
How to record bird calls for AI identification: a field guide

Here is how to record bird calls for AI identification on a phone: get close to the bird without disturbing it and point the microphone at it. Then hold still and record for 30 seconds or longer. In a monitoring project two more things matter. Each recording needs a GPS fix and a timestamp, and the AI's answer has to be read with its confidence score rather than taken as fact.
This guide is for community monitors who make the recordings and for project developers who set up the surveys. It follows the way bird identification works in the CitizenClimate app, and it uses public guidance from the Cornell Lab of Ornithology and from published research on BirdNET.
What the AI needs from a recording
CitizenClimate identifies birds from an audio recording using BirdNET, developed by the Cornell Lab of Ornithology and Chemnitz University of Technology. The model is described in Kahl et al. (2021) in Ecological Informatics. BirdNET is a service the platform calls; it is not a CitizenClimate product.
A few facts about how BirdNET reads audio explain most of the field advice below:
It listens in short windows. The BirdNET-Analyzer documentation implies that detections are reported as 3-second segments by default: its option for merging consecutive detections produces "segments longer than 3 seconds". As an illustration, a 30-second clip holds roughly ten 3-second windows, and a 4-second clip holds one or two.
It reports confidence, not certainty. The same documentation gives a default minimum confidence (`--min_conf`) of 0.25. The CitizenClimate Data Integrity and Methodology Protocol lists, for birds, the "start and end time of the detected call within the audio, minimum confidence threshold of 25%".
What comes back. Each identification returns a scientific name, a common name, a confidence score from 0 to 100% and the species' IUCN conservation status, as the Protocol sets out.
Confidence is a trade-off. In his guide for the British Ornithologists' Union, Cristian Pérez-Granados explains that a higher threshold "increases the percentage of detections correctly classified" but "lowers the proportion of calls detected" (BOU blog, March 2023). A clean, close recording lifts the score for the right species, which is why technique matters more than the phone you carry.

Before the first visit: set up the survey
Good recordings start with the survey a project developer builds. Five decisions are worth making before anyone walks out at dawn.
Add the right question. The App Guide lists the question types: "multiple choice, text, numbers, dates, photos, videos, audio, species identification or GPS". For bird calls, use a species identification question; the monitor will be asked to "record the bird's call".
Require a location. The Protocol states that location "is required or optional per survey type", and when it is required "the system will not allow submission without a valid GPS fix." The app asks the phone for its highest available accuracy ("navigation-grade, typically < 5 metres") and stores the accuracy in metres with each reading, so a reviewer can see how precise each point was.
Choose the audio quality on purpose. The survey builder lets you set audio quality on a recording field, with a low setting meant for slow internet. We have not published tests of how this setting affects identification, so if your sites can upload larger files, try a few recordings at each setting at one point before you fix it for the season.

Fix a field protocol. The biodiversity monitoring page is clear that "The platform does not enforce a specific field protocol" and recommends aligning with whatever your methodology requires, "e.g. point counts for birds". Long-running schemes are strict about this. The North American Breeding Bird Survey instructions say "The official start time is always 30 minutes before sunrise" and, for the count at each stop, "Do not stay less or more than three minutes." You do not need to copy that scheme, but write down your own: the same points, the same time of day, the same season, the same recording length.
Check your standard's requirements. If the project uses Plan Vivo's PV Nature standard, the Protocol notes that terrestrial projects "must include the two mandatory groups (herbaceous & woody plants under 2 m, and birds)", and each group shows its method, "such as field quadrat, acoustic BirdNET or camera trap". Our PV Nature walkthrough shows how that module works with a fictional demo project. CitizenClimate has no affiliation with the Plan Vivo Foundation, and official PV Nature metrics are calculated by an approved Data Analytic Provider, not by CitizenClimate.
Then train people. Project developers can write lessons "with text, images and video", and "Many lessons end with a short quiz" (App Guide). A short lesson on the steps below, plus one practice recording each at a known point, catches most bad habits before the season starts.

How to record bird calls in the field: seven steps
The steps combine the eBird smartphone recording tips from the Cornell Lab with the way the app handles a recording.

1. Wait for "Location obtained"
Open the survey at your point and tick the consent box. If the survey records your position, the App Guide says to "wait for Location obtained". Surveys that require a location "won't start without a GPS fix". Under thick canopy this can take longer; step into a gap if you can and wait rather than walking off.
2. Pick one bird and get closer
eBird's advice is to focus on one bird at a time and to get close if you can do so without disturbing it. Distance matters more than the phone: a bird ten steps away usually sounds clearer than the same bird across a valley on a better device. If the bird moves away or goes quiet, let it go.
3. Aim the microphone and block the wind
eBird's advice is to "Know which microphone your phone is using and keep it clear of any obstructions", to point it at the bird and to point it "away from background noise". On most phones the microphone sits on the bottom edge. We would add two habits of our own: keep your fingers and a bulky case clear of it, and stand with your back to the wind so your body shields the phone. Wind noise across a microphone can swamp a distant call.
4. Stay still and record for 30 seconds or more
eBird asks recorders not to move or talk and to "Aim for 30 seconds or longer when recording a bird." Hold the phone gently or rest it on something steady. Do not whisper notes during the recording; you will add them in a moment.

