Introduction
In the development of "smart" rooms, determining a person's location is only the first step toward understanding human interaction. To discern who is speaking to whom or if a speaker is addressing an audience, we must estimate their head orientation. This paper, originally presented at INTERSPEECH 2012, introduces a robust acoustic method that rivals video-based tracking by analyzing how the human head filters sound.
The Challenge of Acoustic Orientation
Estimating head orientation using only audio is inherently difficult. Unlike light, sound waves—particularly at low frequencies—diffract around the head, creating a complex radiation pattern. Previous acoustic methods often relied on energy levels, assuming the loudest microphone indicates the speaker's direction. However, this approach is sensitive to microphone calibration, room acoustics, and the varying distances between the speaker and the sensors.
Methodology: Coherence over Energy
The authors propose a two-step algorithm to overcome these limitations. First, the system localizes the speaker's 3D position using the Steered Response Power with Phase Transform (SRP-PHAT) algorithm. Second, it estimates orientation by examining the magnitude of the Generalized Cross-Correlation with Phase Transform (GCC-PHAT) peak for every microphone pair.
The core innovation is the use of the GCC-PHAT peak as a measure of signal coherence. When a speaker faces a microphone, the direct sound path is unobstructed, resulting in a sharp, high-magnitude peak. Conversely, when facing away, the head acts as a low-pass filter, and the microphone receives a higher proportion of reflected (reverberant) sound, causing the peak to drop.
As shown in the figure above, the orientation cue is most significant in the 2 kHz to 12 kHz range. High-frequency sounds have shorter wavelengths and are more easily shadowed by the head, creating a more directional pattern that the GCC-PHAT peak can effectively capture.
Experimental Results
The proposed method, GCC-PHAT-P, was evaluated using the UPC Smart-Room Database and the CLEAR Head Pose Database. In controlled environments, the system achieved a Mean Average Error (PMAE) of 11.80°, significantly outperforming traditional energy-based methods.
In real-world seminar settings, the audio-only method was compared directly against established video-based tracking results. The GCC-PHAT-P method achieved a PMAE of 32.52°, which actually slightly outperformed the best video-based algorithm of that era (33.56°).
Conclusion and Future Work
The use of GCC-PHAT peak values provides a robust and calibration-independent cue for head orientation. While the current study focused on single male speakers, it demonstrates that acoustic coherence can match or exceed the performance of visual systems in complex environments. Future research will focus on expanding the system to handle speaker diversity and multi-talker scenarios.
Summary updated