Developing the Crosstalk Filter

The first experiments

The project did not begin with a custom algorithm. The first tests used filter sets published by Ambiophonics. They revealed spatial qualities worth pursuing, but they also introduced artifacts that were too obvious for normal listening. Changing the available attenuation setting altered their severity without solving the underlying mismatch.

That failure was useful. It showed that the basic idea had potential, but a generic filter could not simply be added as a stereo effect. The filter had to be designed around the real acoustic paths between the two loudspeakers and the listener. That is why a binaural measurement head was built — to capture those paths at the ears. It is a tool, not a product.

The question behind the design

The next step was to study the Princeton 3D3A work and the closely spaced “stereo dipole” arrangement, where the loudspeakers may span only about 10°. This geometry has a genuine advantage: a small lateral head movement changes the two path lengths relatively little. However, it also asks the DSP to create separation between two sources that begin in almost the same direction.

Conventional stereo is not a binaural signal. It is a constructed spatial representation created by microphone technique, panning and mixing, and it does not contain one unique pair of listener-specific ear signals. There is therefore no single theoretically correct ear-pressure target for reproducing it. A true binaural recording does provide such a target, and a closely spaced stereo dipole can be well suited to delivering it through loudspeakers. For ordinary stereo, wider-spaced loudspeakers with a crosstalk-cancellation filter optimized for the actual HRTFs and playback geometry are a more appropriate starting point: the system works with the recording’s panning and phantom-image cues instead of treating its two channels as listener-specific ear signals.

The practical question became: why ask the filter to cancel energy that the listener’s head can attenuate naturally? The opposite-ear, or contralateral, path is shaped not only by distance and delay but also by diffraction around the head and filtering by the pinnae. At wider speaker angles this natural separation becomes significant at high frequencies. At low frequencies the wider geometry also improves the conditioning of the two acoustic paths, so less corrective gain is required.

Let geometry shorten the filter

A cancellation signal does not travel to only one ear. Part of it crosses to the other ear and creates a smaller residual that needs another correction. That correction produces another residual, and so on. This creates a sequence of progressively weaker cancellation cycles — the recursive tail of the filter.

Baffless uses a speaker span of about 60° and up to 90° at the listening position. With the loudspeaker geometry and the listener’s head-related transfer function (HRTF) doing part of the separation naturally, each successive cancellation cycle is weaker. The useful filter tail can therefore be shorter, and the low-frequency correction requires less boost.

The custom filter was developed around that geometry instead of a single adjustable attenuation value. In the original development notes, this was the turning point: the spatial effect remained, while the artifacts that made the first filters unusable did not.

What later research clarified

After the custom filter was working, a 2006 study by Mingsian R. Bai and Chih-Chung Lee provided a useful framework for understanding the result. It compared point-source and HRTF models and tested loudspeaker spans of 10°, 60° and 120°. In the HRTF model, natural head shadowing reduced high-frequency ringing and improved the useful separation of the wider arrangements.

There is an important nuance in the paper. Its simulation of an absolute performance region suggested spans around 120°–150°, but the listening experiments compared 10°, 60° and 120°. The 120° arrangement performed comparably to 60° and better than 10°, while the authors still described 60° as a practical compromise and found 120° less effective for frontal images. The paper therefore does not establish one universal best angle. It supports the broader design decision to move away from an extreme 10° span and to consider HRTF, panning, performance and head movement together.

Sources behind the development

• Glasgal, “Ambiophonics, 2nd Edition” — a local copy of the book originally published by the Ambiophonics Institute and the starting point for the first filter experiments.
• Choueiri, “Optimal Crosstalk Cancellation for Binaural Audio with Two Loudspeakers” — a local copy of the Princeton 3D3A paper reviewed during development.
• Bai and Lee, “Objective and subjective analysis of effects of listening angle on crosstalk cancellation” — a local copy of the 2006 paper comparing 10°, 60° and 120° arrangements.