MagpieTTS-LF: Inference-Time Long-Form Speech Generation Without Training on Long-Form data
Researchers have developed MagpieTTS-LF, a novel approach to generating long-form speech with improved coherence and consistency. This method allows the existing MagpieTTS system to produce extended audio without requiring retraining on long-form data. Key innovations include soft attention priors for better alignment, a stateful inference algorithm to maintain prosodic continuity across sentence boundaries, and text encoding that considers past context for discourse-level prosody. AI
IMPACT This research could lead to more natural and coherent long-form speech synthesis for applications like audiobooks and podcasts.