Bird recognition by song using AI and without a network connection

Android Guides » Android Apps » Bird recognition by song using AI and without a network connection

Hearing a bird's trill and wondering what species it is is a very common scene for anyone who spends time outdoors. Until very recently, answering that curiosity required a good ear, a lot of experience, and almost always, a bird guide at hand.Today the landscape has completely changed: artificial intelligence (AI) is able to analyze recordings and tell you, in a matter of seconds, which birds are singing around you.

What's really interesting is that this technology is no longer limited to laboratories or research centers. Applications like BirdNET or Merlin Bird ID put AI models trained with hundreds of thousands of recordings and millions of observations on your mobile phoneAnd they also increasingly offer options to work without a data connection, which is key if you go out into the countryside, away from coverage, or simply want to save battery and data.

How computers learn to recognize birds by their song

When we talk about bird recognition by sound, we are not dealing with a simple "Shazam for birds". Projects like BirdNET are based on deep neural networks that have been trained on labeled recordings of more than 3.000 species on a global scaleThese networks learn to distinguish sound patterns characteristic of each species, even when there is ambient noise or several individuals singing at the same time.

The process does not consist of the computer "listening" as a person would. The key is to transform the audio into spectrograms, that is, images that show how the frequency and intensity of the sound varies over time.These images are analyzed using techniques very similar to those used to recognize objects in photographs, except that what is identified here are specific acoustic signatures of each species.

The power of these models comes from the amount of data they are fed. The Cornell Lab of Ornithology and the Macaulay Library have collected tens of thousands of sound recordings thanks to citizen science networks like eBird.where amateurs and experts upload recordings and note which species appear in them. This combination of audio and validated observations is pure gold for training robust algorithms.

In the case of Merlin Bird ID, the approach also involves spectrograms, but with an interesting nuance: The Cornell team leveraged their previous experience training computer vision models to identify birds in photosInstead of telling the AI ​​"listen to this sound", it tells it "look at this sound image" and recognizes visual patterns associated with each song or call.

All of this is complemented by contextual information. The AI ​​doesn't just look at how the singing sounds, but also at where and when it was recorded.Thanks to eBird's more than one billion observations, models know which species are most likely to be found in a specific place and time, helping to rule out unlikely options even if the sound is similar.

BirdNET: Collaborative AI to listen to the forest

BirdNET is one of the benchmarks when we talk about AI-powered automatic bird song recognitionDeveloped jointly by the K. Lisa Yang Center for Conservation Bioacoustics at the Cornell Lab of Ornithology and Chemnitz University of Technology, its goal is twofold: to help people identify birds and, at the same time, collect valuable data for research and conservation.

The application allows you to use your Android device's microphone to record the environment and analyze in near real-time which species are most likely. Currently, BirdNET is able to identify more than 3.000 common species spread throughout the world, and its model continues to grow as new recordings and annotations are incorporated.

BirdNET's success among the bird-keeping community has been resounding. Since its launch in February 2021, the app quickly surpassed one million downloads on Google PlayIt has become an everyday tool for birdwatchers of all levels. This is partly due to the fact that it is free and available for both Android and iOS.

In addition to providing identification results, BirdNET is designed as a collaborative scientific project. When you submit your recordings, you can contribute to a global database of acoustic observations. This feeds into ecological, behavioral, and species distribution studies. This citizen participation allows for the coverage of areas and times that would be impossible to sample using only professional equipment.

The BirdNET model has also been partially released for research and developer use. It is possible to download the model from GitHub and adapt it to more specific bioacoustics projectsHowever, this requires programming skills and audio data handling experience. This opens the door to customized applications for monitoring biodiversity in nature reserves, urban parks, or forestry operations.

How BirdNET works step by step

BirdNET is quite easy to use, designed so that anyone can go out into the field and start using it without too much trouble. When you open the app for the first time, it will ask for location access in order to calculate the species that are likely to be found in your area., something crucial for refining predictions.

The main screen displays a real-time spectrogram of the sound being picked up by the microphone. Below you'll find a button to start recording and another to pause it.The usual procedure is to wait until the bird sings clearly and then stop the recording to move on to the analysis phase.

Once the audio has been recorded, the next step is to mark the fragment you want to analyze. You slide your finger over the spectrogram to select the portion of time where the singing is best heard.When you release the button, two main options will appear: Analyze or Save. Analyze sends the clip to BirdNET's servers, while Save allows you to keep it for later processing.

The save option is especially useful when you don't have coverage. If you're in a remote area without an internet connection, you can store recordings and, once you have data or Wi-Fi again, start the analysis.This hybrid online/offline workflow makes BirdNET practical even in mountainous or jungle areas, although the actual recognition is performed in the cloud.

