AudioMoth: Evaluation of a smart open acoustic device for monitoring biodiversity and the environment

The cost, usability and power efficiency of available wildlife monitoring equipment currently inhibits full ground‐level coverage of many natural systems. Developments over the last decade in technology, open science, and the sharing economy promise to bring global access to more versatile and more affordable monitoring tools, to improve coverage for conservation researchers and managers. Here we describe the development and proof‐of‐concept of a low‐cost, small‐sized and low‐energy acoustic detector: “AudioMoth.” The device is open‐source and programmable, with diverse applications for recording animal calls or human activity at sample rates of up to 384 kHz. We briefly outline two ongoing real‐world case studies of large‐scale, long‐term monitoring for biodiversity and exploitation of natural resources. These studies demonstrate the potential for AudioMoth to enable a substantial shift away from passive continuous recording by individual devices, towards smart detection by networks of devices flooding large and inaccessible ecosystems. The case studies demonstrate one of the smart capabilities of AudioMoth, to trigger event logging on the basis of classification algorithms that identify specific acoustic events. An algorithm to trigger recordings of the New Forest cicada (Cicadetta montana) demonstrates the potential for AudioMoth to vastly improve the spatial and temporal coverage of surveys for the presence of cryptic animals. An algorithm for logging gunshot events has potential to identify a shotgun blast in tropical rainforest at distances of up to 500 m, extending to 1 km with continuous recording. AudioMoth is more energy efficient than currently available passive acoustic monitoring devices, giving it considerably greater portability and longevity in the field with smaller batteries. At a build cost of ∼US$43 per unit, AudioMoth has potential for varied applications in large‐scale, long‐term acoustic surveys. With continuing developments in smart, energy‐efficient algorithms and diminishing component costs, we are approaching the milestone of local communities being able to afford to remotely monitor their own natural resources.


| INTRODUCTION
Emerging technologies for remote monitoring and species identification bring the promise of more affordable and versatile methods of sampling which are predicted to drive future conservation efforts (Pimm et al., 2015). Current methods often require expensive and complex equipment for aerial imagery (Morgan, Gergel, & Coops, 2010), acoustic sensing (Merchant et al., 2015), bio-telemetry (Cooke et al., 2004), and GPS tracking (Kays, Crofoot, Jetz, & Wikelski, 2015).
The technical know-how and infrastructure needed to implement these devices in large-scale environmental monitoring often requires a total investment beyond the budgets assigned to conservation projects (James, Green, & Paine, 1999). This cost issue is being addressed with an increasingly free availability of online data sources such as satellite images (Kalyvas, Kokkos, & Tzouramanis, 2017). However, such databases cannot capture cryptic biodiversity and exploitation.
For example, events which are hidden by tree cover, or those that are too fine-scale for image resolution, remain unaccounted for without ground-level monitoring (Peres, Barlow, & Laurance, 2006).
Full ground-level monitoring demands many surveyors or devices to cover an ecosystem effectively. Some contemporary sampling methods achieve coverage with semi-automated monitoring technology, for example camera traps triggered by infra-red sensors. Much of the methodology used in acoustic monitoring lags behind this trend, tending to feed large quantities of captured data through detection software after deployment (Mac Aodha et al., 2017). Despite the heavy demand on memory storage, passive acoustic monitoring (PAM) has proved useful for estimating ground-level biodiversity abundance and occurrence, particularly of smaller and more cryptic species (Newson, Bas, Murray, & Gillings, 2017), and it is often employed for analyses of soundscapes . It has also been shown to have potential for monitoring exploitation of natural resources (Astaras, Linder, Wrege, Orume, & Macdonald, 2017). However, PAM devices used for long-term monitoring are limited by their size and weight due to their high power consumption (Wrege, Rowland, Keen, & Shiu, 2017). Moreover, the budget needed to purchase multiple devices and then process the captured data makes them impractical for many research studies or large-scale conservation deployments.
In recent years, researchers have started to look beyond commercially available options to field devices designed and built in partnership with engineers (Kwok, 2017). Research studies have reduced the cost of acoustic monitoring by re-purposing existing technologies (Gross, 2014), or implementing devices based on open-source modular computers with external sensors of fit-for-purpose quality. Recent examples include PAM devices built around the Raspberry Pi computer (Caldas-Morgan, Alvarez-Rosario, & Padovese, 2015;Sankupellay et al., 2016;wa Maina, Muchiri, & Njoroge, 2016) and Arduino computer (Razali et al., 2015;Shafiril, Yusoff, & Yusoff, 2016). For example, the Solo acoustic monitoring platform is based on the Raspberry Pi and an external microphone (Whytock & Christie, 2017), and costs just under ∼US$100. Devices such as Solo are often chosen for their computing power, customisation ability and programming simplicity, using high-level programming languages, such as Python. Despite these advantages, devices based on modular computers present drawbacks for large-scale deployments. They demand considerable investment in time for setting up and configuring each device, involving hobbyist electronics and software development skills. The devices based on the Raspberry Pi have inefficient power optimisation and consequently require large batteries to sustain power over long periods. Solo for example, uses a 12 V car battery to compensate for its low power efficiency in long-term deployments. As with commercial PAM devices, the need for larger battery capacity often makes monitoring tools based on modular computers too bulky for field deployments in remote areas where sensors must be transported manually.
New developments in lightweight commercial detectors are increasing the portability and usability of acoustic devices, such as the Peersonic RPA3 bat recorder costing ∼US$280 (Peersonic pricing page, 2017) or the ARBIMON recorder, which is based on a smartphone, costing ∼US$300 (ARBIMON pricing page, 2017). Despite substantial savings in size and usability, these devices present a high initial cost for largescale studies requiring many devices to cover an area. They also present a lack of customisation compared to the modular computer-based devices. Publications on environmental acoustics to date overwhelmingly report data capture using commercially available, battery-powered, PAM devices. Many acoustic monitoring applications use the commercial Song Meter series from WildLife Acoustics, with best-inclass audio quality at a unit cost of more than US$1000 (Song Meter series product page, 2017).
Here we describe the development and proof-of-concept of a smart, customisable, acoustic monitoring device called AudioMoth (AudioMoth home page, 2017). The device employs a low-power microcontroller and a microelectromechanical systems (MEMS) microphone to perform on-board real-time acoustic analysis, allowing relevant data to be filtered or classified before storage. This smart capability reduces both the storage requirement on the device and the post-processing budget after data collection. With less energy needed to power the device, it can run off smaller batteries. The device addresses the need for a versatile, small, low-cost, low-power monitoring tool for easy deployment in long-term biodiversity and environmental acoustic monitoring. AudioMoth aims to make a substantial step towards the future of acoustic technology, in covering large areas of inhospitable habitats with a network of devices (Browning, Gibb, Glover-Kapfer, & Jones, 2017).

