Current Automatic Speech Recognition (ASR) systems only delve so far into a speech signal, limiting the amount of data available to train a better model to convert speech to text. Our proposition is to deploy a system that uses the speech signal down to its acoustic cues and deeper phonetic information after using relevant speech measurements. This will allow our model to be trained on richer data that current ASR systems do not leverage. Furthermore, because of the modeling methods that our system uses, we are able to achieve modularity in our system that current systems do not offer. The system is Lexical Extraction and Integration (LEXI). The idea is to build a text output from a Lexicon dictionary through acoustic cues that are generated via a new model. The result is a simpler system that can be easily altered, as there is no reliance for fine-tuned specific modeling methods. My project will focus on two critical components to the system’s development and deployment. These include building the modules that classify the acoustic cues that make up speech using predictive methods (development), and additionally, deploying the system on a website that can be used both internally (to train newer UROP scholars, for example) and externally, for the academic/public community.

I have worked with the Speech Communications Group in the past and found both the work and the goal of the research interesting and impactful. Moving forward, I would like to help make the speech system come to fruition by improving the system’s detection rates alongside of deploying it for use.