This thesis focuses on sound source tracking, which aims to extract the spatial positions of sound sources from multichannel microphone recordings. Emphasis is placed on meeting-like scenarios involving speech sources in indoor environments, with a specific focus on the overlooked problem of tracking intermittent and moving speakers, who can alternate between speech and silence periods and may move unpredictably while silent.

In a first contribution, an evaluation pipeline centered on the problem of intermittent and moving speakers is designed. This includes LibriJump, a synthetic evaluation dataset featuring intermittent speakers that move unpredictably while silent, as well as complementary tracking metrics adapted from multi-object tracking to sound source tracking. Experiments using this novel evaluation pipeline confirm that unpredictable movement during silence is a limitation of current tracking approaches.

The remaining contributions investigate the integration of deep speaker embeddings as identity-related features for tracking intermittent and moving speakers. A first solution that leverages a pretrained general-purpose speaker-embedding extractor is proposed, improving performance on the LibriJump dataset and demonstrating the usefulness of speaker embeddings. However, this first solution is sensitive to factors that degrade speaker embedding quality, motivating the need for a more tailored embedding extractor in a second contribution through the design of an appropriate training strategy. Results demonstrate improved short-context embedding extraction and increased robustness to overlap, although further improvements in speech-overlap robustness remain possible.

Keywords: Sound source tracking, speaker recognition, deep learning, speaker tracking, intermittent and moving speakers, deep speaker embeddings