Advancements in Speech Recognition for Regional Languages Using Deep Learning Techniques

Authors:

Abstract
Scientists have been trying to identify and understand impromptu speech since 1970. Many solutions have been adopted to an automated voice identification structure. Users always use English in speech identification since it was the target of much research and implementation. There are several researches conducted to determine the demographic details of a person and his or her terms in the Bangla language but not even a single research has been conducted to find out the dialect of the person speaking. Thus, this work is aimed at identifying and classifying the generated Bangla language, by developing a database of the $\mathbf{3 0}$ hours of Bangla language in five local Bangla languages. The production of Bangla voice was done through the process of first translating the syllables in the Bengali language into English phrases deviation, which is often used as a form of text, and then translating these phrases into an English sentence. The suggested approach was also used to differentiate the regional dialect of the speaker in the audio stream and determine its validity. The processing of this data with deep learning algorithms with convolution tiers can generate feature mappings. The suggested method was widely trained and tested on the basis of unsupervised statistics i.e. the development of new lines with the help of different English and Bangali abbreviated words. The algorithm factored the age category in language category tagging because as age advances physiological changes occur in the brain and it affects how the audio data is handled. This improved classifier accuracy to 95 percent without and with age categories respectively. In the case of generated Bangla language tags, the accuracy in classification is $\mathbf{9 5}$ percent. The proposed solution is also very effective when dealing with audio collections that express English.

This publication has 15 references indexed in Scilit: