Articulatory features describe the properties of speech production, i.e., each sound unit of a language (phone) can be decomposed into a set of features based on the articulators used to produce the sound. Articulatory features (AFs) have been used for automatic speech recognition (ASR) with the aim of better pronunciation modeling, robustness to noise, multilingual and cross-lingual portability of the systems etc. In case of text-to-speech (TTS) conversion systems AFs are explored to achieve emotional speech synthesis. ASR or TTS using AFs poses three main challenges: firstly, the type of AF representation; secondly, the estimation of AFs from the acoustic signal; and finally, the integration into conventional hidden Markov model (HMM) based ASR framework.
The goal of this presentation is to present recent advances made at Idiap Research Institute in addressing the above mentioned challenges in the context of ASR, and open the discussion on natural extension of the work presented here such as, graphemebased ASR, generation of AF based dictionaries, templatebased ASR using AFs, use of deep learning approaches to improve AF estimation].
Cite as: Magimai-Doss, M., Rasipuram, R. (2013) On using articulatory features in posterior-based ASR system: representation, estimation and integration. Proc. Speech Production in Automatic Speech Recognition (SPASR-2013), 22-23
@inproceedings{magimaidoss13_spasr,
author={Mathew Magimai-Doss and Ramya Rasipuram},
title={{On using articulatory features in posterior-based ASR system: representation, estimation and integration}},
year=2013,
booktitle={Proc. Speech Production in Automatic Speech Recognition (SPASR-2013)},
pages={22--23}
}