Unsupervised adaptation of ASR systems: An application of dynamic programming in machine learning

摘要

Dynamic Programming (DP) is used for solving various complex problems. In this paper, it is proposed to use DP for measuring the shortest phonetic distance between two words called Dynamic Phone Warping (DPW) and use DPW engine to classify whether a given pronunciation is a new accent or a new word. Humans learn new accents and words from "Every day speech" whereas Humanoids lack this capability. In this paper, an adaptation framework is proposed using DPW algorithm to enable the Automatic Speech Recognition (ASR) systems to learn from the unlabeled data. The algorithms are implemented using Java language. Data sets are extracted from CMU Pronunciation Dictionary CMUDICT, TIMIT speech corpus and Hindu newspaper. The new algorithms have application to unsupervised learning and adaptation of ASR systems. It makes the ASR systems inexpensive, fast and improves performance of the existing systems.