Slungt: Even Faster Spoken Language Understanding with N-Grams and Tries

Abstract

In the domain of Spoken Language Understanding (SLU) the primary objective is to extract important information from audio commands, like the intent of what a user wants the system to do and specific entities like locations or numbers. This paper presents a simple method that integrates intents and entities into a beam search algorithm, and, in combination with a general-purpose Speech-to-Text model, enables the creation of customized SLU-decoders without any additional training. Constructing such decoders is very fast and only takes a few seconds. It is also completely language-independent. In comparative assessments across multiple benchmarks, this method demonstrates comparable performance to several other SLU strategies, while significantly surpassing them in terms of computational speed.