10.3390/polym8030084

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Low-resource written-query keyword spotting detects a text-specified target in speech without spoken enrollment or full automatic speech recognition. We present Tibetan Phonology-Aware Speech Evidence Matching (Tibetan-PASEM), a method that encodes graphemic, approximate phonological, and dialect-related query information, matches it with local acoustic windows, aggregates window-level evidence, and applies a validation-selected operating point. Querybank198 contains 198 Tibetan written queries derived from four public speech resources. A complete-file and decoded pulse-code modulation (PCM) audit produced a corrected 213,128-pair evaluation with no identified recording-content, source-qualified speaker, or session-proxy overlap across the development-to-held-out boundary. In a three-seed corrected-index re-evaluation, the acoustically strengthened PASEM (PASEM-AS) achieved F1 scores of 66.16 ± 2.14% on seen_test, 26.51 ± 3.42% on unseen_test, and 72.10 ± 1.39% on the confusable-negative diagnostic condition, which comprises 20 training-exposed query forms and is used for negative-set analysis rather than difficulty ranking. Under the original pre-correction pair index, PASEM with transfer regularization (PASEM-TR) increased the unseen_test F1 from 25.80 ± 2.71% to 35.25 ± 4.17% and the Recall from 18.48 ± 2.98% to 33.85 ± 8.22%; the corresponding seed-level 95% confidence intervals were 24.89-45.61% and 13.43-54.27%. These results establish strong familiar-query discrimination, an audited fixed-threshold evaluation protocol, and partial, seed-sensitive transfer to held-out written queries.

Abstract: PubMed · Datensatz