TABR: Tone-Auxiliary Base-Syllable Rescoring for Low-Resource Tujia Speech Recognition

Tujia is a low-resource endangered language, and its digital documentation and preservation are of considerable importance. The scarcity of transcribed speech data leads to low recognition accuracy in Tujia automatic speech recognition (ASR). We propose Tone-Auxiliary Base-syllable Rescoring (TABR), a two-stage framework built on XLS-R-300M. In Stage I, complete tone-marked Tujia Pinyin syllables are the primary recognition targets, while transcription-derived tone labels provide auxiliary supervision to the shared encoder. In Stage II, an independently trained base-syllable recognizer scores the tone-stripped forms of first-pass hypotheses to rerank the original candidates. Neither derived representation requires additional manual annotation. On a coverage-constrained split of a 2.485 h corpus, TABR reduces syllable error rate from 33.53% to 30.89% and Pinyin character error rate from 17.34% to 15.54% compared with the Baseline, with all results averaged over three random seeds. Paired bootstrap analyses support both improvements. Component controls compare auxiliary targets and second-pass representations. Lower mean errors are also observed after retraining on a split without coverage reassignment and when applying models trained on the original split to one unseen speaker within the same book-derived condition. These findings support TABR under the evaluated conditions.