Amrutha Prasad from the Department of Computer Graphics and Multimedia will defend her dissertation in July

We invite you to the public defense of the dissertation by Amrutha Prasad from the Department of Computer Graphics and Multimedia, FIT, BUT, which will take place on Thursday, July 9, 2026, at 10:00 a.m. in conference room L220. The thesis advisor for the dissertation titled “Parameter-efficient models for domain-specific speech and language recognition” is Assoc. Prof. Ing. Petr Motlíček, Ph.D.

Amrutha Prasad collaborates with Assoc. Prof. Motlíček at the IDIAP research institute in Martigny, Switzerland, which is an internationally renowned center focused on artificial intelligence, machine learning, and speech processing. She is a member of the faculty research group Speech@FIT. In her research, she specializes in automatic speech recognition, language and speaker identification, and the use of speech technologies in air traffic control communication. Her research findings have been published at leading academic conferences, such as Interspeech, IEEE ICASSP, IEEE ASRU, Odyssey: The Speaker and Language Recognition Workshop, and others. She has also participated in international research activities related to the European ATCO2 project, focused on air traffic communication, and was a member of the broader Swiss national team participating in the NIST SRE24 speaker recognition evaluation.

Amruth’s dissertation addresses the question of how to reduce the size of large-scale speech recognition models without significantly compromising their accuracy. Current models, such as XLS-R and Whisper, are very effective at transcribing speech to text and identifying the language spoken in a recording. However, they contain a huge number of parameters and therefore require significant memory capacity, high computational power, and consume a lot of energy. This poses a particular problem in applications where the system must operate quickly and reliably, especially on less powerful hardware or with limited energy resources. This dissertation focuses primarily on communication in the field of air traffic control (ATC). This is an exceptionally challenging environment: Recordings are often affected by background noise, speakers have different accents, the quality of the radio signal fluctuates, and communication is concise and highly technical. At the same time, air traffic control is a safety-critical field in which errors can lead to serious consequences.

The dissertation focuses on two main areas: ASR—automatic speech recognition, i.e., the conversion of spoken language to text; and LID—language identification, which determines the language spoken in a recording. The author’s first research approach involves reducing the final components of the model, which, for example, determine which language was detected. A standard time-delayed neural network, referred to as TDNN, is replaced by a variant called TDNN-F, which uses parameters more efficiently. This model achieves similar capabilities in distinguishing between languages while requiring approximately 30 to 50% fewer parameters. The second approach utilizes a method known as normalization flows. This method allows for more flexible modeling of complex speech data distributions than the traditional PLDA method, which is based on the relatively simplifying assumption that the data approximately follow a Gaussian distribution. On two of the three evaluated datasets, this approach achieved higher accuracy than the PLDA method. The third and most significant part of the dissertation deals with direct reduction of the core of the XLS-R and Whisper models. The weight matrices in the forward layers of these models are replaced by a low-rank approximation created using singular value decomposition (SVD). Only the first r principal components are retained. Simply put, the model attempts to preserve the most important information and eliminate less essential parts of the computation. This method is also combined with the LoRA technique, which allows the model to be adapted during training using a significantly smaller number of parameters. The advantage of the proposed solution is that the model operates more efficiently not only during training but also during actual use, i.e., during inference.

What are the main results of the dissertation? For the XLS-R model, the number of parameters was successfully reduced by approximately 23–50%, with only a minimal decrease in accuracy. For the Whisper model, reducing the number of parameters by 50% did not cause a deterioration in performance during automatic speech recognition on a dataset from the field of air traffic control. However, according to the results, a certain decline in performance was observed for other datasets and tasks. When the number of parameters was reduced by 28%, the error rate actually decreased by 1.8%, meaning that transcription accuracy improved slightly. One possible explanation is that compression removed some redundant parameters and acted as a form of regularization—that is, a way to prevent the model from overfitting to the training data. The main conclusion of the study is that large speech processing models are often unnecessarily large. Their size and operational costs can be significantly reduced without substantially limiting their performance. This can facilitate their deployment, for example, directly on local or edge devices, in systems with limited memory and computing power, in real-time applications, and also in safety-critical fields such as aviation, as Amrutha Prasad states in her dissertation.

We wish Amrutha a successful defense and the best of luck in her future career!