<p>State-of-the-art automatic speech recognition (ASR) systems often struggle to maintain reliability and security in noisy or unpredictable environments, especially when trained on limited data. Bayesian approaches introduce uncertainty modeling to address this challenge. However, existing methods often apply a uniform uncertainty modeling strategy across all noise conditions, which may limit performance. In this work, we propose a secure end-to-end reparameterized Bayesian long short-term memory (LSTM)-ASR framework that captures epistemic (model) and aleatoric (data) uncertainties by modeling network weights and Mel-frequency cepstral coefficients (MFCCs) as Gaussian distributions, respectively. We evaluate the proposed framework on the AudioMNIST and Google’s Speech Commands datasets in the presence of real-world noises collected from the Microsoft Scalable Noisy Speech Dataset (MS-SNSD). Our findings indicate that epistemic uncertainty is more effective for stationary background noises, such as air conditioning and traffic, while aleatoric uncertainty is superior for transient and unpredictable noises, including keyboard typing and neighboring speech. These results show how important it is to choose the right uncertainty modeling method based on the type of noise to make ASR systems more secure and reliable in noisy settings.</p>