MagpieTTS-LF: Inference-Time Long-Form Speech Generation Without Training on Long-Form data

A new inference-time method, MagpieTTS-LF, enables the MagpieTTS neural speech model to generate coherent long-form audio without retraining, addressing persistent issues of prosodic drift and speaker inconsistency that plague conventional text-to-speech systems [1]. Neural Text-to-Speech systems deliver high-quality output for short utterances but degrade when synthesizing longer passages, exhibiting prosodic drift, speaker inconsistencies, and unnatural sentence boundaries [1][2]. Standard workarounds compress sequences, expand context windows, or naively concatenate independently synthesized chunks, each introducing its own artifacts [1][2]. The MagpieTTS-LF approach, detailed in a paper submitted to the arXiv preprint repository on June 16, 2026, avoids these trade-offs by operating entirely at inference time [1]. arXiv, which began on August 14, 1991, now hosts over two million e-prints and receives roughly 24,000 submissions per month, serving as a primary distribution channel for computer science and engineering research [6]. The paper introduces three components: soft attention priors that guide monotonic alignment while preserving surrounding context, a stateful inference algorithm that carries acoustic context across sentence chunks to maintain prosodic continuity, and history-aware text encoding that leverages preceding text for discourse-level prosodic planning [1][2]. Experiments reported in the paper show significant gains in long-range intelligibility, prosodic coherence, speaker consistency, and boundary naturalness compared to baseline methods [1][2]. The work appears within arXiv’s Sound (cs.SD) category, and the abstract page links to community-built tools such as the Bibliographic Explorer and Connected Papers, which are part of the arXivLabs framework launched in 2020 to foster third-party innovation on top of the repository [4][5]. arXivLabs projects operate under guidelines that require partners to uphold openness, community engagement, and user data privacy, and collaborators receive only minimal, anonymized data necessary for their tools to function [4]. The MagpieTTS-LF paper has not undergone formal peer review, consistent with arXiv’s role as a moderated but not peer-reviewed preprint platform [6].