Automatic Speech Recognition (ASR) systems trained on native English accents often underperform for Indian-accented English due to distinct phonetic and prosodic variations. This study presents a comprehensive, architecture-controlled evaluation of acoustic modeling approaches for Indian-accented English ASR, leveraging a 126-hour annotated speech corpus (91 h from the NPTEL2020 Corpus and 35 h from a manually curated In-House Corpus) and a 45,000-word text corpus spanning six major Indian linguistic regions. We evaluate five acoustic-modeling systems in total: a conventional GMM-HMM baseline, refined progressively through four training stages (Mono-Phone, Tri-Phone, LDA+MLLT, and LDA+MLLT + SAT), and four discriminative neural architectures — DNN-HMM, TDNN, TDNN-BLSTM, and TDNN-LSTM-Attention — using Word Error Rate (WER) as the primary metric, and we additionally report a zero-shot, inference-only open-source Whisper (large-v3) baseline on the identical held-out test set to directly situate the Kaldi hybrid systems against a current open-source state-of-the-art model. All trained systems use identical 118.5-hour training / 7.5-hour test split of the combined corpus to isolate the effect of acoustic-model architecture. Results demonstrate that hybrid models (TDNN-BLSTM) achieve the lowest WER among the trained hybrid systems (8.22
Bishnupriya Manipuri, an Indo-Aryan language, is spoken by more than 120,000 people in Northeast India and Bangladesh and has been designated as a vulnerable language by UNESCO. This language lacks an annotated voice corpus, and no automatic speech recognition system applying deep learning has been established for it. This article confronts both shortcomings by two main contributions: (1) Established a read speech corpus consisting of 2,800 utterances, featuring 28 words and 10 sentences, recorded from 10 native BM speakers within the 20 to 35 age range; and (2) a Convolutional Neural Network-based isolated-word recognizer, trained on 20-dimensional Mel Frequency Cepstral Coefficient attributes, achieving a word recognition accuracy of 92.8 ± 8.6 ± standard deviation across the ten held-out speakers). Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) baselines trained on identical features degrade far more severely under the same speaker-independent protocol (39.3 ± 9.5 ± 7.4
Tamil is a complex language with many different dialects and frequent mixing with English, especially in spoken conversations. Because of these unique features, building reliable AI systems that can understand and generate Tamil speech is particularly challenging. Most current Tamil speech technologies—like automatic speech recognition (ASR), conversation processing, and text-to-speech (TTS)—work separately and often don’t handle the variety of dialects or the natural flow from speech to speech. In this work, we introduce a multitask framework that brings these components together into one seamless system. Our model includes a speech recognition system that’s specifically fine-tuned to handle different Tamil dialects and code-mixed speech. For conversation, we use a smart retrieval system that pulls relevant information from domain-specific documents to answer queries accurately. For speech synthesis, we use an advanced neural TTS model (VITS-based) to produce natural-sounding Tamil speech. Our experiments show that this integrated approach leads to better transcription accuracy for different Tamil dialects and code-mixed speech, demonstrating its effectiveness in real-world scenarios.
The purpose of this study was to evaluate the effectiveness of modern machine learning architectures for the automatic generation of audio descriptions in the field of audiovisual translation. The study was theoretical and applied in nature and was conducted in January–August 2025. The study used the You Described, We Archived (YuWA) corpus of more than 1.2 million segments (2013–2022) and a control subsample of 200 segments according to European Broadcasting Union standards. The corpus analysis revealed syntactic deviations: attribute omissions (35–42
The field of Multilingual Sentiment Analysis (MSA) has made tremendous progress due to deep learning and cross-lingual representation models; yet, the effectiveness of such methods declines considerably in scenarios where grammatical mistakes prevail. Although there is ample proof indicating that grammatical mistakes alter sentiment features, current research on MSA tends to overlook the role of Grammatical Error Detection (GED) in sentiment analysis as a separate preprocessing task. The present systematic literature review brings together 84 peer-reviewed papers from 2020 to 2025 and explores the relationship between MSA and GED in hybrid AI models through a rigorous methodological approach. This approach combines lexicon-based, deep learning, multilingual, and XAI methods to investigate interactions between grammar and sentiment. The study shows that grammar errors affect sentiment interpretation in a structured way and not uniformly as noise and based on the type of grammar error and the model architecture. Incorporating GED into the process increases the robustness of sentiment detection mainly for those that involve critical polarity aspects such as negation and tense. In some cases, indiscriminate correction of grammar errors might remove pragmatically relevant sentiment information. While multilingual embeddings increase the lexical robustness of sentiment detection among different languages, there is a robustness problem with syntax and logical errors especially in code-mixed and low-resource settings. XAI seems to be crucial in debugging sentiment failures because of grammar issues. This review repositions GED from its common use as a preprocessing tool to an aspect of sentiment modulation and suggests Error-Stratified Evaluation Protocol (ESEP) to diagnose robustness in controlled grammar error setting for sentiment detection.
