Soundscapes contain rich acoustic information associated with animal behaviours, environmental characteristics and human activities, providing opportunities for predicting biodiversity changes and associated drivers. However, assessing the diversity of animal vocalizations remains challenging due to the interference of environmental and anthropogenic noise. A tool for separating sound sources and delineating changes in acoustic signals is crucial for an effective assessment of acoustic diversity.
We present soundscape_IR, an open‐source Python toolbox dedicated to soundscape information retrieval in which nonnegative matrix factorization is applied. This toolbox provides algorithms for supervised and unsupervised source separation (SS). It also enables the use of a snapshot recording for model training and subsequently applying adaptive and semi‐supervised SS when target species produce sounds with varying features and when unseen sound sources are encountered.
Our results demonstrated that SS could enhance the vocalizations of target species, characterize the complexity of vocal repertoires and investigate the spatio‐temporal divergence of soundscapes. In tropical forest soundscapes, the application of SS effectively detected the rutting vocalizations of sika deer and revealed a graded structure in their acoustic characteristics. In subtropical estuarine soundscapes, SS automated the process of identifying distinct biotic and abiotic sounds, and the result uncovered divergent sound compositions between inshore and offshore waters.
Implementation of SS in soundscape analysis offers a promising method for streamlining the assessment of acoustic diversity in diverse environments. Future application of SS will open new directions to acoustically quantify ecological interactions across individual, species and ecosystem levels.
聲景中充滿了與動物行為、 環境特徵和人類活動相關的豐富訊息, 提供一個絕佳機會, 讓人們得以透過聲音預測生物多樣性的變化趨勢以及相關的驅動機制。 然而, 環境聲音和人為噪音的干擾導致動物聲音多樣性的分析極富挑戰性。 因此,一個能夠分離各種聲音並釐清音訊變化的分析工具, 對於分析聲音多樣性將有至關重要的幫助。
soundscape_IR 是一個 Python 開源工具箱, 運用非負矩陣分解法擷取聲景中的各種訊息, 其提供了基於監督式學習和非監督式學習的聲源分離演算法, 使用者也能夠透過一小段錄音訓練模型, 在後續的分析中, 運用適應性和半監督式聲源分離找出目標物種不同特徵的聲音和其他未曾遇過的聲音。
研究結果顯示, 聲源分離演算法可以在吵雜的錄音中強化目標物種的聲音訊號, 解析動物聲音曲譜的複雜性, 辨別聲景的時空變化趨勢。 在熱帶森林的聲景錄音中, 此工具可有效偵測梅花鹿的發情叫聲, 並揭露其聲學特徵的階層性結構。 在亞熱帶河口的聲景錄音中, 此工具可以自動存別出各種生物和非生物聲音, 揭露近岸和離岸水域在聲音組成的差異。
本研究開發的聲源分離演算法與分析工具提供一個極具前景的聲景研究方法, 能夠簡化聲音多樣性的評估過程, 未來可望有效協助研究者透過聲音訊息量化個體、 物種和生態系等各種層級的交互作用。