About me

Researcher at Norwegian Institute for Nature Research (NINA)

Currently mainly working with bioacoustics

... but also HPC, Deep learning

Background

Eco-acoustic monitoring is increasingly being used to map biodiversity across large scales

Little thought is given to the privacy concerns and potential scientific value of inadvertently recorded human speech

We developed an end-to-end VAD pipeline and show how VAD models can be used for anonymisation & human noise quantification!

Method

We developped an end-to-end pipeline (ecoVAD) and compared it with 2 state-of-the-art VAD models:

• Pyannote
• Google's WebRTC VAD

Note that in our repo we provided wrappers for using both models

Method

Training : We collected soundscape data that we mixed with human voices & other noises.

Testing : playback experiments with speech samples from a man, woman, and child at 1,5,10 and 20 meters.

Using: we collected soundscape from a recreational area to evaluate the ability of the models to detect human speech.

Results

• All model performed well on the playback data

• Our custom model performs better

Results

Discussion

The importance of training a model for specific purpose

But ... there is a trade of to make with generalisation (Pyannote performed better on a more "anthropogenic")
area

Speech detections as a direct measure of anthropogenic noise pollution and indirect proxy of human disturbance

Discussion

Have a look at the ecoVAD GitHub repo!

And very soon the paper in MEE!

Current work on snowscooter detections

Please reach out to us for any questions / collaboration!