Towards the implementation of an Attention-based Neural Machine Translation with artificial pronunciation for Nahuatl as a mobile application

There are great translation systems online. However, even though this technology is available for the majority of languages, it is not the case of Nahuatl [1]. For this reason, this paper outlines a master's degree thesis proposal which is aimed to use a Neural Network for Translation with an attention mechanism and Long Short-Term Memory (LSTM) like the one used by Google [2]. In addition, it seeks to implement an artificial Text To Speech (TTS) system trained with a Neural Network with a given dataset of Mel spectrograms in [3] from a person speaking in Nahuatl and it attempts to achieve a natural voice output as a spectrogram, then process it and obtain the sound desired. Finally, once trained, these models can be prepared for being used within mobile devices, and even taking advantage of the neural engine some of them are equipped with. In this way, this technology can reach more people and help to preserve and even spread the language. The early results showed how the limited resources of this language could cause a strong bias in the outputs and also how there could be some loss of information given the morphemes Nahuatl has, given its polysynthetic nature. This also highlights the way it can be tokenized, playing an important role in how the results turn out obtaining a BLEU score of 0.34 at best. Finally, this application and research can be an interesting framework of how a polysynthetic language can be manipulated to be used for fusional languages like Spanish or English. This research work was carried out at the "Tecnologico Nacional de Mexico" (TecNM), campus of the "Instituto Tecnologico de Apizaco" (ITA)