Abstract
Segregating an audio mixture containing multiple simultaneous bird sounds is a challenging task. However, birdsong often contains rapid pitch modulations, and these modulations carry information which may be of use in automatic recognition. In this paper we demonstrate that an improved spectrogram representation, based on the distribution derivative method, leads to improved performance of a segregation algorithm which uses a Markov renewal process model to track vocalisation patterns consisting of singing and silences.
Introduction: The Avian Cocktail Party Problem
In natural environments, birds rarely sing in isolation. Bioacoustic recordings are frequently "polyphonic," featuring overlapping songs from multiple individuals. For researchers monitoring populations or studying individual behavior, this presents a significant challenge: accurately assigning specific vocal "syllables" to the correct source.
Most automated recognition tools are designed for monophonic audio. Previous attempts to solve the polyphonic problem using probabilistic tracking have been limited by the quality of initial detection. Standard spectral analysis, such as the Short-Time Fourier Transform (STFT), assumes signal stability over short windows. However, birdsong is characterized by "chirps"—extremely rapid frequency modulations (FM) that often appear blurred in traditional representations. This research addresses this gap by enhancing both the representation of fast-changing sounds and the clustering of these sounds into distinct "streams."
Methodology
The proposed system integrates two advanced mathematical frameworks to improve tracking accuracy:
1. Distribution Derivative Method (DDM)
To better capture non-stationary signals, the authors replace standard spectrograms with a DDM-based representation. While rapid FM causes blurring in STFT, DDM utilizes the "distribution derivative rule" to infer fine details of modulated sinusoids across spectral bins. This allows for a more accurate estimation of the "slope" of a chirp.
2. Markov Renewal Process (MRP) Clustering
Detected syllables are linked using an MRP model. Unlike standard Hidden Markov Models, an MRP explicitly accounts for the duration between events. This is particularly suited for birdsong, where the rhythmic timing of syllables and the silences between them serve as key identifying features.
3. Enhanced State Representation
The authors further refined the tracking by moving from a single-value frequency offset to a "vector-valued" state—a 5-point snapshot of a syllable's frequency evolution. This provides the tracking algorithm with more "texture" to distinguish between similar vocalizations.
Experimental Setup
The study utilized recordings of the Common Chiffchaff (Phylloscopus collybita), a species known for its rhythmic, repetitive song, sourced from the Xeno Canto database. Solo recordings were used for training, while synthetic mixtures (ranging from 1 to 5 simultaneous birds) were created to evaluate the algorithm's ability to segregate overlapping voices.
Key Results and Discussion
The system was evaluated using the F-measure for two primary tasks: signal-noise separation ($F_{SN}$) and the recovery of transitions between syllables ($F_{trans}$).
The results demonstrate that the combination of DDM spectrograms and FM information leads to an approximate 20 percentage point improvement in the F-measure compared to baseline methods. DDM consistently outperformed the standard STFT across all mixture densities. Furthermore, providing the tracker with a 5-point vector representation significantly improved signal-noise separation, confirming that high-level tracking is heavily dependent on the quality of low-level feature encoding.
Conclusion and Future Work
This research highlights that rapid pitch changes, often viewed as noise or a challenge, actually carry vital information for source segregation. By explicitly modeling frequency slope and temporal rhythm, the system effectively untangles complex avian audio.
Future work will focus on reducing the computational complexity of DDM to enable large-scale bioacoustic analysis. Additionally, the authors plan to test the robustness of this approach on species with more irregular singing patterns and in real-world environments with high levels of ambient noise.
Summary updated