Large pretrained acoustic models have become foundational to modern speech technologies, achieving state-of-the-art performance across a variety of tasks. However, their massive parameter counts pose significant challenges for deployment in resource-constrained and safety-critical environments. This doctoral thesis addresses these limitations by proposing and evaluating model compression and parameter-efficient methodologies for Automatic Speech Recognition (ASR) and Language Identification (LID). The research primarily focuses on the Air Traffic Control (ATC) domain—a highly demanding scenario characterized by reduced linguistic context, elevated background noise, and high speaker diversity, making it an ideal testbed for evaluating robust speech processing under real-world constraints.

To mitigate computational costs without sacrificing performance, the thesis first targets task-specific back-end discriminators built on top of the XLS-R model. For the LID task, the traditional Time-Delay Neural Network (TDNN) backend discriminator is replaced with Factorized Time-Delay Neural Networks (TDNN-F), achieving a 30–50% parameter reduction while maintaining full discriminative capacity. Furthermore, the thesis explores the use of Normalizing Flows to model complex embedding distributions. This effectively circumvents the traditional Gaussian assumptions of Probabilistic Linear Discriminant Analysis (PLDA) scoring and removes the need for length normalization.  For two of the three datasets, the Flow model consistently yielded higher accuracy than the PLDA scoring method.

The thesis then investigates structural compression within the foundational models themselves. Low-rank matrix factorization via Singular Value Decomposition (SVD) is applied to the feed-forward layers of transformer architectures (XLS-R and Whisper), both independently and in conjunction with Low-Rank Adaptation (LoRA). Unlike standard parameter-efficient methods that constrain the parameter budget only during fine-tuning, this factorization reduces model size during both training and inference.  The effectiveness of the proposed techniques is tested on three datasets for LID and on two datasets for ASR. For the XLS-R model, parameters are reduced by 23–50% with minimal to no performance degradation. For Whisper, a 28% parameter reduction yields a 1.8% absolute improvement in Word Error Rate (WER), while a 50% reduction introduces no degradation.