5. Read the matches before you confirm
When the phone is online, the app "shows a ranked list of possible matches with confidence scores", and you can tap a result to read a short species summary (App Guide). Look before you tap. If the top match is a species nobody has seen in your area, or the confidence is low, do not pick it just because it is first. Record a second clip, or say what you doubt in the notes.
6. Add the count and notes
The App Guide's next step is "Confirm the match, and add a count and any notes. Your location is attached automatically." Use the notes for what the audio cannot show: whether you saw the bird, its behaviour, the weather. eBird suggests ending a recording with a spoken note of the species and behaviour; in a project survey, the notes field does that job without adding speech to the audio.
7. Submit, with or without signal
Tap Submit. With no signal, "the photo or recording is saved and identified when you reconnect" (App Guide). One rule here: never submit the same audio file twice, for example at a second point. The Protocol describes a daily check under which "the same photo or recording reused from another submission", detected "by an exact digital fingerprint of the file", is rejected and archived with a reason, and the points earned for it are reversed.
What happens after you tap Submit
The Protocol sets out what each bird record keeps. Every species observation "is stored as a permanent record containing the scientific name, common name, category, observation count, IUCN status, the submission it came from, the timestamp of observation, GPS coordinates, observer name, and URLs to the supporting photos or audio."
Offline recordings wait with the status `ai_pending`. Once the phone reconnects the identification completes. The Protocol says the original observation timestamp, the moment the species was actually recorded in the field, "is preserved regardless of when identification occurs." The Sync Status screen shows what is still pending.
The Protocol also describes a daily automated check of new submissions with photos, recordings or free text. For audio it notes that "Bird recordings are re-analysed by BirdNET, because audio cannot be judged from a picture." A rejection needs "at least 85% confidence"; anything less goes to a person. A rejected record is archived with a reason, never deleted, and the points earned for it are reversed.
Records then feed the biodiversity dashboard, which shows species and observation totals and diversity indices "such as Shannon-Wiener, Simpson's and Pielou's evenness", and counts and flags species listed by the IUCN as Vulnerable or worse (App Guide). Data can be exported as CSV or JSON. The biodiversity survey walkthrough shows a whole project, set up with a fictional demo, from survey to dashboard.
Making bird recordings a verifier can trust
A verification body with the read-only VVB role sees the record, not the morning in the forest. Every field habit should leave something in the record that a reviewer can check.
Field habit | Why it matters | What the record keeps |
|---|---|---|
Same points, same time of day, each visit | Visits can be compared fairly | GPS coordinates, GPS accuracy in metres, observation timestamp |
Location required in the survey | No bird record without a place | Coordinates that "cannot be modified after submission" (Protocol) |
One bird, close, quiet, out of the wind | Fewer missed calls and wrong matches | The audio file itself; media files "cannot be deleted through the application" |
30 seconds or longer | More windows for the model to hear the call | The full recording, and the AI result's start and end time for the detected call |
Notes on what you saw | Context for a reviewer or an ecologist | Count and notes |
One file, one submission | The daily check treats a reused file as duplicate media | The daily check's verdict, category and confidence |
None of this makes the AI infallible. The BirdNET team's own paper on its app says "classification accuracy will never be perfect" (Wood et al. 2022, PLOS Biology). Pérez-Granados goes further in his BOU guide: validation "should be paramount for any study aiming to use BirdNET to validate its performance", and at the time he wrote "all published studies have been carried out in Europe or North America". His review in Ibis likewise found that the studies it covered focused on North American or European species. If your project is in the tropics, that gap is a reason for more checking, not less.
In practice, build a validation sample into the plan. Each season, ask someone who knows local birds by ear to listen to a set of recordings: some at random, all the low-confidence ones, and every record of a threatened species. Our biodiversity monitoring page says "the platform does not treat AI output as ground truth", and your monitoring plan should not either.
Common mistakes to avoid
Recording on the move. Footsteps and clothing rustle sit right next to the microphone.
Short clips. Three or four seconds may hold one partial call, which gives the model little to work with.
Chatting at the point. A second person talking quietly behind you is often louder on the recording than the bird you came for.
Taking the first match on trust. A rare species at the top of the list deserves a second recording and a note.
Moving the points. Shifting a point a few hundred metres to "where the birds are" breaks comparisons between visits.
Re-using a good file. The daily check the Protocol describes treats a recording submitted twice as duplicate media: the second record is rejected and archived with a reason.
Frequently asked questions
How long should a bird call recording be?
eBird recommends aiming for 30 seconds or longer. BirdNET's documentation implies that detections are reported as 3-second segments by default, so a 30-second clip holds roughly ten such windows in which the model could hear a clear call. In a monitoring project, agree one recording length for every point and every visit, so that results from different days and different monitors can be compared on the same footing.
Does AI bird call identification work offline in the CitizenClimate app?
No. Recording works offline, but identification does not. The App Guide says the recording "is saved and identified when you reconnect", and the Protocol marks it `ai_pending` until then. The original observation timestamp is kept, so the record shows when the bird was heard rather than when the phone found signal. Monitors can carry on with the survey and sync later.
What confidence score is good enough?
The Protocol sets a minimum confidence threshold of 25% for bird identifications, which matches the default in BirdNET's own analysis software. That is a floor for reporting a detection, not a mark of certainty. Pérez-Granados notes that higher thresholds give more correct detections but miss more calls, so choose with your goals in mind and check low-confidence records by ear.
Can phone recordings replace a trained bird surveyor?
Not on their own. Phone recordings let more people collect more data at more points, and each record keeps its audio, timestamp and GPS point for later checking. But published reviews of BirdNET stress validation, and performance outside Europe and North America has been studied less. The strongest set-up pairs community recordings with a regular expert check of a sample.
Start recording
Good bird data is mostly patience: wait for the GPS fix, then give one bird half a minute of quiet. Set the survey up once, train your monitors on the seven steps, and keep a validation sample in the plan. Download the free app to try a bird survey on your own project, or read the App Guide for the full walkthrough.



Comments