Detect Sounds in Audio and Get Timestamps

Upload a handful of examples of a sound — a dog barking, a gunshot, a doorbell, a bird call, a cough — and this trains a detector for it in your browser. Point it at a long recording and it returns every moment that sound occurs, with start and end times you can download as a subtitle file or a spreadsheet.

How to Detect Sounds in Audio and Get Timestamps

  1. 1

    Name your sound and add examples

    Create a class, call it what you like, and add five to ten short clips containing the sound. One or two seconds each is ideal — trim them so the sound fills most of the clip.

  2. 2

    Fill the Background class

    Add clips of ordinary audio that does not contain your sound: room tone, speech, traffic, whatever the recording normally sounds like. This is the step people skip, and it is the one that stops the detector from labelling everything a match.

  3. 3

    Train the detector

    Training runs in your browser and takes a moment. You will see how well the detector fit your own examples — if that number is low, two classes probably sound too alike.

  4. 4

    Scan a recording and export

    Drop in the file you want searched. Adjust the confidence threshold and minimum event length until the results look right — both update instantly — then download an SRT or CSV of the timestamps.

How it works

Detection runs on YAMNet, a Google convolutional network trained on AudioSet, a corpus of about two million human-labelled clips from YouTube. YAMNet is not retrained here — it is used as a listener that converts each 0.96 seconds of audio into a 1024-number description of what it sounds like. Your example clips are turned into the same kind of description, and a small classifier is fitted on top of them in your browser, which is why a few clips are enough where training a network from scratch would need thousands. The recording you scan is then stepped through in 0.48-second hops, each step scored, and neighbouring matches merged into events. It is the same transfer-learning idea behind Google’s Teachable Machine, but it runs entirely on your own device and works on audio files rather than a live microphone.

Credits

The research this tool runs on, credited in full — with BibTeX and licence notices — on the credits page.

  • YAMNet

    Apache-2.0
    Shawn Hershey, Sourish Chaudhuri, Daniel P. W. Ellis, Jort F. Gemmeke, Aren Jansen, R. Channing Moore, Manoj Plakal, Devin Platt, Rif A. Saurous, Bryan Seybold, Malcolm Slaney, Ron J. Weiss, and Kevin Wilson. 2017. CNN Architectures for Large-Scale Audio Classification. In Proceedings of ICASSP. arXiv:1609.09430.Read the paper

  • AudioSet

    CC BY 4.0 (ontology and labels)
    Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. 2017. Audio Set: An ontology and human-labeled dataset for audio events. In Proceedings of ICASSP, New Orleans, LA.Read the paper

  • MobileNets

    Apache-2.0
    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv:1704.04861.Read the paper

  • ONNX Runtime

    MIT
    ONNX Runtime developers. 2021. ONNX Runtime: Cross-platform accelerated machine learning.

Supported formats

WAV, MP3, M4A, AAC, OGG and FLAC. Files are decoded by your browser, so anything it can play will work here. Output is exported as WAV.

Frequently asked questions

How many example clips do I need?
Five to ten per class is usually enough, and more helps. What matters more than the count is variety: if you want to catch a dog barking, include different dogs, distances and rooms. Include roughly as many Background clips as sound clips.
What sounds can it learn?
Anything with a consistent character — animal calls, alarms, gunshots, door slams, coughs, engine noise, glass breaking, a specific instrument. It is weaker at distinctions that depend on meaning rather than texture, like telling two people’s voices apart, which is a different task called speaker diarization.
How accurate are the timestamps?
The model listens in 0.96-second windows that step every 0.48 seconds, so an event is placed to the nearest half second. The reported span deliberately errs wide: a window only has to overlap the sound to count it, so a cue can open up to a second before the sound starts and close up to a second after it ends. That is the right trade for a subtitle — the cue is always on screen while the sound plays, rather than clipping its start — but it means the durations in the CSV are an upper bound, not a precise measurement. It will not give you a sample-accurate transient.
Can I save a detector and use it again?
Yes. Export the trained detector as a small JSON file and import it later, or on another machine, without needing the original clips. The file contains only the trained numbers, not your audio.
Why is everything being detected as my sound?
Almost always too few Background examples. The detector has to choose between the classes you gave it, so if the only thing it has ever heard is a meow, everything sounds relatively meow-like. Add more varied Background clips and retrain, or raise the confidence threshold.
Is this the same as Teachable Machine?
It uses the same idea — freeze a pretrained network, train a small classifier on your own examples — but it is built for this job. Teachable Machine’s audio projects are designed around a live microphone and one-second speech commands. This runs on audio files, uses a model trained on general environmental sound rather than speech, and gives you timestamps out of it.
Is my audio uploaded anywhere?
No. Everything runs inside your browser using the Web Audio API, so the file never leaves your device. You can prove it: load the page, turn off your Wi-Fi, and the tool still works.
Does this use AI?
Yes. It runs YAMNet, a neural network Google trained on AudioSet, to turn sound into a form a classifier can work with, and then trains a small classifier of its own on the clips you supply. Both the training and the detection happen on your device — nothing is uploaded, and there is no server doing the work.
Do I need an account?
No. The free tools work straight away with no sign-up. Creating a free account raises your daily processing limit from 20 to 60 minutes.
Is it free?
Yes. Every tool on this page is free to use, with a daily limit on total audio processed. There is no watermark and no trial period.

Other free audio tools