QILLQAQ

The current paradigm, hidden Markov model has generated commercial tools (Google Voice, Amazon Echo, Cortana or Siri) with an impressive level without achieving perfection but an operational level. However the hidden model of Markov practically reached the ceiling and any new advance will be only incremental. It should be remembered that the commercial tools mentioned are only available for approximately fifty languages. If one opts for this technology, perhaps the current level, operational but imperfect, of the ASR systems in other languages could be reached for the Quechua language, nevertheless the risk would be that the developed product quickly becomes outdated and outdated.

It was found that there is an increasing tendency to use models based on neural networks; the learning curve takes more time but it is clear that in the long run this is the best option.

Steps for creating an ASR:

The voice recognition system that directly transcribes audio data with text without the need for an intermediate phonetic representation is based on a combination of "deep bidirectional LSTM recurrent neural network architecture" and "Connectionist Temporal Classification objective function".