model-zoo VS NaiveGAflux.jl

If your task is limited to isolated words or words that are separated by spaces, you can reverse the [CMU Pronouncing Dictionary](www.speech.cs.cmu.edu/cgi-bin/cmudict). In a programmatic environment with a dictionary/hash map, you will need to have the values be an extensible list of some sort to account for homophones. You could also train an ML model to convert phonemes to graphemes, like reversing this neural network model.

You can also do some deep learning in Julia if you like. Flux.jl is Julia's deep learning library, and they have a model zoo of easy to follow working examples as a good starting point.

Secondly, flux.train! computes epochs automatically and is designed to take the entire data set. If you want to use the DataLoader, I think you probably need to write a custom training loop. It's not that complicated thankfully. The model zoo has some nice examples like: https://github.com/FluxML/model-zoo/blob/master/vision/mlp_mnist/mlp_mnist.jl

Check out the Flux.jl model zoo: https://github.com/FluxML/model-zoo#examples-listing

Fwiw (and shameless plug disclaimer), I made a library for easily toying around with evolution based NAS. It uses mixed integer programming to align the neurons (e.g. when resizing a layer) which it gives much higher degrees of freedom over what architectures one can allow in the search space over anything else I have seen: https://github.com/DrChainsaw/NaiveGAflux.jl

Maybe not exactly what you are looking for, but: https://github.com/DrChainsaw/NaiveGAflux.jl