A Speaker-Independent Hybrid ASR System for the Under-Resourced Nyishi Language Using Self-Supervised Feature Representations

The presented work introduces the first speaker-independent hybrid Automatic Speech Recognition (ASR) system for the Nyishi language of Arunachal Pradesh, which is an under-resourced language belonging to the Tani subgroup of the Tibeto-Burman language family. Nyishi presents significant challenges in the development of ASR because of a scarcity of annotated speech data, the use of nonstandard orthography, and limited digital resources. To address these challenges, a read-speech corpus comprising approximately two hours of speech from six native speakers was developed, with four speakers used for training and two previously unseen speakers used for testing. All utterances were manually transcribed and verified to ensure transcription accuracy. A comparative analysis was performed using traditional Mel-Frequency Cepstral Coefficients (MFCC) and the self-supervised speech representations of the HuBERT and XLS-R models. Hybrid DNN-HMM models were implemented using Time Delay Neural Networks (TDNN) and factorized TDNN (TDNN-F) acoustic models with consistent triphone GMM-HMM alignments across all feature representations. Experimental results indicate that self-supervised representations consistently outperform MFCC-based systems on the Nyishi ASR task. MFCC-TDNN and MFCC-TDNN-F had a Word Error Rate (WER) of 24.71% and 21.47%, respectively, as compared to HuBERT-TDNN-F and XLS-R-TDNN-F with a much lower WER of 14.85% and 15.33%, respectively. Statistical significance analysis using paired bootstrap resampling confirmed that the improvements of the SSL-based systems over the MFCC baseline were statistically significant. These findings establish an initial benchmark for speaker-independent Nyishi ASR and demonstrate the effectiveness of self-supervised speech representations under the low-resource conditions investigated in this study.