Copyright © 2025 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{185536,
        author = {P Bhanumathi and N V Muthu Lakshmi},
        title = {MULTILINGUAL DEEP LEARNING FOR ENHANCED BLOCK DETECTION IN BILINGUAL STUTTERING SPEECH},
        journal = {International Journal of Innovative Research in Technology},
        year = {2025},
        volume = {12},
        number = {5},
        pages = {1972-1976},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=185536},
        abstract = {Millions of people all over the world suffer from the common speech disorder stuttering. This paper aims at detecting the occurrence of blocks in bilingual stuttering speech by proposing a multilingual Long Short-Term Memory (LSTM) framework. The main idea was to use clip-level Fluency Bank-style labels (Block, Prolongation, Sound/Word Repetition, and Interjection) for the framework to fuse self-supervised multilingual acoustic embeddings with weakly supervised multiple instance learning and a conditional random field decoder. In this work, a preliminary study on the annotations provided is done in order to uncover the imbalance of label frequencies and the presence of co-occurrence patterns. The use of LSTM (Long Short-Term Memory) as a multilingual model turns out to greatly improve recognition results because of the language independency, nevertheless, the model still confronts problems of linguistic diversity and data. Its design allows it to cope with the temporal dependencies very efficiently and also extract the features clearly, thus helping it to recognize the stuttering patterns even in different languages and, consequently, to be applicable in the field of multilingual speech analysis and stuttering intervention. The goal of this work is to find the different stuttering behaviors happening concurrently, including blocks; therefore, a sole LSTM model was chosen to be trained on multiple languages spoken by young children. This paper serves as a record of the preprocessing, feature extraction, imbalance mitigation, and a reproducible training/evaluation protocol that can be used for benchmarking purposes.},
        keywords = {Stuttering, Block Detection, Bilingual Speech, Multilingual ASR, Self-Supervised Learning, MIL, CRF, Fluency Bank.},
        month = {October},
        }