A Tool for Easily Integrating Grammars as Language Models into the Kaldi Speech Recognition Toolkit

International audienceIntroduction & Motivation Language Models (LMs) represent a crucial component in the architecture of hybrid Automatic Speech Recognition (ASR) systems, as far as the linguistic regularities that they describe guide the prediction of the most likely sequence of uttered words (Adda-Decker and Lamel, 2000). An important interest in LM design has been cultivated in the last few years. It is not in vain that we have witnessed the transition from statistical models into neural-based approaches, which have proven to be a solid strategy for capturing deeper lexical and semantic representations (Naseem et al., 2021). Current trends in ASR point to the creation of high-performing and increasingly robust systems thanks to the exploitation of data-driven approaches, the continued improvements in computing infrastructure and the sophistication of new Deep Learning techniques (Huang et al., 2014). This suggests that the implementation of grammars and the role of formal approaches, which constitute an important precedent for the later development of LM resources, seems to be questionable in the context of NLP-related tasks. However, their use may be advantageous in some of today's ASR applications, especially when an efficient control of the generated hypotheses is needed. Providing a deliberately constrained transcription can be more easily achieved using formal-based models, where the use of unseen rules in the training data is not allowed (Post and Gildea, 2009), so that only the utterances that can be produced by the grammar may be output. Unlike probabilistic models, grammar-based approaches favor the direct injection of knowledge into LMs and thus a broader span of lexical, semantic, and syntactic constraints between words. This may be of high interest in settings where the quality of the ASR system is particularly dependent on the correct recognition of semantically and grammatically sound constructions, as can be observed in speech-enabled medical translation devices. Due to the criticality of a correct transcription in such contexts (Dew et al., 2018), a natural language representation by means of grammars seems convenient for producing only reliable outputs. Moreover, the use of these resources may prove to be an inexpensive palliative solution to building LMs for domains where there are subject-matter experts to help encode grammars but not enough corpora to infer a LM from