Termeh Taheri

Website: https://termehtaheri.github.io/

Research keywords: Audio & Speech Processing, Explainable AI (XAI), Signal-Aware Representation Learning, Multimodal Machine Learning

Bio:

I have a background in computer engineering, where I initially focused on medical image analysis for breast cancer detection and computational neuroscience, analyzing EEG brain signals for imagined speech decoding to assist non-verbal individuals. Following my undergraduate studies, I spent two years working as a software engineer.

In 2024, I was awarded a Chevening Scholarship to pursue my master’s in Artificial Intelligence in London. It was during this time that I developed a strong interest in audio signal processing, conducting research on symbolic audio reasoning within large language models. After graduating, I spent a year working as a machine learning engineer across startups, building practical AI systems. I have now moved to Edinburgh to begin my PhD, where I aim to build on this foundation by developing transparent, signal-aware audio understanding and generation systems.

PhD research:

My research focuses on designing structured representations for explainable and signal-aware audio and speech understanding and generation.

Rather than relying solely on opaque dense embeddings, my work captures acoustically meaningful features, such as prosody, timing, and harmonic structure, and integrates them with large language models to enable transparent reasoning.

Explanatory outputs will be investigated in both textual and non-textual modalities, including annotated spectrograms and auralisations that highlight the signal cues underlying model decisions. On the generation side, the project will develop controlled audio generation methods that provide explicit accounts of acoustic choices such as word pronunciation and prosodic patterns. A particular focus is placed on detecting and explaining audio manipulation and deepfakes. Evaluation addresses both core technical performance and the intelligibility and trustworthiness of explanations to real-world, non-technical audiences.

Supervisors: Simon King, Catherine Lai