Google have ignored all warnings from the Terminator and opened sourced their AI code. This has huge potential for use in conservation. How can we use it?
20 December 2015 7:05pm
"TensorFlow, you see, deals in a form of AI called deep learning. With deep learning, you teach systems to perform tasks such as recognizing images, identifying spoken words, and even understanding natural language by feeding data into vast neural networks. "
Would this be applicable to an acoustic monitoring network? For example. my research has shown tigers have unique, identifiable vocalizations down to the individual and sex. If this software is applied to my recording network for tigers, would it be able to automatically recognize and categorize these individuals?
For example: when it hears Tiger 108, it would know and then input that it heard Tiger 108 at a particular time and date.
11 January 2016 12:38pm
The catch will be (and for any neural network or AI type learning I would expect the same) the training phase. If you are able to tell the sounds apart or identify a specific sound as belonging to a certain individual, the AI should afterwards be able to automatically identify the critical factors needed to distinguish the voices of the individuals. But it will need enough input from each individual as well as the different vocalizations used by tigers. AFAIKT it will be able to do this automatically afterwards, but I am not sure if (a) you will get enough identifiable vocalisations and (b) with a wide enough range of typical tiger vocalisations for it to be really reliable. Training on zoo animals might work? I am also interested in this, but for jackals instead of tigers.
11 January 2016 2:30pm
I'd like to suggest our open source package Wildbook (http://www.wildbook.org) as a base data management platfor for this. I agree with the above that there are a number of challenges around the vocalizations themselves, but having the identity information in a good database and data model is a great foundation. That's what we're doing for our computer vision/deep learning project at www.IBEIS.org.
Our non-profit WildMe.org is running both. Feel free to contact us with questions. We have played with time series matching (often used for speech recognition)...but actually for whale flukes. Would be happy to discuss potential for audio ID.
Courtney Dunn
The Prusten Project