Repetition, Not Length: Isolating the Counting Failure in Neural Text-to-Speech

2026-09-29

AI Summary

This study addresses the failure modes of neural text-to-speech (TTS) models—specifically looping, truncation, and counting errors—when processing repetitive phrases. We introduce a novel analytical framework that decouples textual repetitiveness from sequence length. Through paired controlled-set testing, multi-architecture benchmarking, and cross-validation across multiple automatic speech recognition (ASR) systems, we demonstrate that textual repetitiveness, rather than length, is the primary driver of model failure, further revealing a smoothing effect of periodicity on accuracy. Our analysis quantifies a precipitous drop in accuracy to 18.2% under highly repetitive conditions and shows that the proposed framework can predict the performance of emerging architectures with an error margin within one percentage point.

Abstract

Text-to-speech models loop, truncate and lose count on text that repeats a phrase many times. We show that repetition itself is what breaks them, not the length that comes with it. Every repeated sentence in our test set is paired with a control of matched sentence and word count in which no word ever repeats back-to-back. Six models from three architectures render the controls almost perfectly and fail the repeated twins: 94.3% against 18.2% exactly right at k >= 6. The gap survives greedy decoding, repetition-penalty sweeps, four independent speech recognisers and 420 analysis specifications without once reversing sign; a held-out fourth architecture lands within a point of its predicted gap, and one of two non-autoregressive baselines shows the same failure. Varying the period of the text shows the failure grows smoothly with periodicity, half of it surviving when no word is adjacent to itself.

Problem

• Text-to-Speech
• Repetition Failure
• Counting Failure
• Neural TTS
• Periodicity

Innovation

• Text-to-Speech
• Repetition Failure
• Counting Mechanism
• Periodicity
• Controlled Experiment