Low-Resource TTS Synthesis

Low-resource text-to-speech (TTS) aims to produce intelligible, natural speech for languages and dialects with few recordings, often including tonal languages and non-Latin scripts. Common approaches combine curated speech corpora with cross-lingual transfer from multilingual or foundation models, parameter-efficient fine-tuning, and synthetic data or pseudo-labeling; text normalization and language-specific phoneme, tone, and prosody modeling are also important. Evaluation typically combines listening tests such as mean opinion scores with intelligibility, speaker-similarity, and acoustic measures, increasingly testing long or complex text and out-of-domain use. More data and larger pretrained models can improve quality, but synthetic-data errors, weak language coverage, and trade-offs among intelligibility, expressivity, and speaker identity remain open challenges.