Beta We are beta testing now, and open for business soon.
The analysis underneath
Acoustic Analyzer finds the sounds in your recordings and says how sure it is and why. Your reviewers' answers go back into it, so it gets better on your own data.
Acoustic Analyzer
Acoustic Analyzer runs over a recording and writes what it found beside it, in the documented detection format. It will be released as open source under the Apache 2.0 licence: the code and the file format.
Portable
Its model runs as ONNX: no deep-learning framework is needed to analyse a recording.
Any length
From a clip to a recording of hours; a long recording can carry hundreds of detections.
Every decision explained
Each detection comes with a sentence a reviewer can read, built only from the numbers beside it.
What it listens for
The current underwater model sorts each stretch of sound into seven families:
- Baleen whales
- Toothed whales and dolphins
- Seals
- Man-made sound
- Pingers
- Rowing
- Background
Working in another field? We can train a model for it with you, on its own sounds and your reviewed recordings.
Its evidence, not just its answer
For every detection, the analyser works out which part of the sound its answer rests on: it blanks each part in turn, in time and frequency, and measures how much its answer drops. Each part that carries the answer becomes a box.
| It tells you | As |
|---|---|
| What its answer rests on | a box in time and frequency, and its share of the evidence at that moment |
| How sure it is | in words: very sure, fairly sure, leaning that way, a guess |
| What else it might be | the runner-up, when it comes close |
| Whether the sound is outside what it was trained on | where the sound's energy sits against the band the class was trained on |
| How loud the moment was | silent, quiet, typical or loud, against the recording's own level |
Detectors and the catalogue
Detectors
Fast scans for pingers, steady tones and porpoise clicks, run before or beside the model.
The catalogue
Animals, devices and ambient sounds, each with its measured band, pulse, interval and shape, and the source of every number.
Measured, then matched
What a reviewer measures in Analyse is matched to the catalogue with a percentage fit; the pinger and click detectors use the same catalogue.
Better on your own data
Each round makes the next one better:
- The analyser flags what it finds.
- Your reviewers listen and answer in Sound Onion.
- The Publish results button gathers their answers into a
learningsfile. - The analyser's thresholds are set again from those answers, and the corrected examples are gathered for training.
- It flags better next time, and your reviewers' time goes further.
Every result names the learnings file it came from, so a wrong review can be traced to what it changed, and left out of the next set.
The formats it reads and writesTalk to us
We are beta testing now and will open for business soon. Tell us what you record and how much of it there is: we will show you your own data in Sound Onion, and you will be among the first.
accounts@sound-onion.com