Enhancing Data-Driven Phone Confusions Using Restricted Recognition

Abstract

This paper presents a novel approach to address data sparseness in standard confusion matrices and demonstrates how enhanced matrices, which capture additional similarities, can impact the performance of spoken term detection. Using the same training data as for the standard phone confusion matrix, an enhanced confusion matrix is created by iteratively restricting the recognition process to exclude one acoustic model per iteration. Since this results in a greater amount of confusion data for each phone, the enhanced confusion matrix encodes more similarities. The enhanced phone confusion matrices perform demonstrably better than standard confusion matrices on a spoken term detection task which uses both HMMs and DNNs.

Domain-Informed Probing of wav2vec 2.0 Embeddings for Phonetic Features

“…O'Neill and Carson-Berndsen (2019) demonstrate that embeddings derived purely from text using a grapheme-to-phoneme mapping and applying a word2vec approach exhibit similarity between phoneme classes. These phoneme embeddings were subsequently integrated with the data-driven acoustic similarities of Kane and Carson-Berndsen (2016) to generate a similarity matrix for use in phonemically driven spell checking (O'Neill et al, 2021).…”

The Influence of Regional Pronunciation Variation on Children’s Spelling and the Potential Benefits of Accent Adapted Spellcheckers

“…The costs of these operations are taken from a phoneme distance matrix that models the similarity between phonemes. Since similarity can be considered a function of confusability (Gallagher and Graff, 2012), this distance matrix was generated based on the confusability of phonemes both acoustically and distributionally (Kane and Carson-Berndsen, 2016;O'Neill and Carson-Berndsen, 2019). If two phonemes are likely to be confused, then they are considered highly similar and thus have a low distance score and low substitution cost.…”