[Paper Review] Bird Species Classification And Acoustic Features Selection Based on Distributed Neural Network with Two Stage Windowing of Short-Term Features

Nahian I. Hasan|arXiv (Cornell University)|Jan 1, 2022

Animal Vocal Communication and Behavior4 citations

TL;DR

This paper proposes a two-stage windowing approach with a distributed 2D CNN-LSTM network for bird species classification, combining traditional signal processing and deep learning. It achieves 90.45% accuracy using 13 MFCCs and 12 Chroma vectors, with high specificity (98.94%) and AUC (94.09%), demonstrating effective acoustic feature selection and temporal-spectral feature learning for real-time applications.

ABSTRACT

Identification of bird species from audio records is one of the challenging tasks due to the existence of multiple species in the same recording, noise in the background, and long-term recording. Besides, choosing a proper acoustic feature from audio recording for bird species classification is another problem. In this paper, a hybrid method is represented comprising both traditional signal processing and a deep learning-based approach to classify bird species from audio recordings of diverse sources and types. Besides, a detailed study with 34 different features helps to select the proper feature set for classification and analysis in real-time applications. Moreover, the proposed deep neural network uses both acoustic and temporal feature learning. The proposed method starts with detecting voice activity from the raw signal, followed by extracting short-term features from the processed recording using 50 ms (with 25ms overlapping) time windows. Later, the short-term features are reshaped using second stage (non-overlapping) windowing to be trained through a distributed 2D Convolutional Neural Network (CNN) that forwards the output features to a Long and Short Term Memory (LSTM) Network. Then a final dense layer classifies the bird species. For the 10 class classifier, the highest accuracy achieved was 90.45\% for a feature set consisting of 13 Mel Frequency Cepstral Coefficients (MFCCs) and 12 Chroma Vectors. The corresponding specificity and AUC scores are 98.94\% and 94.09\%, respectively.

Motivation & Objective

• To address challenges in bird species classification from real-world audio recordings, including overlapping species, background noise, and long-duration recordings.
• To identify the most effective acoustic feature set from 34 candidates for robust and efficient bird species classification.
• To develop a hybrid deep learning framework that integrates temporal and spectral feature learning for improved classification performance.
• To enable real-time classification by optimizing feature selection and network architecture for computational efficiency.

Proposed method

• Voice activity detection is applied to raw audio to isolate bird vocalizations before feature extraction.
• Short-term features are extracted using 50 ms Hamming windows with 25 ms overlap (first-stage windowing).
• Features are then reshaped via non-overlapping second-stage windowing to form 2D input for a 2D Convolutional Neural Network (CNN).
• The 2D CNN learns spatial and spectral patterns from the reshaped feature maps.
• Output features from the CNN are fed into a Long Short-Term Memory (LSTM) network to model temporal dependencies in the audio sequence.
• A final dense layer performs the classification of bird species from the learned representations.

Experimental results

Research questions

• RQ1 Which combination of acoustic features from a set of 34 candidates yields the highest classification accuracy for bird species identification?
• RQ2 How does two-stage windowing (overlapping then non-overlapping) improve feature representation for bird sound classification compared to single-stage approaches?
• RQ3 To what extent does combining 2D CNN and LSTM architectures enhance performance in classifying bird species from complex audio recordings?
• RQ4 Can the proposed method achieve high accuracy and specificity in real-time applications despite background noise and overlapping species?

Key findings

• The highest classification accuracy of 90.45% was achieved using a feature set comprising 13 Mel Frequency Cepstral Coefficients (MFCCs) and 12 Chroma vectors.
• Specificity reached 98.94%, indicating strong performance in correctly identifying non-target species.
• The Area Under the Curve (AUC) score was 94.09%, reflecting robust discriminative capability of the model.
• The two-stage windowing strategy effectively preserved temporal and spectral information, enhancing feature representation for deep learning.
• The hybrid architecture of 2D CNN and LSTM outperformed standard models by capturing both local spectral patterns and long-term temporal dynamics.
• The study provides a validated feature selection pipeline, identifying the optimal 25-feature set (13 MFCCs + 12 Chroma) for real-time bird species classification.