What the tool listens for
- Animals. Birds, frogs, insects and some mammals, named by the BirdNET+ V3.0 model. Each species adds the countries where it lives BirdNET+ V3.0 model files, Zenodo (read 21 Nov 2026).
- Season and place. With a date and a lead, the BirdNET+ Geomodel says whether each animal is expected there that week.
- Mains hum. 50 Hz or 60 Hz electrical hum splits the world roughly in two Wikipedia, mains electricity by country (read 21 Nov 2026).
- Old TVs. The whine of a tube TV shows the old 525-line or 625-line TV system.
- Grid frequency. Clear hum is traced second by second. Experts can match the trace with grid records to check the time.
- Speech. VoxLingua107 names the spoken language of each voice SpeechBrain VoxLingua107 model card (read 22 Nov 2026). For English, an accent model adds a weak hint about the speaker’s country.
- Voices and words. The voices are told apart, and Whisper base can write down what each one says. Place names said aloud become clues.
- Music. Music is marked with its times. Save a 15-second clip and play it to Shazam or SoundHound to name the song.
- Passing traffic and aircraft. A tone that drops in pitch gives a rough speed and distance. A slow, high pass is probably a plane.
- Other sounds. Sirens, trains, gunfire, explosions, fireworks and music are noted with their times.
- Sounds tied to regions. Calls to prayer, siren styles, train horns, tuk-tuks, temple bells and more are matched by the CLAP sound model. Play each one and check it by ear.
- A map of the sound. A waveform and spectrogram mark where each animal, sound and passing vehicle was heard.
- Hidden data. Recordings from phones in M4A or MP4 files can keep a time and sometimes a GPS position.
One click, one report
Press “Find location clues” once. Each ticked check runs in order, and the report opens at the top of the results. It covers the bottom line, the evidence, the checks and your next steps.
What to do with the results
- Listen yourself. Every sound has a play button. Compare animal calls with recordings on xeno-canto before you rely on them.
- Match aircraft. With the recording time, flight-tracking history can show which plane flew over. Our aircraft noise guide explains how.
- Read more. The birdsong and ambient sound guide covers what else a soundtrack can reveal.
- Protect your own recordings. The Doppler Scrubber hides plane tones and mains hum before you publish.
What leaves your browser
Your recording never leaves the tab. With “Online checks” ticked, one date and position go to Open-Meteo, and Wikimedia Commons is asked for photos near the top lead. The models are checked against SHA-256 hashes before they run.
Limits
Short clips, wind and music hide most sounds. Sound can be added in editing, and pets, zoos and recordings played aloud can mislead. The speed of a passing vehicle is a rough estimate.
This tool is not made or endorsed by the BirdNET project. Its use policy asks that nobody use it to harm wildlife or to watch people in secret BirdNET licence and acceptable use (read 21 Nov 2026).
Models and data
| Part | What it does | Download | Licence |
|---|---|---|---|
| PP-OCRv5 mobile | Finds and reads text; one reader per script | 13 MB, plus 8 to 17 MB per extra script | Apache 2.0 |
| SigLIP base | Visual clues: scenery, plants, 55 building styles, vehicles, roads, street details, flags, script style; also vehicle makes | 100 MB | Apache 2.0 |
| GeoCLIP | AI location guess over 26,958 map cells | 307 MB model + 14 MB index | MIT |
| OSV-5M baseline (Astruc and others) | Second AI location guess over 11,399 map cells, trained on street-level photos; off unless you tick it | 224 MB, served from this site (4-bit weights) | MIT (trained on CC BY-SA data) |
| Pl@ntNet MobileNetV4-Conv-S | Names plants among 7,806 species of southwestern European flora; off unless you tick it | 50 MB, plus a 1.4 MB species list | Apache 2.0 |
| GBIF.org occurrence counts | Which countries have records of each plant species (read 22 Nov 2026) | In the species list | Counts only; GBIF data are CC0, CC BY or CC BY-NC per dataset |
| RT-DETR r18 (Peking University; ONNX by onnx-community) | Finds people and vehicles; each vehicle is then compared with 94 written descriptions | 21 MB | Apache 2.0 |
