Character Embedding Priors and Script Structure in Low-Resource Bangla Text-to-Speech Adaptation

Character-level text-to-speech (TTS) models pretrained on Latin-script data are usually adapted to a new language by adding its characters to the vocabulary. We ask why this works poorly for Bangla, an abugida, when F5-TTS is fine-tuned on one hour of speech for 3,000 updates: native Bengali characters give a character error rate (CER) of 67.96%, under an ASR model that scores 1.93% on real recordings. Controlled input representations separate two factors. On identical phonemic Latin text, pretrained embedding rows lower CER by 22.04 percentage points (pp) relative to freshly initialized rows, while assigning the same rows to the wrong sounds raises it by 51.53 pp. With fresh rows, a lossless re-encoding of Bengali text (explicit inherent vowel, merged vowel forms, split aspirates) lowers CER by 47.56 pp, and pretrained rows help this restructured encoding less than a standard transliteration. All five tests specified before the runs replicate on a second corpus of another Bangla variety, and, where tested, the directions hold across three random draws of the new embeddings; the effect sizes, however, vary strongly with the draw. In an exploratory run with five times more updates, the gap between the two Latin representations shrinks to 0.54 pp, so the findings concern adaptation under a limited training budget.