When you click on analyze, the application takes a few seconds to process the fragment. If it detects a species, it will display its name and an indicator of confidence in the identification.In recordings where several birds sing, BirdNET is able to list several detected species, which is great for forest or wetland scenes with a lot of sound movement.

The overall design is quite clean. The interface focuses on the spectrogram and basic recording controls, without overwhelming the user with technical data.However, the app does hide a lot of advanced options in its menu for those who want to fine-tune things further.

BirdNET main menu and advanced options

Accessing the menu (the typical three horizontal lines in the upper left corner) reveals a set of sections and settings that allow you to customize the experience. One of the most useful is “Show observations”, where you can review all the analyzed recordings along with their date, location and result.

From each observation you can share the original audio. Simply tap the corresponding entry and then tap the share button to save the file outside the app or send it via messaging or email.This is very practical for documenting field trips, compiling personal lists, or sharing findings with other observers.

Another interesting section is “Explore your area”. Here BirdNET shows you a list of species that you are likely to find in your area based on your GPS location and data from eBird.orgIn a specific example, for the Farallones del Citará region (Antioquia, Colombia), the app offers up to 323 detectable species, which gives an idea of ​​the potential in megadiverse regions.

In terms of customization, BirdNET incorporates several key options. It is possible to choose the language for the common names of birds from more than 20 available.This is very useful if you want to see the names in Spanish but also check the name in English or in the local language of another region.

The selection of likely species can be adjusted according to temporal criteria. BirdNET allows you to choose between a weekly and an annual mode to filter which species are considered common at each time of year.In this way, the model takes into account migrations and seasonal changes in the presence of birds.

There is also a detection sensitivity control. With high sensitivity, the algorithm becomes more "bold" and detects more things, but at the cost of increasing the chances of errors.With low sensitivity, it will be more conservative and will only report particularly clear identifications. Adjusting this parameter depends on the environment and how you intend to use the app.

Depending on the phone, BirdNET lets you choose the audio input source. You can select different channels, such as the standard microphone, RAW (unprocessed) audio, or even the voice recognition channel.The best option, when available, is to use unprocessed audio to prevent the operating system from filtering background noise and distorting the singing.

For those who enjoy interpreting spectrograms, there are additional aesthetic and technical options. The app allows you to change the color palette of the spectrogram (Viridis, Plasma, Magma, Parula, Jet, and Gray), adjust the visible length between 10 and 30 seconds, and modify the amplitude gain between -20 and 20 dB.A high value will make faint sounds appear and be heard better, but if you go too far it can cause automatic cuts in the recording.

It is also possible to set the maximum visible frequency, between 3 and 18 kHz. A wide range helps to visualize species with very high-pitched songs, while a narrower one can focus on the vocal range of most common birds.The contrast of the spectrogram, for its part, serves to highlight color differences and facilitate the reading of sound patterns.

In addition to all this, the menu includes access to an initial tutorial, the privacy and licensing policy, a section to claim authorship of submitted observations (so that your name can be associated with the records) and an "About us" section with project information. It's a fairly comprehensive set of tools, but integrated in a way that doesn't hinder basic use..

Actual performance of BirdNET with Colombian birds

Testing BirdNET in high-diversity regions, such as the western Andes of Colombia, is a good way to test its limits. Field experience shows that the application runs quite smoothly and takes up little space on the devicewhich is appreciated on phones with limited storage.

In terms of accuracy, the algorithm performs very well with frequent species and defined songs. Accurately identifies common birds such as the yellow-breasted (Myiozetetes cayanensis) or the mountain oriole (Icterus chrysater)as well as very characteristic calls such as those of the common spoonbill (Todirostrum cinereum) or the collared trogon (Trogon collaris).

Things get more complicated when very similar songs come into play between different species. Some groups of tanagers with simple and similar calls may cause confusion in the appThis makes sense if we consider that even for an experienced ornithologist, certain notes are almost indistinguishable without a good context.

Another sensitive point is the species with a very wide vocal variation. Wrens of the genus Henicorhina, for example, exhibit highly variable repertoires that continue to pose headaches for the modelAlthough BirdNET covers many species, it still does not reach all regional variations and song dialects.

It should be emphasized that the list of species that the app can recognize is not exhaustive worldwide. There are birds that have not yet been incorporated into the model, or for which there is hardly enough acoustic data.It is expected that these gaps will be filled in future updates, especially thanks to global database initiatives for singing.

