Identification and classification of speech disfluencies: A systematic review on methods, databases, tools, evaluation and challenges
Alana S. Luna, Ariane Machado‐Lima, Fátima L. S. Nunes · Journal of the Brazilian Computer Society · 2025
With the advancement of multimedia technologies, human-computer conversational interfaces are becoming increasingly important and are emerging as a highly promising area of research. Vocal representations, facial expressions, and body language can be used to extract various types of information. In the context of vocal representations, the complexity of human communication involves a wide range of expressions that vary according to grammatical rules, languages, accents, slang, disfluencies, and other speech events. In particular, the detection of disfluencies, i.e., interruptions in the normal flow of speech characterized by pauses, repetitions, and sound prolongations, is of interest not only for improving speech recognition systems but also for potentially identifying emotional aspects in audio. Several studies have aimed to define computational methods to identify and classify disfluencies, as well as appropriate evaluation methods in different languages. However, no studies have compiled the findings in the literature on this topic. This is important for both summarizing the motivations and applications of the research, as well as identifying opportunities that could guide new investigations. Our objective is to provide an analysis of the state of the art, the main limitations, and the challenges in this field. Eighty articles were extracted from four databases and analyzed through a systematic review. Our results show that research into the detection of disfluencies has been conducted for various purposes. Some aimed to improve the performance of translation tools, while others focused on the summarization of spoken dialogues, speaker diarization, and Natural Language Processing. Most of the research was oriented toward the English language. F-score, precision, and recall were the most commonly used evaluation measures for the reported methods. Statistical and machine learning techniques were widely applied, with CRFs (Conditional Random Fields), MaxEnt (Maximum Entropy), Decision Trees, and BLSTM (Bidirectional Long Short-Term Memory) being especially prominent. In general, newer approaches, such as BERT and BLSTM, have demonstrated higher performance. However, several challenges remain, opening up new research opportunities.