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

2026-06-16

AI Summary

This work addresses the challenges of prosodic drift, speaker inconsistency, and sentence-boundary artifacts that commonly arise in long-form speech synthesis with neural text-to-speech systems. To enable MagpieTTS to generate coherent extended utterances without retraining, the authors propose a novel inference-time approach featuring three key innovations: a soft attention prior to encourage monotonic alignment while preserving cross-sentence context, a stateful inference algorithm to maintain prosodic continuity across segments, and a history-aware text encoding mechanism that supports discourse-level prosody planning. Experimental results demonstrate that the proposed method substantially improves long-range intelligibility, prosodic coherence, speaker consistency, and naturalness at sentence boundaries in synthesized speech.

Abstract

Neural Text-to-Speech (TTS) systems achieve remarkable quality on short utterances but long-form speech generation shows prosodic drift, speaker inconsistencies and sentence boundary artifacts. Existing approaches either compress sequences, increase context length or naively concatenate independently synthesized chunks. We present an inference-time approach called MagpieTTS-LF that enables MagpieTTS to produce coherent long-form speech without model retraining. Our method introduces three key innovations: (1) soft attention priors to guide monotonic alignment while preserving past and future context; (2) a stateful inference algorithm that maintains context across sentence chunks, ensuring prosodic continuity; (3) history-aware text encoding that uses past text for discourse-level prosodic planning. Experiments on long texts show significant improvements in long-range intelligibility, prosodic coherence, speaker consistency, and boundary naturalness compared to other baselines.

Problem

• long-form speech generation
• prosodic drift
• speaker inconsistency
• sentence boundary artifacts
• neural TTS

Innovation

• long-form TTS
• inference-time adaptation
• prosodic coherence
• stateful inference
• history-aware encoding