| MATERIALS AND METHODS
Here we describe the design of the AudioMoth hardware and its customisable software, in the context of two ongoing monitoring studies aiming to achieve large-scale and long-term coverage with a large number of smart devices. The first study tests for presence of the New Forest cicada (Cicadetta montana Scopoli, 1772), an elusive species last sighted in the UK over 22 years ago (Pinchen & Ward, 2002, p. 134). The second study investigates the detection of gunshot events within tropical forests in Belize, Central America, in an area under pressure from poaching.

| Design
AudioMoth is built around an ARM Cortex-M4F microcontroller. The M-series processors are some of the most energy efficient microcontroller cores currently available. The M4F core used by AudioMoth has on-board floating-point signal processing functionality, allowing efficient processing of acoustic data at high speeds. AudioMoth can process data at sample rates up to 384 kHz in real-time, made possible by an additional 256-KB SRAM chip, which increases the amount of available processing time at ultrasonic frequencies. AudioMoth stores uncompressed WAV files to microSD card, with a capacity limit of 32 GB. AudioMoth can accommodate extensions to the board, such as external sensors or a wireless network unit, using a 6-pin peripheral module interface (PMOD) header that connects four general purpose input/output pins to the processor.
AudioMoth can easily be deployed as a scheduled recorder, without any requirement to code or to learn a computer programming language; however, to unlock the main advance that the device brings, users are encouraged to customise and design their own on-board software for filtering or classifying sounds as they happen. The user can modify and distribute AudioMoth's software for specific applica- and data storage requirements. In conjunction with its low power consumption while listening, and its full-spectrum frequency response, AudioMoth creates a unique opportunity for users to design specific classification algorithms for individual projects. In order to realise its performance capabilities, however, AudioMoth employs the low-level programming language, C, which requires a greater level of technical expertise than the less efficient high-level programming language, Python. Despite this constraint, AudioMoth achieves ultra-efficient power optimisation, high-speed data processing and a wide spectrum acoustic performance through the greater control over low-level processes.
AudioMoth uses the Goertzel filter for real-time classification algorithms. This filter evaluates specific terms of a fast Fourier transform on temporarily buffered audio samples without the computational expense of a complete transform. The outcome of each algorithm is used to trigger recordings to a microSD card.
To apply a Goertzel filter to an audio recording, the samples are split into N windows of length L given by: (s 1,1 , …, s 1,j , …, s 1,L , …, s N,1 , …,

