Abstract This work presents a comprehensive review of the Viterbi algorithm's applications in pattern recognition and natural language processing (NLP). Originally developed for decoding convolutional codes, the Viterbi algorithm has become a cornerstone technique for maximum a posteriori (MAP) sequence estimation in Hidden Markov Models (HMMs). The paper outlines its theoretical foundation as a dynamic programming method that efficiently identifies the most probable sequence of... This work presents a comprehensive review of the Viterbi algorithm's applications in pattern recognition and natural language processing (NLP). Originally developed for decoding convolutional codes, the Viterbi algorithm has become a cornerstone technique for maximum a posteriori (MAP) sequence estimation in Hidden Markov Models (HMMs). The paper outlines its theoretical foundation as a dynamic programming method that efficiently identifies the most probable sequence of hidden states given observed data. It then explores numerous domain specific adaptations such as modified trellis structures, context aware decoding, beam search pruning, and finite state machine optimisations that enhance performance across diverse tasks. These applications span distorted pattern recognition, named entity recognition (NER), stochastic grammar parsing, sign and gesture recognition, contextual text recognition, language testing, letter to phoneme conversion, and multilingual translation. In each domain, Viterbi based approaches consistently outperform traditional baselines by improving accuracy, reducing computational complexity (often from exponential or cubic to linear or quadratic time), and increasing robustness to noise and temporal distortions. The review also highlights recent innovations integrating Viterbi decoding with deep learning, parallel computing (e.g., GPU acceleration), and multimodal inputs. A unified conceptual framework illustrates how multimodal inputs are processed through feature extraction and probabilistic modelling before Viterbi based decoding yields optimised decisions. Despite its maturity, the algorithm remains highly relevant due to its optimality, interpretability, and efficiency. However, challenges persist including scalability in large state spaces, integration with modern neural architectures, and the need for explainable, user centric decoding. The paper concludes that the Viterbi algorithm is not obsolete but evolving, with ongoing research poised to further extend its utility in intelligent, real time systems. Read More Read Less

Abstract This paper proposes a novel Speech to Phoneme to Text (SPG) framework that integrates speech recognition with Large Language Models (LLMs) through phoneme level intermediation. This approach aims to enhance LLMs' ability to process spoken input accurately, especially in challenging conditions such as accented speech, background noise, or low resource languages, while enabling real time, multimodal, and speaker aware voice interaction. The system comprises four... This paper proposes a novel Speech to Phoneme to Text (SPG) framework that integrates speech recognition with Large Language Models (LLMs) through phoneme level intermediation. This approach aims to enhance LLMs' ability to process spoken input accurately, especially in challenging conditions such as accented speech, background noise, or low resource languages, while enabling real time, multimodal, and speaker aware voice interaction. The system comprises four core components: (1) a Speech to Phoneme (S2P) module, built on self supervised models like Wav2Vec 2.0 or Wav2VecBERT 2.0, fine tuned to output International Phonetic Alphabet (IPA) sequences; (2) a Phoneme to Text (P2G/LLM) module, which uses multilingual LLMs (e.g., mT5) to convert phonemes into fluent text, leveraging contextual understanding for better disambiguation; (3) a Voice Recognition Enhancer, integrating speaker embeddings (e.g., from ECAPA-TDNN) for diarization and personalization; and (4) innovative training and inference strategies including Data Augmentation with Noisy Phonemes (DANP), Top-K Marginalized (TKM) training, LoRA based fine tuning, and delayed fusion to reduce latency and computational load. Evaluation targets include <5% Word Error Rate (WER) on CommonVoice, <10% Phoneme Error Rate (PER), <3.5 tokens/sec latency, and <5% speaker verification Equal Error Rate (EER). The framework is validated using datasets such as CommonVoice, TIMIT, and LibriSpeech, as well as a multilingual PR corpus. TF-IDF analysis of this corpus reveals language specific expressions of concepts such as "sustainability," underscoring the need for semantic, rather than lexical, alignment in multilingual ASR. The SPG architecture advances accessibility, cross lingual transfer, and real world robustness while reducing token usage by 86% compared to direct audio to LLM baselines. Read More Read Less

Abstract This study explores the integration of cognitive science and artificial intelligence to enhance English writing instruction and assessment. It frames writing as a complex, recursive cognitive process involving planning, generation, and revision mirroring functional mechanisms in modern AI systems, such as large language models (LLMs). The paper proposes a conceptual framework that aligns human cognitive models with AI architectures through components such as goal... This study explores the integration of cognitive science and artificial intelligence to enhance English writing instruction and assessment. It frames writing as a complex, recursive cognitive process involving planning, generation, and revision mirroring functional mechanisms in modern AI systems, such as large language models (LLMs). The paper proposes a conceptual framework that aligns human cognitive models with AI architectures through components such as goal representation, memory (long-term and working), attentionbased context handling, and feedback driven revision. This alignment is formalized mathematically using state transition systems and probabilistic generation models. An intelligent writing system was implemented in a 12-week online doctoral course (N=22) to support idea generation and coherence development. Text coherence was measured via average degree centrality in graph based representations. Experimental results showed that the proposed feature selection method significantly outperformed LSTM across training set sizes (85-99% vs. 40-90% accuracy), particularly with limited data. Adversarial training further improved robustness. Frequent subgraph analysis enabled effective discrimination between coherent and incoherent texts, with low performing subgraphs filtered to preserve reliability. The study concludes that hybrid AI-cognitive models enhance writing quality, engagement, and efficiency while underscoring the need for ethical, explainable, and human centred AI design in educational contexts. Read More Read Less