| The browser’s built-in AI (optional) | Where the browser offers them: its AI model describes each frame and summarises the report; its translator, language detector and QR reader help too | None from us; the browser fetches its own model | The browser’s terms |
| Depth Anything V2 Small (ONNX by onnx-community) | Estimates depth for the 3D view and walk-through; only the Small model is Apache 2.0 | 26 MB | Apache 2.0 |
| opus-mt (Helsinki-NLP; ONNX by Xenova) | Translates signs and speech into English from 14 languages, off unless you tick it | About 110 MB per language | CC BY 4.0 or Apache 2.0 |
| VoxLingua107 ECAPA (SpeechBrain) | Spoken language, 107 languages, for each voice | 22 MB, served from this site | Apache 2.0 |
| CommonAccent ECAPA (Zuluaga-Gomez and others) | English accent, 16 accents, only when the speech is English | 21 MB, served from this site | MIT |
| ECAPA-TDNN VoxCeleb (SpeechBrain) | Telling voices apart within one file | 21 MB, served from this site | Apache 2.0 |
| Whisper base (OpenAI; ONNX by onnx-community) | Transcription, off unless you tick it | 77 MB | MIT |
| YuNet (OpenCV Zoo) | Finds faces so each voice can be matched to the one whose lips move; it does not recognise anyone | 0.2 MB | MIT |
| Whisper tiny | Spoken language, only if the main language model cannot load | 41 MB | Apache 2.0 |
| BirdNET+ V3.0 (The BirdNET Team) | Birds, frogs, insects, mammals and 16 human-made sounds: 11,709 species in all | 69 MB, served unchanged from this site | Apache 2.0 |
| CLAP (LAION; ONNX by Xenova) | Sounds tied to regions, matched against short descriptions | 33 MB | Apache 2.0 |
| AST AudioSet (MIT; ONNX by Xenova) | General sound labels, off unless you tick it | 87 MB | BSD 3-Clause |
| Libraries | transformers.js (sound feature extraction), wavesurfer.js (waveform), Meyda (audio features), exifr (photo metadata), mp4box.js (MP4 metadata) | Small, from jsDelivr at fixed versions | Apache 2.0, BSD 3-Clause, MIT, MIT, BSD 3-Clause |
| BirdNET+ Geomodel V3.0.5 | Where and in which week each species occurs | 7.4 MB, plus a 1.2 MB range list | Apache 2.0 |
| GeoNames | 34,248 towns and cities with 15,000 or more people, 2,548 region names, and country facts | 2.7 MB, with the page | CC BY 4.0 |
| Natural Earth | Country outlines for the map | 0.1 MB, with the page | Public domain |
| libphonenumber geocoding data | 14,079 phone area codes in 45 countries | 0.5 MB, with the page | Apache 2.0 |
| Name Suggestion Index | 43,683 brand, operator and transit names and the countries they trade in | 2.7 MB (0.6 MB compressed), with the page | BSD 3-Clause |
| IANA time zones | Which countries used a clock offset on a date | With the page | Public domain |
| OpenStreetMap via Photon and Overpass | Street crossings, sign names and map searches, with “Online checks” ticked, or when you press a button | None | ODbL |
| Zippopotam.us | Postcode and town checks, with “Online checks” ticked, or when you press a button | None | Service |
| Wikimedia Commons | Geotagged photos near the top lead, with “Online checks” ticked, or when you press a button | None | Per photo |
| Open-Meteo historical weather | Rain, snow and cloud on the day, with “Online checks” ticked, or when you press a button | None | CC BY 4.0 |
Frequently asked questions
What files does it take?
MP3, WAV, M4A, OGG, FLAC and most other audio your browser can play. Videos work too; only the sound is used.
Is my recording uploaded?
No. Everything runs in your browser tab. Online checks send only a date and map positions, never the sound.
Can a bird call show where a recording was made?
It can narrow the region. Many species live on one continent or in a few countries. A date makes the check stronger, because many birds migrate.
How big is the download?
Animal sounds are 76 MB, regional sounds 33 MB, and language and accent 43 MB. Transcription adds 77 MB and telling voices apart 22 MB. Ticking “Keep downloaded models” saves them for later visits.
Can it name the music?
It finds where music plays, but it cannot name songs. Save the 15-second clip and play it to a song-ID app such as Shazam or SoundHound.
Can it tell the speakers apart?
Usually, for two to four clear voices. It groups stretches of speech by how they sound. It cannot tell who anyone is, and similar voices or crosstalk can merge or split speakers.