Incorporating alignments into Conditional Random Fields for grapheme to phoneme conversion

Abstract

Conditional Random Fields (CRFs) are a state-of-the-art approach to natural language processing tasks like grapheme-tophoneme (g2p) conversion which is used to produce pronunciations or pronunciation variants for almost all ASR pronunciation lexica. One drawback of CRFs is that for training, an alignment is needed between graphemes and phonemes, usually even 1-to-1. The quality of the g2p result heavily depends on this alignment. Since these alignments are usually not annotated within the corpora, external models have to be used to produce such an alignment in a preprocessing step. In this work, we propose two approaches to integrate the alignment generation directly and efficiently into the CRF training process. Whereas the first approach relies on linear segmentation as starting point, the second approach considers all possible alignments given certain constraints. Both methods have been evaluated on two English g2p tasks, namely NETtalk and Celex, on which state-of-the-art results have been reported in the literature. The proposed approaches lead to results comparable to the state-of-the art.

Formulation of Syllable Based Pronunciation Models for Tamil Text-to-Speech Synthesizer

“…CRF can handle overlapping context of features with a relative cohesion in an excellent way Apparently, to train CRF a training data with appropriate alignment and a template with details on the contextual frame size is required. One detriment in using CRF is the need of a monotonic alignment between the source and target symbol for resolving the one-to-many alignment [39]. This alignment process is not an integral part of a CRF and should be fabricated separately.…”

Hidden conditional random fields with M-to-N alignments for grapheme-to-phoneme conversion

“…Hidden Dynamic Conditional Random Fields [13] (HD-CRFs). In [14] it is shown that the computational cost for calculating a HCRF is not significantly changed compared to a CRF with given Alignment, as the numerator is a subset of the denominator. In the latter publication, the alignment is realized by adopting the BIO scheme [6], where the abbreviation stands for "begin" (B), "inside" (I), and "outside" (O) markers (cf.…”

Deep Bidirectional Long Short-Term Memory Recurrent Neural Networks for Grapheme-to-Phoneme Conversion Utilizing Complex Many-to-Many Alignments

“…In [10], a joint n-gram model is combined with a CRF model. In [11], many-to-1 alignment constraints are used with CRF models. However, in [12,7], many-to-many alignments are used.…”