Speech recognition and voice command control system for Luganda language
Muwebwa, Solomon Eltonjim
MetadataShow full item record
A Speech Recognition and Command Control Interface is software that allows the user control over specified computer functions by voice. Our implementation of this system is unique in that it responds to commands made in the most commonly spoken native tongue in Uganda, which is Luganda. The system is a composition of four different co-dependent processes; Signal processing which filters and cleans the audio to be used such that only the necessary frequencies are present, Acoustic modelling-which detects specific features present in the signal and constructs a likely text representation of the signal, and Language Modelling-which pieces the detected words together in a manner that approximates the most likely combination to create coherent sentences, and finally keyword mapping-which looks out for specific vocabulary present in the sentences in order to determine which command is being issued. As a result of our study, we have created a sample Luganda keyword dataset for specific computer commands, with over 1400 audios from over 80 different speakers, and used this data to train a neural network model that at the moment has an accuracy of 60%, but is improving as we add more data samples. In conclusion, we are satisfied with the existing project, as it is capable of training a model to detect almost any grouping of keywords, speech or otherwise, and has simplified the process by only requiring a user to upload keywords. For future users, we recommend heavy testing in order to optimize the parameters of the model, as well as creation of a large enough dataset, from a wide variety of sources.