Deep Learning for Head Orientation Estimation Using Phase Spectrograms from a Single Microphone Array
Year: 2025
Type: dissertation
Abstract: Estimating a speaker's head orientation from audio provides valuable contextual information for applications such as smart environments, meeting analysis, and driver monitoring.This thesis presents a data-driven method that estimates head orientation from the audio signals captured by a single compact microphone array, using only phase information extracted from the shorttime Fourier transform.The proposed deep neural network combines convolutional, recurrent, and self-attention layers to learn spatial, temporal, and contextual patterns directly from phase spectrograms.In contrast to prior work relying on handcrafted, physics-based features or raw waveform inputs, the presented approach enables robust learning from both simulated and real data.A large-scale simulated dataset was generated using measured voice directivity patterns and room acoustic simulations, allowing the model to learn orientation-dependent cues under varied acoustic conditions.The network is then fine-tuned on real recordings, resulting in improved adaptation to real-world variability.The proposed system achieves state-of-the-art accuracy, outperforming baseline models under both clean and noisy conditions.In reverberant environments, it attains a mean angular error of 26.0°, and further personalization to individual speakers and environments reduces this to 13.9°.Evaluation on a real dataset confirms the method's generalization capability, reaching 71.1% classification accuracy after fine-tuning.These results demonstrate that accurate head orientation estimation can be achieved with a single microphone array, without the need for cameras or dense array setups.
Author: Turi, Balint
Institution: Tampere University
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7133029270