In day-to-day use, BirdNET is perceived as a very useful tool to complement visual observation. Its ability to show several detected species in a single recording helps to understand what is happening in a forest saturated with vocalizationsAnd the fact that you can then calmly review the observations also makes it a good learning resource.

BirdNET's errors, bugs, and strengths

Like any complex app, BirdNET is not immune to occasional glitches. Occasionally, when you press the "Analyze" button, a message appears indicating problems with the servers.These are usually temporary connection or saturation errors, which are resolved after a while or by repeating the attempt.

Another bug observed in the field is the sudden change of location. In some cases, the app has jumped to a very distant location (for example, Canada), which completely alters the likely species it can detect.Fortunately, this behavior usually corrects itself within a few minutes when the GPS recalculates accurately.

Beyond these details, one of BirdNET's great advantages is the ease with which it allows you to select and trim the audio fragment that interests you. The intuitive gesture on the spectrogram and the system's quick response make the process quite convenient.This is something that's appreciated when you're on the road and don't want to waste time with complicated menus.

The overall design of the app is another highly valued aspect. Its simple appearance, with few visual distractions, helps to focus on what's important: recording and analyzing singing.At the same time, the ability to customize the spectrogram and microphone settings offers scope for advanced users.

Overall, BirdNET is perceived as a tool that, while not perfect, offers a very balanced relationship between power, simplicity, and usefulness in the field. For someone starting out in birdwatching, it can be the final push to get hooked on recognizing species by ear.For more advanced users, it becomes an interesting complement for documenting and reviewing records.

Suggestions and potential of BirdNET for offline use

Among the ideas put forward by advanced users, one request stands out as very frequent: to be able to download custom models for specific geographic areas, similar to how Merlin does with its regional packagesThis would allow the use of recognition systems adapted to a country or region without the need for a continuous internet connection.

Another desirable improvement would be the ability to upload pre-recorded audio files, manually choosing the location of each recording. There is currently a related web tool, but it does not allow you to change the recording location.which limits its usefulness if you want to evaluate songs captured in different places.

It is clear that the BirdNET development team has room to add features in future versions. Today, it is already possible to install and run the model in environments like Google Colab to analyze previously recorded audio.This opens up a range of possibilities for researchers and enthusiasts with some technical skill.

Specifically, the implementation of BirdNET in Google Colab has made it possible to process collections of old sounds and compare them with the model's outputs. These types of semi-offline workflows, where analysis is done in the cloud but in a controlled manner, are a kind of intermediate step towards more autonomous solutions., such as field recording devices that run the model locally.

If the downloading of models for offline use becomes widespread, we could soon see autonomous monitoring stations powered by solar energy that record and identify birds for months. These scenarios fit very well with the objective of strengthening biodiversity monitoring without depending so much on intensive on-site campaigns..

Merlin Bird ID: Bird identification by sound, photo, and description

Merlin Bird ID, also developed by the Cornell Lab of Ornithology, is another key tool within the AI-powered bird recognition ecosystem. It was created primarily to help beginners and intermediate users identify species from the United States and Canada.But over time it has expanded worldwide and added features.

In the sound field, Merlin is able to identify more than 400 species by their song in the United States and Canada, with the promise of incorporating more regions and birds. This capability is in addition to the already available photo recognition and guided description functions., which brings the total to approximately 7.500 identifiable species on a global scale.

The internal workings of sound recognition in Merlin are also based on artificial intelligence and the work of thousands of citizen scientists. The recordings uploaded to the Macaulay Library and the observations from eBird serve as the basis for training the model, in a very similar way to what happens with BirdNET.

One relevant difference is how the user interface has been designed. Merlin listens with you in real time and displays a list of birds it thinks are singing or calling.with photos and names that update as the recording progresses. It's like having an expert guide whispering in your ear who's performing at the concert.

After recording, you can select a specific species and jump directly to the point in the recording where its song was heard. The recordings are saved automatically, allowing you to listen to them multiple times and use them as learning material.This helps a lot in training the ear, by comparing what you hear with the name and image of the bird.

AI, spectrograms, and machine learning within Merlin

Merlin's technical approach to sound puts an interesting spin on the identification problem. Instead of trying to "understand" the audio directly as a wave, it transforms each recording into a spectrogram and applies computer vision methods., similar to those he uses to identify birds in photographs.

Thus, the model learns to associate visual patterns in those sound images with specific species. This allows it to recognize individual songs even when they overlap with others, something very common in nature.The key is that each species leaves a characteristic "footprint" in terms of frequency, duration, and rhythm, which the neural network eventually distinguishes.

