However, it is still time-costing to train a deep network model by these DL algorithms for a large-scale sound dataset.
However, as paradoxical as it may sound, mobile phone datasets are not appropriate to observe real propagations of information.
The MAPS (standing for MIDI aligned piano sounds) dataset [34] provides high-quality recordings of a Yamaha Disklavier, that is, an automatic piano.
However it does raise 2 issues, one being that IMHO this generation should focus on providing sound traceable datasets upon which those who follow can base their research/decisions.
While our initial experiments were focused on very specific recordings such as the ones in the Sound Events dataset, adding more examples allowed us to generalize to a wide variety of recordings.
However, this approach is notoriously prone to introduce errors in itself, due to the limitations in reproducibility for human operators 7. Furthermore, to create a statistically sound ground-truth dataset, especially when working with biological data and its basically infinite variations, it is often necessary to manually analyze huge quantities of data.
The goal of this work is to provide the capability to accurately detect and classify whale sounds in large acoustics datasets collected using underwater sensors.
First, the initial dataset of sound samples is roughly clustered using the k-means algorithm with the k-means++ seeding procedure in order to find the cluster best representative of the structure.
To evaluate the performance of the hybrid network we adopt a two-fold cross validation approach where all of the sound files in our dataset are partitioned into two nonoverlapping subsets.
Since the model represents the most regular sound shape in the dataset, irregularities, i.e. distant sound samples can be recognized as characteristic of defects on the structure, distance to the regular model can be expressed as a similarity value (Fig. 4).
