Ruru/moreporks automatically identified now

We have just released our automatic ruru/morepork identification software. This software runs over every file uploaded from a Cacophony Bird Monitor and identifies any ruru/moreporks in your recording and adds tags noting at what time in the recording morepork can be found.

To access these tags you need to export your recordings to a spreadsheet. This is the same way you access your cacophony index. You can see details on how to do this in our post about the cacophony index. The spreadsheet shows the birds identified (only morepork at this stage) and the time in the recording that they occurred. When we identify a morepork we calculate a likelihood that we are correct. At the moment we don't include this in the output but are considering including that.

The software uses deep learning to recognise the ruru and is significantly better than anything else we've managed to produce previously. It correctly identifies about 77% of ruru. The deep learning models learn from examples. The more examples we have, the better we are able to train our model.

This problem has been worked on by open source contributors to the Cacophony Project: Dennis Sosnoski and Tim Hunt. The model we are using now was developed by Dennis and we would like to thank both him and Tim for all the mahi that has gone into this. For those that are interested here is the morepork detection software.

Here's an example of a recording that was identified with multiple ruru calls.

In order to expand this to other birds we need to get more data to train our deep learning model. This data simply consists of examples of different bird calls. Please let us know if you have data you think we could use to expand this to other birds. We are looking to use the NZ section of the xeno-canto bird song collection.  This collection was used for a recent BirdCLEF kaggle competition for birdcall identification, but didn't include any NZ birds.

Part of our plan is to ask people to tag some of the almost half a million recordings we have collected. To do this we need to update our tagging interface to make it easier to tag these recordings. We have specified what this needs to look like, but have not secured funding to build this interface yet. We have specified the improvements to our tagging interface that we would like to make. If you know of a developer who may be interested in building this as an open source contribution then please get in touch.

Related reading

• An article by Tim published last year about some of his early work on Morepork detection.