SEAM Cuts Low-Resource TTS Training Data Needs by 75% via Spectral Embedding Initialization
ITMO · hf · 2026-10-23
Researchers at ITMO University propose SEAM (Spectral Embedding Alignment & Merging), a zero-audio initialization method for expanding vocabulary in conditional TTS models.
- Treats subtoken embeddings as discrete signals and applies spectral low-pass filtering to preserve temporal phonetic compositionality
- Projects results onto the pretrained embedding manifold via closed-form ridge regression, grounded in Lipschitz continuity of conditional generation
- Reduces training data requirements by 75% for low-resource and cross-lingual TTS vocabulary expansion
More from Research
- TokenGate e2e-learns to drop tokens on the fly, opening a new frontier for LLM prefill — CShorten30 · 2026-10-23
- ProtTale learns to self-assess reliability of AI-generated protein function descriptions — anshulkundaje · 2026-10-23
- Tübingen study: top-down feedback only robustifies vision models when paired with dropout — maier_ak · 2026-10-23
- Study: LLMs chime in at unwelcome moments 46.4% of the time in group chats — windx0303 · 2026-10-23
- Google Cloud's Physical-Engineering Last Exam: all frontier models score under 50% — Afinetheorem · 2026-10-23
- 'Wait' markers are inherited from training data, not invented by RL, HKUST study finds — huawei-noah · 2026-10-23