We present AURA-ST, a three-stage modular pipeline for low-resource speech-to-text translation submitted to the IWSLT 2026 African-Celtic Track 1. The architecture bypasses traditional cross-attention between audio and text modalities by treating projected acoustic representations as a native token prefix to a frozen large language model. A dual-stream encoder captures linguistic and paralinguistic features via a jointly trained semantic and a paralinguistic encoder. A convolutional subsampler then bridges the modality gap through a 4x temporal compression and a linear projection into the LLM embedding space. Finally, a MLP-targeted Low-Rank Adaptation adapter fine-tunes the frozen Gemma-4-E2B backbone for translation without catastrophic forgetting of base language model knowledge. We further identify and resolve the incompatibility between standard PEFT attention-level adapter injection and the Gemma-4 Per-Layer Embedding architecture that tends to cause gradient isolation. Trained on the IWSLT 2026 Track 1 data covering Hausa, Igbo, and Yoruba, the final system achieves a best proxy teacher-forced SacreBLEU of 91.29 on the validation set at Phase 3, with Phase 1 speech encoder validation loss converging to 0.651.

Speech recognition is known to be a specialized application of speech processing. Automatic speech recognition (ASR) systems are designed to perform the speech-to-text task. Although ASR systems have been the subject of extensive research, they still encounter certain challenges when speech variations arise. The speaker’s age, gender, vulnerability, and other factors are the main causes of the variations in speech. In this work, we propose a fine-tuned speech recognition model for recognising the spoken words of vulnerable individuals in Tamil. This research utilizes a dataset sourced from the LT-EDI@EACL2024 shared task. We trained and tested pre-trained ASR models, including XLS-R and Whisper. The findings highlight that the fine-tuned Whisper ASR model surpasses the XLSR, achieving a word error rate (WER) of 24.452, signifying its superior performance in recognizing speech from diverse individuals.

Accented speech classification plays a vital role in the advancement of high-quality automatic speech recognition (ASR) technology. For certain applications, like multi-accented speech classification, it is not always viable to obtain data with accent variation, especially for resource-poor languages. This is one of the major reasons that contributes to the underperformance of the speech classification systems. Therefore, in order to handle speech variability in Indian language speaker accents, we propose a few-shot learning paradigm in this study. It learns generic feature embeddings using an encoder from a pre-trained whisper model and a classification head for classification. The model is refined using LLM’s fine-tuning techniques, such as LoRA and QLoRA, for the six Indian English accents in the Indic Accent Dataset. The experimental findings show that the accuracy of the model is greatly increased by the few-shot learning paradigm’s effectiveness combined with LLM’s fine-tuning techniques. In optimal settings, the model’s accuracy can reach 94% when the trainable parameters are set to 5%.