Conceptual

Data-Driven Text Pronunciation Correlation from Speech Embeddings in Speech Recognition

A method that measures how similarly different written symbols are pronounced by comparing the speech embeddings of their spoken occurrences, rather than relying on a manually designed pronunciation lexicon. It aligns text to speech, embeds each symbol occurrence with a speech encoder, and computes symbol-to-symbol pronunciation correlations from embedding distances; the resulting correlations improve end-to-end speech recognition under contextual biasing and transfer to languages lacking lexicons.