Automatic rain and cicada chorus filtering of bird acoustic data
Introduction
Recently, there has been research interest in designing processes to detect and monitor animal species using unattended environmental recordings [1]. A key challenge in achieving this goal is interference from noise which can mask vocalisations of interest [2], [3]. This makes vocalisations more difficult or impossible to detect, and can result in wasting resources on examining audio that cannot be feasibly analysed.
While some work has gone into filtering stationary (i.e. constant) background noise for both speech [4], [5] and bioacoustics processing [6], [7], there are many sounds, such as rain, wind, car sounds, and animals that are not of interest, that cannot be removed using these filters. This work examines and evaluates techniques to filter noise from rain and cicada choruses for use in a theoretical bird sound analysis. These sources are common in the environmental recordings being analysed, interfere significantly with signals of interest, and have distinct characteristics that can help in detecting them [8], [9].
There has been limited research into removing these noise sources, and the research that has been done has focused on detecting these sources, though not for the purposes of removal [8], [9], [10]. Only the presence or raw intensity of noise from some sources of environmental noise, such as rain, have been considered rather than the extent of which noise interferes with any signal in the recording. For example, light rain might interfere significantly with a quiet animal call, but a loud animal might be very clear even in the presence of heavy rain. Furthermore, existing research does not consider that users might desire to filter noise sources with different sensitivities. For example, a researcher might only want to keep the cleanest samples possible, but another researcher might want to keep everything that might contain a vocalisation of interest.
To address current limitations, two filters are proposed: one for cicada choruses and another for rain. These filters utilise multiple acoustic indices in combination with Mel Frequency Cepstral Coefficients (MFCCs). The most effective filtering configuration is determined by evaluating the ability of combinations of acoustic features, classifiers, and other filters to detect rain and cicada choruses, using Area Under the Receiver Operating Characteristic Curve (AUC) as the primary metric. Feature sets, classifiers, classifier hyperparameters, and the effect of other filters are compared on a level that is deeper than previous work [8], [9] and this results in more accurate classification and filtering. While samples classified as containing rain are removed, a second step is introduced to filter cicada choruses which removes only the frequency range containing the choruses.
The filters are designed to work with thresholds based on the probabilities of samples containing the noise source of interest (i.e. rain or cicada chorus). This allows users to determine the sensitivities of the filters. They are also trained to classify samples based on how much they interfere with sounds of interest, which are more suited for filtering than intensity-based classifiers in previous work.
In the next section, current works on filtering rain and cicada sounds and their limitations are discussed. A methodology for cicada chorus and rain detection is then presented in Section 3. Then, in Section 4, the results of the rain and cicada detection are presented and discussed. In Section 5, a filter to remove cicada choruses from environmental recordings is proposed and evaluated for its effectiveness. Finally, this work is concluded and future directions are proposed in Section 6.
Section snippets
Related work
Current work into processing rain and cicada sounds has been limited to detection approaches. These existing approaches are discussed, particularly in terms of how they could be improved upon to detect more accurately and to remove these sounds.
Methodology for targeted sound detection
The approach for filtering rain and cicada choruses used here employs machine learning, much like previous research. The selection of classifier, feature set, and additional pre-processing tasks play a significant role in the classification accuracy. As such, many combinations of feature sets, classifiers, and filters are tested using a training dataset to determine the best rain and cicada chorus filters.
Rain classification
Results from the testing for the best rain classification configuration are shown and discussed, before being applied to a larger dataset. Existing approaches are then compared against the new approach.
Cicada Chorus Filter
Cicada choruses are typically very loud and occupy a narrow frequency band. While a regular noise filter, such as the MMSE STSA filter, will reduce the cicada noise somewhat, it is not aggressive enough to remove the cicada sound completely because, while a cicada chorus sounds stationary when listening, it is actually non-stationary in reality. This is shown in Fig. 9 which only looks at the frequency region containing cicadas, showing that intensity does indeed fluctuate.
A more aggressive
Conclusions and future directions
In this work, the use of acoustic indices with MFCCs in combination of classification algorithms was investigated for reducing noise from non-stationary sources (rain and cicada choruses) in bioacoustic recordings. Many configurations were evaluated with different acoustic features, classifiers, and filters to detect cicada choruses and rain in bioacoustic recordings. Using a MLP classifier and applying a high-pass filter, an AUC of 0.9911 was achieved. This classifier can have thresholds
Acknowledgement
We thank the Samford Ecological Research Facility (SERF) for providing us with environmental recordings used as sample data for this work. The lead author is supported by the Australian Government’s Research Training Program (RTP) .
Source code
The source code used is available at sourceforge.net/projects/rain-and-cicada-noise-removal/.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (44)
- et al.
Automatic anuran identification using noise removal and audio activity detection
Expert Syst. Appl.
(2017) - et al.
The use of acoustic indices to determine avian species richness in audio-recordings of the environment
Ecol. Inform.
(2014) - et al.
Automatic identification of rainfall in acoustic recordings
Ecol. Inform.
(2017) - et al.
A new methodology to infer the singing activity of an avian community: the acoustic complexity index (ACI)
Ecol. Inform.
(2011) - et al.
Cross-validation for selecting a model selection procedure
J. Econometrics
(2015) - et al.
Automatic bird sound detection in long real-field recordings: Applications and tools
Appl. Acoust.
(2014) Computational Bioacoustics: Biodiversity Monitoring and Assessment
(2017)- et al.
Bird detection in audio: a survey and a challenge
Suppression of acoustic noise in speech using spectral subtraction
IEEE Trans. Acoust. Speech Signal Process.
(1979)- et al.
Speech enhancement using a minimum-mean square error short-time spectral amplitude estimator
IEEE Trans. Acoust. Speech Signal Process.
(1984)