Should the testing of my learning algorithm be restricted only on standard datasets or can I use any dataset to publish my results?

Question

I've trained a speech recognition model using the LibriSpeech database which isn't as widely used as other datasets like TIMIT or MNIST. Does that factor in any way or can I publish the results irrespective of what dataset I use?

The reason I ask this is, if we only use datasets like TIMIT, models could be made to overfit those datasets for greater accuracy.

Is there a thing such as standard datasets in speech recognition?

Or else, which are some other easily available speech datasets for multiple languages?

Answer 1

The advantage of using a standard dataset is that it allows direct comparison of results and performance. In essence, standard datasets allow benchmarking by providing a common, accepted measure. That said, I don't imagine any journal editor or reviewer would scoff if you included results of a different dataset in addition to the standard ones in your discipline.