Robust automatic continuous speech recognition for 'Adi', a zero-resource indigenous language of Arunachal Pradesh

This article depicts an automatic speech recognition system (ASR) for the continuous speech of 'Adi,' a zero-resourced endangered indigenous language of Arunachal Pradesh, India. This ASR system uses a speech corpus of 40 native Adi speakers of Arunachal Pradesh. Mel frequency cepstral coefficients (MFCC) features extracted from recorded Adi speech samples. Different speech recognition models, such as Sub Space Gaussian Mixture Model (SGMM), Monophone, and Triphone (tri1, tri2, tri3), were applied in this ASR system. The monophone model's word recognition accuracy (WRA) was 58.5%. In triphone models, the recognition efficiency of tri1, tri2, and tri3 was enhanced at 78.18%, 83.08%, and 88.62%, correspondingly. SGMM was the most proficient model in this ASR system, with a minimum word error rate (WER) of 10.12%. This proffered ASR model may be beneficial in the physical world to set up various ASR applications of man-machine interfaces in the Adi language.