Challenges and solutions pertinent to machine learning-based audio recognition

Abstract

The problem of transcribing data from an acoustic waveform has yielded multiple approaches over the last decades. The current preferred approach involves a three-stage model that breaks the problem into its constituent stages, each with an equivalent model. The first stage divides pure acoustics and language study using a Bayesian model. The second stage focuses on the acoustics model; this model makes the most sufficient and efficient division. The third stage focuses on the acoustics model; this model provides complete instruction on how to compute the probability required by the first stage, given the division specification declared within the second stage.