| Purchase and configuration
Open-source, custom-designed and simply constructed hardware provides cost-effective access to technology for all. Simply constructed hardware can be bought at close to component prices, because the number of fabrication steps can be minimised. To enable simple construction of the device, its parts must be readily available and simple to fit together. Accordingly, the circuitry for AudioMoth uses online accessible components, which all fit on one side of a two-layer printed circuit board (PCB). This permits the acquisition of devices from a single PCB assembler. Such simply-constructed hardware can be manufac- LCD screen or manual assembly. Experience from users of devices such as Mataki (Mataki product page, 2017) suggests that conservation technology puts a premium on ease of configuration in the field.
Accordingly, our open-source code includes a cross-platform configuration application, which can configure device settings in the field using a laptop and a USB cable. The configuration application is built on a free, open-source framework called Electron (Electron, 2017).
Together with the configuration application, AudioMoth's default firmware enables the device to be used as a scheduled recorder. The configuration application features adjustable recording schedules, gain levels and sample rates. Using this configuration application makes it easier to configure large quantities of AudioMoths as PAM devices for multiple large-scale applications.

| Deployment
Portability in the field requires minimising the size and weight of devices, so as to maximise the number that can be carried on foot AudioMoth is designed to make efficient use of its available storage by on-board real-time audio processing. The energy consumption while processing or during a calculation is negligible, consuming from 10 to 25 mW between the lowest and highest sample rates.
Deployments in locations likely to trigger large quantities of recordings will more likely be limited by the microSD card capacity than the battery life.

| New Forest cicada
The

| Gunshots in tropical forests
The second study aimed to test the detection range of AudioMoth The wide array of false-positive sources and variation in gunshot amplitude due to factors such as distance and topography means that an algorithm for detecting gunshots must accommodate various components of the acoustic pulse, such as the initial muzzle blast, and the various stages during its sound propagation. When recorded at close range, the initial muzzle blast consists of a loud impulse covering a wide range of frequencies. As the sound propagates from the gunshot to the detector, the high frequency components start to decay as they are absorbed into the air and the surrounding environment. The gunshot detection algorithm for AudioMoth used the characteristic rate at which select frequencies peak and then decay from the initial muzzle blast, determined by ground-truthing trials in the forest.
To characterise the gunshot features, we developed a four state hidden Markov model and used the Viterbi algorithm (Forney, 1973)

| Detection capabilities
For the first case study, the detection capabilities of the device were For the second case study, gunshot amplitude diminished with distance from the sound source as expected, and was affected further by the orientations of the device and the gun. The rate of decline with distance in gunshot amplitude increased substantially when the device faced away from the source, and it increased slightly when the gun was facing away from the device ( Figure 5). In earlier pilot trials in the same area during 2016, the probability of an audible signal in continuous recordings was 98% at ≤300 m, declining to 93% at ≤1 km, from a total of 120 gunshots. The captured data were run through the detection algorithm after deployment, which identified gunshots at up to 500 m with a success rate of 66%, decreasing to 50% at 1 km. At this furthest distance, devices facing towards the gunshot were 80% more likely to detect it than devices facing away. In the available 1-s time interval for processing each buffer of samples, the algorithm took just 40 ms to run, using just 4% of the available processing time. Future iterations of the algorithm will make use of the remaining 960 ms of computational time to take into account variations in acoustic structure due to orientations of the device and sound source.

| Deployment logistics, field configuration and durability
A single AudioMoth has a build cost of ∼US$43 (Table S1). Group purchasing brings the price down to ∼US$30, on an order of 500 devices delivered assembled, pre-programmed and ready for deployment. This economy of scale is particularly relevant to large-scale monitoring, for example of forest exploitation where many devices are needed to cover large tracts of protected forest. With a gunshot detection distance of up to 500 m, a single AudioMoth would monitor an area of ∼0.8km 2 .
Thirty AudioMoths, bought with group purchasing, would have a total cost of ∼US$900, and the capability for monitoring an area of ∼24 km 2 .
The low cost of AudioMoth allows researchers to deploy more devices with their budgets, allowing them to ask bigger research questions.
With AudioMoth's small dimensions when using AA-cell batteries, more than 100 devices can fit into a standard field backpack with a 25-L capacity. For deployments in rough terrain, such as areas of tropical forest, this ability to carry multiple devices greatly facilitates field deployments and reduces the infrastructure needed to achieve a large-scale study.
F I G U R E 5 Ratio of gunshot peak amplitudes relative to the maximum possible amplitude, from continuously recording AudioMoths. (a) Devices facing towards the gunshot source, demonstrating a higher performance of audio capturing ability; (b) devices facing away from the gunshot source, demonstrating a lower performance of audio capturing ability During the preparation stage in the first case study, the user required on average 10 s to configure each device to a pre-set configuration over USB. Eighty-seven devices were configured in total, taking under 20 min to prepare them all for a deployable state. Further possibilities exist to program multiple devices simultaneously by audio signals played through a computer speaker (Jewell, Costanza, & Kittley-Davies, 2015). This could bring substantial time savings for large-scale deployments.
During the New Forest deployment, 23% of devices suffered some level of water damage, due to heavy rain and failure of the grip sealed bag. The rainforest deployment trialled a commercial waterproof electronics enclosure, combined with an acoustic waterproof permeable membrane and silica gel sachets to absorb moisture. All these additional parts cost a total of ∼US$8 (Table S1). This combination provided an effective protection for the duration of the study, with the membrane allowing sound penetration to the microphone.