The Merlin project not only focuses on identification, but also seeks to promote user learning. With a single tap you can access detailed fact sheets for each species, with identification tips, distribution maps and a huge library of photos and sounds (more than 80.000 resources)It is, in practice, an interactive bird guide with AI capabilities.

Another very useful function is the search by name or by family through “Browse all birds”. This allows you to explore species even without a recording, ideal for studying before a trip or for resolving doubts afterward.All of this is available for free on iOS and Android, with bird names translated into Spanish in the Android version.

An important point for those who spend a lot of time outdoors is that Merlin is designed to work offline as well. The app allows you to download regional data packs with sounds, photos, and basic models.so that you can continue to identify birds even if you lose coverage, which is very common in natural environments.

Image identification and offline use in Merlin

In addition to sound, Merlin stands out for its image recognition tool. From version 1.2 onwards, Bird Photo ID can be downloaded directly in the app, which until then was only available on the web.This function analyzes a photo of a bird and suggests which species are the best fit.

The process is very simple. Simply frame the bird within the frame that appears on the screen and let Merlin do the rest.The model compares the silhouette, colors, and other features with its huge database of labeled images, and offers you a list of possible matches with varying degrees of probability.

The great advantage is that all of this can work offline, as long as you have downloaded the necessary packages. This makes Merlin an all-rounder: it can identify birds by song in real time, by a photo taken at the moment, or even by a simple guided descriptionAnd all this in areas where there is no trace of mobile coverage.

In the context of recognizing birds by their song offline, Merlin offers a mixed approach. While some advanced capabilities may require a connection, the ability to download regional models and databases means that much of the work can be done locally.This reduces dependence on the cloud and lays the groundwork for almost entirely offline use in the medium term.

Combined with its free nature and educational approach, Merlin Bird ID has become a fantastic gateway to ornithology for thousands of people. The feeling of pointing your phone at the trees and seeing the names and photos of the birds you're hearing appear is simply addictive..

The new global database of annotated bird songs

All this deployment of artificial intelligence would not be possible without quality raw material: well-labeled recordings. In this regard, a team led by the Forest Science and Technology Centre of Catalonia (CTFC) has taken a huge leap forward by publishing the world's first database of detailed annotated bird songs..

This work, published as a data paper in the journal Ecology, compiles recordings from 72 locations spread across the planet. In each file, local ornithologists have manually marked the exact moment when each species sings.covering more than 1.100 different species and exceeding 90.000 individually recorded vocalizations.

The database is available in open access through the Zenodo platform. Its value lies in combining wide spatial coverage with highly accurate annotation by experts, giving it exceptional reliability.Having this type of data is essential for the development and evaluation of recognition algorithms.

Thanks to this resource, researchers around the world can create or test acoustic identification models, either by improving existing tools like BirdNET or by designing new architectures from scratch. In fact, this data collection has already been used to evaluate BirdNET's performance and optimal execution parameters on a global scale., thus refining its practical use in different regions and conditions.

The initiative is led by the research group “New tools for monitoring biodiversity” of the CTFC. Its director, Cristian Pérez-Granados, highlights precisely that component of open, efficient and reproducible science, in which expert knowledge is shared so that anyone can build upon it.

AI, citizen science and biodiversity conservation

The combination of field observation, citizen science, and artificial intelligence is changing how biodiversity is monitored, and there are also options for Identify plants and animals with Google Photos and Google Lens. Tools like BirdNET or Merlin allow for the continuous and massive collection of information on the presence of birds in very different regions, multiplying the scope of ornithologists' work.

Far from replacing experts, these automated systems act as capacity multipliers. A small team can deploy acoustic sensors or promote the use of these apps and receive a volume of data that would otherwise be impossible to collect manually.This translates into better detection of changes in populations or the arrival of invasive species.

In a context of biodiversity crisis and accelerated climate change, detecting warning signs in time is crucial. Having AI-supported monitoring networks makes it easier to see trends, identify problem areas, and prioritize conservation efforts.The sound of birds, which is often the first thing to change in response to habitat alterations, thus becomes a very sensitive indicator.

The philosophy of open resources and global collaboration that underlies these projects also makes a difference. The fact that a database as valuable as the CTFC's is available free of charge, or that projects like BirdNET allow their models to be downloaded, fosters a more participatory science.where large and small institutions can contribute and benefit equally.

Ultimately, the great achievement is not just that your mobile phone can tell you which bird is singing, but that millions of those small individual identifications become a dynamic map of the state of our ecosystems. While you enjoy listening and learning, you are contributing data that helps protect the very species you are amazed to hear about.That combination of technology, passion for nature, and collective collaboration is probably the most powerful aspect of this new era of AI-powered bird song recognition, even when we go out without a network connection.