r-G2P: Evaluating and Enhancing Robustness of Grapheme to Phoneme Conversion by Controlled noise introducing and Contextual information incorporation

TL;DR

This paper evaluates the vulnerability of neural G2P models to orthographical noise and proposes methods to improve their robustness by introducing controlled noise and leveraging contextual information, resulting in significantly better performance.

Contribution

It introduces three noise-adding techniques and a robust training strategy to enhance G2P model resilience against spelling variations and errors.

Findings

Robust G2P outperforms baseline by -2.73% WER on benchmarks.

Significant improvement of -9.09% WER on real-world data.

Neural G2P models are highly sensitive to orthographical variations.

Abstract

Grapheme-to-phoneme (G2P) conversion is the process of converting the written form of words to their pronunciations. It has an important role for text-to-speech (TTS) synthesis and automatic speech recognition (ASR) systems. In this paper, we aim to evaluate and enhance the robustness of G2P models. We show that neural G2P models are extremely sensitive to orthographical variations in graphemes like spelling mistakes. To solve this problem, we propose three controlled noise introducing methods to synthesize noisy training data. Moreover, we incorporate the contextual information with the baseline and propose a robust training strategy to stabilize the training process. The experimental results demonstrate that our proposed robust G2P model (r-G2P) outperforms the baseline significantly (-2.73\% WER on Dict-based benchmarks and -9.09\% WER on Real-world sources).