Project Overview:
A hallmark of effective prediction applications is access to a large, high-quality dataset that captures the underlying dynamics and variables. Bird enthusiasts have been recording bird songs for decades, creating deep repositories of labeled audio data. These resources enable the development of models that can match new audio recordings to specific bird species. This project used a European dataset from Xenocanto to improve upon benchmark prediction accuracy traditionally achieved through manual identification.

Approach:
Although audio files are naturally represented as time-series data (suited for recurrent neural networks), they can also be transformed into images using Fast Fourier Transform (FFT). These spectrograms allow for more compact analysis with convolutional neural networks (CNNs). Audio data was segmented into uniform lengths, with silences removed and white noise filtered out. To address class imbalance and enrich the dataset, additional spectrograms were generated by varying FFT parameters. Multiple neural network architectures and techniques, including autoencoders, were tested and tuned to maximize predictive performance.

Project Results:
The team was able to achieve 95% validation accuracy for the Top 10 bird songs, surpassing the benchmark manural idenfication accuracy of 85%. The work was then extended to cover 78 bird songs and calls, with an intial accuracy of 70% and with further refinement possible to improve results. Finally, a multi-output neural networks was developed to predict not only bird species but also call type (song vs. call).

• Milestone 4: Baseline Model and Refinement Discussion
• Milestone 5: Final Model Discussion and Areas for Expansion
• Final Model (Full Python/Jupyter Code)

Project Partners: 

Imran Naskaki

Sandra Forro

Notes on Data Used for Project:

The dataset is acquired from Global Biodiversity Information Facility’s website:
https://www.gbif.org/dataset/b1047888-ae52-4179-9dd5-5448ea342a24
This data shared by GBIF is infact a subset of the entire collection of the Bird sound collection of Xeno-canto (XC), the Foundation for NatureSounds in the Netherlands, available at
https://xeno-canto.org/?gid=1 . There are more than 700k occurences of bird sound recordings, verifi edand shared by Xeno-canto.

Featured Image: