Multilingual Speech Identification Framework (MSIF) A Novel Approach in Language Identification

Abstract

Multilingual language detection is the process of automatically identifying the language(s) present in a given speech corpus that may contain multiple languages. Several approaches have been proposed for multilingual speech corpus detection, including statistical methods, machine learning algorithms, and deep learning models. These models have difficulty determining specific language, especially when dealing with biased towards certain accents, dialects, or languages, and reduce the accuracy of the model. Hence a novel framework named “Multilingual Speech Identification Framework (MSIF)” is developed to solve this problem by data augmentation and increase the accuracy of language identification. There is a limited amount of datasets available for languages except English makes it difficult to train the Indian regional language. So the proposed framework uses a novel Superintendence Neuvised Network, which combines GAN and CNN for data augmentation and transfer learning for feature extraction. The existing multilingual models have been implemented to identify the languages but these models were not able to detect dialect variations because these model does not utilize the attention mechanism. For this reason, the proposed model uses a novel Duel Atenuative memory network, which integrates a Generalized self-attention mechanism with bi-LSTM to understand dialect variations thereby providing better language detection in the Indian regional language.