Not sure what that noise is? Let the microphone listen, or open a recording, and an on-device AI model names the sounds it hears, from barking dogs and sirens to speech, music and individual instruments, with a confidence score and a timeline.
Last updated:
AI model (YAMNet, about 15 MB) loads when you start. Audio never leaves your device.
Spectrogram appears when you start listening
Hearing now
Detected sounds
Sounds that cross the threshold are listed here with the time they were heard.
A short synthetic clip: a siren followed by beeps.
Top sounds in this file
Sound by sound
Time
Sound
Peak
It can recognise, among others:SpeechLaughterBaby cryDog barkCat meowBird songRoosterSirenCar hornTrainAircraftDoorbellSmoke alarmPianoGuitarDrumsViolinSingingRainThunderWater tapTypingApplauseWhistling
Advertisement
Advertisement
How to use the Sound Identifier
Choose a source: Microphone (live) to identify sounds around you, or Audio or video file for a recording.
Press Start listening and allow the microphone. The model downloads once (about 15 MB) and then runs offline in the page.
Read the Hearing now list: the five most likely sounds with their confidence. Sounds above the threshold are added to the Detected sounds log.
For a file, choose or drop it (or press Analyze demo sound). The timeline and table show which sounds occur when.
Click the timeline or a table row to play that moment, and download the results as CSV if you need them.
How the AI recognises a sound
Resampling. The audio is mixed down to mono and resampled to 16,000 samples per second, the rate the model was trained on.
Spectrogram. YAMNet computes a short-time Fourier transform with 25 ms windows every 10 ms and groups the frequencies from 125 Hz to 7,500 Hz into 64 mel bands, a scale that resembles human hearing.
Patches. 96 consecutive spectrogram frames (0.96 seconds) form one patch. A new patch starts every 0.48 seconds, so the patches overlap by half.
Classification. A MobileNet-style convolutional network with about 3.7 million weights scores each patch against 521 AudioSet classes. A sigmoid output gives every class its own score from 0 to 1.
This page shows the highest-scoring classes for each patch, merges consecutive patches with the same top sound into one entry, and reports the peak score.
Reading the confidence scores
Score
What it usually means
Above 70%
A clear, typical example of the sound. Very likely correct.
40% to 70%
Probably right, or a close relative (for example Bark versus Dog).
20% to 40%
A plausible guess. Check the next few labels and listen again.
Below 20%
Weak match: the sound is faint, mixed with others or not well covered by the classes.
Labels are organised in a hierarchy, so general and specific labels often appear together: Animal, Domestic animals and Dog can all score high for one bark. The most specific label with a high score is usually the most informative.
Tips for better results
Get closer to the source and keep the phone or microphone still to reduce handling noise.
Record at least two or three seconds of the sound; very short clicks may fall between analysis windows.
Turn off noise suppression in conferencing apps that share the microphone, as it removes non-speech sounds.
For long recordings, use the timeline to find the moments where the sound you care about appears, then click to listen.
Raise the logging threshold in noisy places so only confident detections are listed.
Example uses
Find out whether your dog barked while you were out by analysing a recording.
Check what a mysterious noise in the house or car is likely to be.
Locate applause, laughter or music sections in a long video's soundtrack.
Teach children about sounds of animals, instruments and vehicles.
Tag field recordings and sound effect libraries with a first set of labels.
Frequently asked questions
How does the sound identifier work?
It uses YAMNet, a neural network published by Google that was trained on the AudioSet collection of labelled 10-second clips. The audio is resampled to 16 kHz mono, turned into a 64-band log-mel spectrogram and scored in 0.96-second windows that advance every 0.48 seconds. Each window gets a score between 0 and 1 for each of 521 sound classes.
Is my audio uploaded?
No. The model files are downloaded once from Google's model hub and then the analysis runs inside your browser with TensorFlow.js. Microphone audio and files you open never leave your device.
What sounds can it recognise?
The 521 classes cover human sounds (speech, laughter, coughing, snoring, a crying baby), animals (dogs, cats, birds, roosters, insects), vehicles and alarms (cars, trains, aircraft, sirens, smoke alarms), household sounds, nature (rain, wind, thunder) and music, including genres and individual instruments such as piano, guitar, drums and violin.
Why can several sounds score high at the same time?
YAMNet is a multi-label model: each class is scored independently, so a single moment can be Dog, Bark and Animal at once, or Music, Guitar and Singing. Scores do not add up to 100 percent; read them as how strongly each label matches.
How accurate is it?
It is good at common, clear sounds like speech, music, barking, sirens and alarms. It is less reliable for rare sounds, quiet or distant sources, heavy background noise and sounds that resemble each other. It names the type of sound, not a specific song, person or animal breed, so treat results as a helpful guess.
Can it identify a song or a bird species?
No. It recognises broad categories such as Music, Pop music or Bird vocalization, but it does not identify song titles, artists or individual bird species; dedicated song and bird recognition apps are needed for that.
Why does live mode show a delay?
Each prediction uses the last 0.975 seconds of audio and runs every half second, so a sound usually appears in the list within about one second after it starts.
Are all the tools on FreeToolCabin tested for quality and accuracy?
Yes! Every tool on our site undergoes automated Q/A testing in real browser environments to ensure high precision, correct output, and zero errors. If you happen to spot a bug or have a suggestion, please contact us at [email protected] and our team will review it right away.