OFFLINE INFERENCE
BirdNET Demo
Try the latest BirdNET models right in your browser: identify species in recordings, explore where and when they occur, and draw boxes around bird calls. Everything runs on your device; your audio is never uploaded.
1
Load a recording
Drop a file, record with your microphone, or try the sample soundscape.
2
Watch it analyze
Species appear while the model runs on several CPU cores in parallel.
3
Explore and export
Listen to each detection, filter by location and date, and download a CSV.
Drop a recording here
WAV, FLAC, MP3, OGG or M4A, up to one hour. Audio is processed in your browser and never uploaded.
Recording...
0:00
Up to 5 minutes. Hold your device still and keep quiet for the best results.
Click to listen · Ctrl + scroll or the buttons to zoom · drag to pan
Change colors, brightness and frequency scale
Species groups
What do the percentages mean?
- What it measures
- How closely the sound in one stretch of the recording (a 3-second window by default) matches what BirdNET learned about a species from many example recordings. Only the sound counts; the model does not know where or when you recorded.
- What it is not
- It is not the chance that the identification is right. A detection at 80% is not correct 8 times out of 10, and the same number means different things for different species: some reach high scores easily, others rarely do, even when they are right.
- Why it goes up or down
- Clear, loud, complete calls close to the microphone score high. Distance, wind, rain, traffic, overlapping voices or a call cut off at the edge of a window lower the score. Each species is scored on its own, so the numbers do not add up to 100%, and similar-sounding species can both score high.
- On this page
- A species card shows its highest score in the recording; the list inside shows the score of each window. Min. confidence hides everything below it: higher values give fewer but more reliable detections, lower values find more but also more mistakes. The location filter does not change the scores; it only hides species not expected at that place and time.
- Check before you trust
- Listen to the detections and look at the spectrogram, especially for low scores and for rare or unexpected species.
BirdNET+ V3.0 developer preview 3.1 recognizes 9,789 species of birds, frogs, insects and mammals in 32 kHz audio. The 64 MB model downloads once, stays cached in your browser and runs locally with ONNX Runtime Web.
1
See where species are
The map shows how many species are expected in each area. Drag the week slider or press play to follow the seasons.
2
Click a place
See the species expected there and when they occur through the year.
3
Open a range map
Click a species in that list, or search one, to see where it occurs.
Map
…
Computing…
Loading geomodel…
Research preview: the maps show the relative likelihood of occurrence learned from citizen-science observations, so they also reflect where and when people observe. Do not use them as the only basis for conservation or legal decisions.
Click a species to see its range map.
SpeciesJan – DecThis week
BirdNET+ Geomodel V3.0.4 predicts which of 9,789 species occur at a place in each of 48 weeks; the 13 MB model runs locally. GitHub
1
Load a recording
Open a file, record, or enter a Xeno-canto number. Audio from the V3.0 tab is available here too.
2
Review the boxes
Click a box to hear it, then verify or adjust it, or draw boxes the detector missed.
3
Share with Xeno-canto
Check species and sound types, add your details, then download Annota-JSON.
Drop a recording here
WAV, FLAC, MP3, OGG or M4A, up to one hour. Audio is processed in your browser and never uploaded.
Recording...
0:00
Up to 5 minutes. Hold your device still and keep quiet for the best results.
Change colors, brightness and frequency scale
- Select and listen
- Click a box on the spectrogram or its number in the table.
- Verify
- When a box, its species and sound type are right, click in the table or press V. Boxes you verify, adjust or draw are exported under your name; the others under BirdBox.
- Move and resize
- Drag a selected box to move it; drag its edges or corners to resize it.
- Draw a box
- Turn on the box tool or press B, then drag across the sound.
- Remove
- Press Delete or use × in the table. Detected boxes turn dashed and can be restored; drawn ones can be undone.
- Zoom and scroll
- Ctrl + wheel or pinch zooms, Shift + wheel or dragging the background scrolls, Space plays.
- Colors
- detected verified by you drawn by you removed
Calls in the same frequency band closer than this are joined into one phrase box. 0 keeps single calls.
What do the numbers mean?
- Conf.
- How sure BirdBox is that the box holds a bird sound, from 0 to 1. It does not say which species. A box joined from several calls shows their average; hover over the number to see the highest.
- Species %
- BirdNET V3.0's confidence score for the sound around the box: how closely it matches that species. Below 35%, the box keeps no species, but you can still pick one from the suggestions.
- Not a probability
- A high number does not prove a box right, and a low one does not prove it wrong. Faint, distant or overlapping calls score lower. Listen to a box before you verify it.
| # | Time | Conf. | Species | Sound type | Remove |
|---|
BirdBox Just-Bird finds bird vocalizations as time-frequency boxes. One 35 MB ONNX graph computes the PCEN spectrogram and runs the detector, calls can be joined into phrases; BirdNET V3.0 suggests the species. Exports Xeno-canto Annota-JSON. Documentation · GitHub
1
Load a recording
Drop a file, record with your microphone, or try the Blue Jay sample.
2
Run the analysis
Choose the label language, overlap and an optional location filter, then click Run Analysis.
3
Explore the results
Expand a species to play its detections, or download the results as CSV.
0:00 / 0:00
Waiting for audio…
Run BirdNET locally with TensorFlow.js. All audio is processed within your browser—nothing is uploaded. Code adapted from georg95/birdnet-web.
Found a bug or have an idea for the demo?