Effect of language independent transcribers on spoken language identification for different Indian languages

Abstract

Automatic spoken language identification systems are usually built by extracting suitable acoustic features from speech samples, or by using language dependent speech-to-text/phone synthesizers. Acoustic features are vulnerable to recording environments and often affect the identification performance. At the same time, building language dependent speech to text synthesizers is expensive and often not feasible for low resource languages. Instead of using language dependent speech to text transcriber, this paper investigate the effect of using language independent speech-to-text transcribers (trained on languages different from target languages) on language identification for various Indian languages. The language independent transcriber transcribes the audio samples without prior knowledge of the input language. From various experimental setups over speech samples recorded under controlled and uncontrolled environment, it is evident that transcription text using language independent transcriber can be effectively used for spoken language identification tasks.