Reproduce “Speaking Longer Than Learned: Training-Free Long-Form Generation for Full-Stream Text-to-Speech” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jeseung Lim, Jiseong Woo, Heeseung Kim

The authors modify pretrained VoXtream2 for continuous long-form speech by shortening its context window and increasing attention to the acoustic prompt. On LibriSpeech-Long, these training-free changes reduce overall word error rate from 8.63% to 3.58% and raise speaker similarity from 0.659 to 0.759. Mean opinion scores remain 4.03, comparable to the original; however, WER remains higher than long-form-trained baselines.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.