| Data storage and energy consumption
In the first case study, the 5-month total period of field deployment Power consumption reduced to 25 mW or less when classifying audio in real time, and to 80 μW when sleeping between samples, outside scheduled wake-up periods, or in standby mode. The low demand in standby would allow AudioMoth to keep track of time for approximately 6 years in that mode, using three AA-cell lithium batteries. In the first case study, the devices remained powered for the total 2-and 3-month periods they were deployed for.
The complexity of the processor for an intelligent device directly affects its power consumption, which in turn affects the size and weight requirements of its power supply. A modular computer-based PAM device such as Solo continuously runs a Linux operating system during operation (Whytock & Christie, 2017). These devices consume anywhere from 400 mW to 1,000 mW when idle ( Figure   6), with minimal to no power management available during operation. In contrast, AudioMoth's processing comes from an ultra-low energy microcontroller, which has complete control over its power management, meaning it can run embedded code fast enough to power down and sleep between individual microphone samples.
Even during recording to microSD card at 48 kHz, AudioMoth is ∼15 times more energy efficient than the most energy efficient modular computer-based PAM device, and ∼4,000 times more energy efficient during its idle state. Further developments are exploring the potential for networking AudioMoth by LoRa radio, to link them to a base station for real-time signalling of acoustic events triggered by the detection algorithm.

| WIDER APPLICATIONS
Although this capability adds ∼US$30 to price, the devices are sufficiently cheap to make it a potentially cost-effective option for capacity building. AudioMoth also has the ability to record alternative types of data to memory, instead of memory inefficient uncompressed WAV files. For example, AudioMoth can summarise the important characteristics of sounds with measurements known as acoustic indices (Towsey, Wimmer, Williamson, & Roe, 2014). Acoustic indices can summarise recordings into meaningful characteristics, such as the frequency distribution and acoustic power, which can be viewed as false colour images to aid the assessment of biodiversity (Sueur, Farina, Gasc, Pieretti, & Pavoine, 2014). As these indices require less space to store than raw audio, devices using them are less constrained by limited storage capacities. In addition, significant energy benefits accrue from writing small summary files to the SD card, rather than raw audio files. We are currently developing real-time acoustic indices that make use of the fast processing available on-board the AudioMoth hardware. Future work will continue to test the feasibility of deployments by drone (Project Erebus AudioMoth flight test page, 2017), again only possible for small and cheap devices.
While the configuration software enables a basic level of device customisation with minimal technical expertise, knowledge in programming low-level C is required to achieve full use of AudioMoth's flexibility and produce new detection algorithm implementations.
Bringing technology such as AudioMoth to less technically skilled users remains an ongoing challenge in the area of conservation technology. We see possibilities for progress in the future with solutions such as compilers that generate C code from simplified implementations of digital signal processing techniques.

| CONCLUSION
The purchasing opportunities available for simply designed, opensource and configurable hardware can dramatically reduce the financial cost and time commitment required for environmental monitoring on large spatial and temporal scales. Monitoring projects can address bigger questions with access to smart, small and power-efficient devices such as AudioMoth. We are now close to being able to flood large areas with these devices, for improved coverage of obscured, remote or inhospitable ecosystems. High initial investment costs remain the biggest barrier for conservation projects in poorer areas.
AudioMoth provides opportunities for groups with limited budgets to perform systematic bioacoustics research, for example by benefiting from economies of scale in group purchases. With further developments in the new technologies described here, we are getting closer to achieving a basic requirement of sustainable development, that local communities can afford to monitor their own natural resources.