Redpoll as an example, its sound consists of very high frequency amplitude modulation. However, it shows up as the lowest in the rms variance scale because there are no fast changes in the overall envelope, it has a very consistent loudness throughout.

and the Common Redshank:

and the Eurasian Jay.

As we can see, the Eurasian Jay shows fast changes in frequency while the Dunlin doesn't. However,  The Dunlin applies very intense and fast amplitude modulation to its sound, as seen on the waveforms.

On the other end of the spectral flux spectrum we see the Eurasian Collared Dove and the Common Snipe :

Results

Classification was done using the k-means clustering unsupervised learning algorithm. The features were first normalized using ‘l2’  norm and then passed through the classifier.

Setting the classifier for three classes finds three main groups of birds that lie on the 3d space.

2d views:

The largest group is of low RMS variance, low spectral flux, and all around consistent bandwidth. This group contained the highest number of birds. 23 out of 36 which is 63.9 percent of birds.

Then there is the smallest group, which are the RMS variance outliers. 4 out of 36 which is 11.1% of birds.

And finally, the high spectral flux, low RMS variance class. 9 out of 36 which is 25% of birds.

The largest gap in the bird distribution was the area with high spectral flux and high rms variance. It is inconclusive whether the gap was due to a lacking dataset or it really is an area of sound types where there are no living bird examples of.

Conclusion

In conclusion: bird sounds are extremely diverse. That may be an obvious statement but sitting down and quantifying their differences leads to many interesting observations. This work barely scratches the surface of the amount of diversity in bird sounds. For example, bird vocabulary is an important subject that demands more consideration. Some birds have a diverse vocabulary that may well end up putting them in many clusters at once. Other birds have very unique sound making mechanism that may need a whole different feature set and other considerations to be  modeled.

Drawbacks

Dataset may not be able to accommodate the huge vocabulary of come birds. a short excerpt of song/call was used for each species so it only represents a single word.

More data could either fill the feature space better of just populate existing found clusters. It would be interesting to see how the clustering changes as more species and different words are introduced to the dataset. It may be

Further work

There is a huge amount of interesting features that we can extract from birds. For example, rate of their calls, extent of their vocabulary, and identifying calls versus song. However, designing the algorithms required to extract those features is a whole different task which may be further explored. Also, There needs to be better ways of using the huge amounts of data at our disposal. Data is unreliable, and finding faster and better ways of cleaning it up, organizing, and labeling it is always useful.

References

Stowell, D., Gill, L. and Clayton, D., 2016. Detailed temporal structure of communication networks in groups of songbirds. Journal of the Royal Society Interface, 13(119), p.20160296.

Briggs, F., Huang, Y., Raich, R., Eftaxias, K., Lei, Z., Cukierski, W., Hadley, S.F., Hadley, A., Betts, M., Fern, X.Z. and Irvine, J., 2013, September. The 9th annual MLSP competition: New methods for acoustic classification of multiple simultaneous bird species in a noisy environment. In 2013 IEEE international workshop on machine learning for signal processing (MLSP) (pp. 1-8). IEEE.

McAdams, S., Winsberg, S., Donnadieu, S., De Soete, G. and Krimphoff, J., 1995. Perceptual scaling of synthesized musical timbres: Common dimensions, specificities, and latent subject classes. Psychological research, 58(3), pp.177-192.