Development Of a Speech-To-Text System for Sanskrit: Leveraging Linguistic and Computing Techniques

Abstract

Objectives: Speech-to-text (STT) is a technology for converting spoken words into printed text. It's also called Automatic Speech Recognition (ASR). STT systems analyze audio data and detect spoken words using complex algorithms and machine learning models. The natural-language idea has been dubbed an "AI-complete problem" since it appears to necessitate both the ability to use language and a thorough grasp of the surrounding world. In the case of Indian languages, there are 1652 dialects and 22 official languages, with more than 30 languages spoken, including six Indian languages on Google Assistant.
Methods: Sanskrit is highly valued in literature, culture, and traditions. Languages such as Assamese, Dogri, Punjabi, Kashmiri, Sanskrit, and others remain understudied. Only a fraction of India's 22 constitutional languages have had NER work done. Some obstacles include a scarcity of resources, verbal ambiguity, and morphological diversity. It is challenging to create a speech recognition system for a language with low resources. This paper presents a lightweight audio recognition method and a text generation technique for the Sanskrit language. A unique algorithm for recognizing Sanskrit speech-to-text is built using a Sanskrit parser. For input purposes, male and female voices are used for selected Sanskrit words. After taking the voice, the Sanskrit Parser is utilized to generate the relevant text.
Findings: 
This study used an innovative approach to convert Indian language speech to text. Speech-to-text for Indian languages is difficult due to their rich morphology. The pronunciation of one language cannot be matched with another, resulting in a pure case of gender-based language voice diversity. By implementing gender-specific voice input, the system produced a variety of performance results. 
Novelty: The Eighth Schedule lists 22 Indian constitutional languages. Few languages are spoken in many states with various dialects, and the number of speakers is very small or none. Sanskrit is one of the languages with the fewest speakers, yet due to its sophisticated grammar and syntax, it plays an important part in contemporary technical advances with AI. We propose a novel technique for Sanskrit speech-to-text. Because there are so few Sanskrit language speakers, we created our own dataset by selecting 100 words from the Sanskrit lexicon dictionary and utilizing them as voice input, gender-specific.