Spoken Language Translation in Low‐Resource Language
S. Shoba, A. Sasithradevi, S. Deepa · 2024
The real world challenge of individuals to communicate between each other is the language barrier. Most of the languages in India are low-resource ones. People in India use different languages for communication. Various research were carried out by translating the source languages to target language through devices. The technologies used are automatic speech recognition and machine translation (MT) system. There are multiple machine translator devices available in the market and shows considerable performance for high resource languages (HRL). But very few devices are available in use for low-resource languages (LRL). The primary drawback in LRL are the lack of annotated dataset, complex grammar-based system, missing deep analysis, and low production quality. This chapter develops an automatic MT system for LRL by boosting the dataset, improving the quality by evaluating the samples through deep analysis, and increasing the production quality by marketing the need of translation. Neural MT uses artificial neural network to predict the target language for increasing the accuracy. Applying supervised, semi-supervised, and transfer learning model to the spoken word will translate and predict the target language. This work explores different translation dataset collections and in addition to that, the performance of the model are evaluated by metrics like BLEU, chrF, and latency. The efficiency of LRL can be improved by voice activity recognition, speech transcription, speech recognition, and validation of target language.