by Sara Heimlich, Dave Mellinger and Holger Klinck

About MobySound.org

MobySound is a sound database for research in automatic call recognition (ACR). Many ACR methods require training data to work correctly; for instance, neural networks need to "learn" to distinguish calls from non-calls. MobySound provides that training data. MobySound contains recordings of a number of species of marine mammals. MobySound is not intended to be a general-purpose library of animal sounds; there are animal sound libraries that do that job well. Rather, MobySound is targeted at ACR research. There are several things that make MobySound different from a more general-purpose animal sound library:

  1. The recordings here are often poor quality. This is because ACR algorithms are often used in situations where the animals to be detected are distant or are in a noisy environment -- where the sound signals are of low quality. To recognize sounds in these situations, ACR algorithms need training data that is of similarly low quality.
  2. The recordings are annotated. Each recording containing a certain species calls has one or more annotation files associated with it. These annotation files specify where in the recording the calls of that species occur. This information is used both in training call recognition algorithms, so that the training method knows what the call of interest sounds like, and in testing algorithms, so that anything that a recognition algorithm registers as a detection can be checked to see if it really is the call of interest.
  3. Recordings are long and continuous, often containing long periods of time when no calls of interest occur. Again, ACR methods must often work in such situations, where the calls of interest are widely scattered, so the training data here must be arranged similarly.

Types of Objects in MobySound