Developing an AI-Powered Tool for Real-Time Multilingual Speech Transcription

Thesis Abstract

The increasing global interconnectedness and multicultural communication necessitate efficient and accurate transcription of multilingual speech in real-time, addressing challenges such as language barriers, diverse phonetic structures, and speech variability. Despite advancements in speech recognition technologies, current systems often exhibit limitations in supporting multiple languages simultaneously with high precision, particularly in noisy environments or when handling dialectal variations. This study aims to develop an AI-powered tool capable of real-time multilingual speech transcription, thereby enhancing communication efficiency across diverse linguistic contexts. The specific objectives include designing an integrated speech recognition framework leveraging deep learning models, evaluating the system's transcription accuracy across selected languages, and assessing its usability in real-world settings. Employing a mixed-methods research design, the study integrates quantitative model evaluation with qualitative usability assessment. The population comprises 300 multilingual speakers across three major languages—English, Mandarin, and Swahili—recruited from urban linguistic research centers. A stratified random sampling technique was applied to ensure representative inclusion of dialectal variants and speech contexts. Data collection involved recording speech samples under controlled laboratory conditions and real-world environments, using standardized speech elicitation protocols. The developed machine learning models, primarily convolutional neural networks (CNNs) and transformer architectures, form the core of the transcription system, trained on a corpus of 10,000 annotated speech recordings per language, sourced from publicly available datasets and proprietary collections. The performance of the AI transcription tool was evaluated through metrics such as Word Error Rate (WER), Sentence Error Rate (SER), and Bleu scores, using cross-validation techniques. Additionally, user experience and system usability were examined through thematic analysis of feedback obtained via semi-structured interviews with 50 end-users, employing NVivo software for qualitative data coding. Descriptive statistics summarized transcription accuracy, while inferential statistics, including ANOVA and multiple regression analysis, examined factors influencing system performance and user satisfaction. It is anticipated that the developed model will achieve a transcription accuracy with a WER below 15% for each supported language, outperforming existing multilingual speech recognition systems by at least 10%. The findings are expected to demonstrate significant improvements in processing speed and contextual comprehension, especially in challenging acoustic environments. The qualitative analysis aims to reveal key usability features and user perceptions, influencing future refinements of the tool. This research contributes to the growing body of knowledge at the intersection of speech processing, machine learning, and multilingual communication technology. It extends theoretical frameworks such as the Speech Chain Model and the Multilingual Deep Learning Theory to explain the system's capability to adaptively recognize and transcribe speech across languages and dialects. It also provides empirical evidence on the feasibility and effectiveness of integrating advanced neural network architectures for real-time transcription in multilingual settings. The study concludes that AI-driven multilingual speech transcription tools can significantly enhance cross-cultural communication, linguistic accessibility, and digital inclusivity. It recommends continued refinement of the models to incorporate more low-resource languages, expansion of the speech corpus to better handle variability, and deployment in diverse real-world applications such as international conferencing, language learning, and assistive technologies. Future research directions include exploring multimodal inputs, such as lip-reading and contextual cues, to further improve accuracy in adverse acoustic conditions.