Stavros Papadopoulos | Searcharxiv

arXiv2025

Retrieval Augmented Generation based context discovery for ASR

This work investigates retrieval augmented generation as an efficient strategy for automatic context discovery in context-aware Automatic Speech Recognition (ASR) system, in order to improve transcription accuracy in the presence of rare or out-of-vocabulary terms. However, identifying the right context automatically remains an open challenge. This work proposes an efficient embedding-based retrieval approach for automatic context discovery in ASR. To contextualize its effectiveness, two alternatives based on large language models (LLMs) are also evaluated: (1) large language model (LLM)-based context generation via prompting, and (2) post-recognition transcript correction using LLMs. Experiments on the TED-LIUMv3, Earnings21 and SPGISpeech demonstrate that the proposed approach reduces WER by up to 17% (percentage difference) relative to using no-context, while the oracle context results in a reduction of up to 24.1%.

cs.CL↗

arXiv2017

Engineering Methods for Differentially Private Histograms: Efficiency Beyond Utility

Publishing histograms with $ε$-differential privacy has been studied extensively in the literature. Existing schemes aim at maximizing the utility of the published data, while previous experimental evaluations analyze the privacy/utility trade-off. In this paper we provide the first experimental evaluation of differentially private methods that goes beyond utility, emphasizing also on another important aspect, namely efficiency. Towards this end, we first observe that all existing schemes are comprised of a small set of common blocks. We then optimize and choose the best implementation for each block, determine the combinations of blocks that capture the entire literature, and propose novel block combinations. We qualitatively assess the quality of the schemes based on the skyline of efficiency and utility, i.e., based on whether a method is dominated on both aspects or not. Using exhaustive experiments on four real datasets with different characteristics, we conclude that there are always trade-offs in terms of utility and efficiency. We demonstrate that the schemes derived from our novel block combinations provide the best trade-offs for time critical applications. Our work can serve as a guide to help practitioners engineer a differentially private histogram scheme depending on their application requirements.

cs.DB↗