There is no official percentage. The BCT guidelines say to always verify a proportion of analysed data, but they give no figure for noise files. A defensible method is a random sample from each detector and survey period: if you check 300 random noise files and find no bats, you can be about 95% confident that fewer than 1% hold a bat. In BioSonic you can show the files the AI marked as noise, open and listen to them, and correct the species.

We make BioSonic. The method on this page works with any bat analysis software.

How many noise files to check

If a random sample contains no bats, the rule of three gives the highest miss rate that is still plausible at 95% confidence: 3 divided by the number of files you checked.

Random noise files checked, no bats foundMiss rate is likely belowIn a noise folder of 10,000 files
605%Up to about 500 files could still hold bats
1003%Up to about 300 files
3001%Up to about 100 files
6000.5%Up to about 50 files

The table is statistics, not BCT guidance. It holds only if the sample is random and you can recognise the bats you are looking for. A very rare species can still be missed.

A method you can write into a report

  1. Split the noise files by detector and survey period, because a windy week or a faulty microphone behaves differently.
  2. Draw a random sample from each part. The first files in a folder are not a random sample.
  3. Check more where a quiet species is likely, such as barbastelle, long-eared bats and some Myotis.
  4. If you find bats, record the species and how many, then check that detector and period in more depth.
  5. Report how many noise files there were, how many you checked and how many held bats.

Wording for the method section

Files classified as noise were sampled at random for each detector and survey period. Of [total] noise files, [number] were checked by [name, role] and [number] contained bat calls. [Detectors or periods] were checked in more depth as a result.

Checking noise in BioSonic

Common question

Does BCT say how much of the noise folder to check?

No. The 4th edition says a proportion of data should always be verified for quality assurance, but we found no figure for noise files in it. Choose a sample, give the reason and report the result.

Sources

Related: Why barbastelle calls end up as noise · Cut the time spent checking calls · What BCT says about auto ID