Neural models for Japanese pronunciation estimation often suffer from errors such as hallucinations (generating pronunciations that are not grounded in the input) and omissions (skipping parts of the input). Although attention-based alignment has been used to detect such errors, selecting reliable attention heads is difficult, and developing methods that can both detect and correct these errors remains challenging.

In this paper, we propose a simple method called existence-based alignment check. In this approach, we consider alignment candidates independently extracted from all attention heads, and check whether at least one of these candidates satisfies two conditions derived from the linguistic properties of Japanese pronunciation: monotonicity and pronunciation length per character. We generate multiple hypotheses using beam search and use the alignment check as a filtering mechanism to correct hallucinations and omissions.

We apply this method to a dataset of Japanese facility names and demonstrate that it improves pronunciation estimation accuracy by over 2.5\%.