Deconstructing a Commercial Spatial Audio Effect

Back when I was utilizing a different operating system, I relied on a specific piece of software that processed audio in real-time to enhance the listening experience. The effect was subtle yet impactful: it made the sound feel lighter and shifted the perceived source further away from the listener, while still preserving all the essential sonic details.

Because I wanted a similarly intuitive tool for Linux, I decided to build my own.

The Goal: Create a processed version of audio that feels less "blocked" and reduces ear fatigue during long listening sessions by increasing the perceived distance of the sound.

---

🛠️ The Initial Approach

Since I was developing for Linux, I chose PipeWire as my foundational stack. Being new to the world of audio DSP, I used an LLM as a conceptual guide while experimenting with psychoacoustic methods and standard filters.

During this phase, I discovered HRTFs (Head-Related Transfer Functions), which model how sound interacts with a listener's physical geometry to reach the ears from a specific coordinate in space.

Database Comparison

I tested two primary datasets to achieve this effect:

As I pushed further, the LLM began suggesting highly technical concepts:

• Perceptual envelopment
• Late-field decorrelation
• Spectral diffusion
• Diffuse spatial energy

I tried to implement these suggestions blindly. Unfortunately, while I could hear that my results were "off," these abstract terms didn't provide a concrete path toward improvement. The project eventually stalled.

---

🔍 Shifting Strategy: Black-Box Analysis

I realized that instead of guessing the internal algorithms trying to deduce the exact parameters, I should simply analyze the behavior of the filter itself.

This was a challenge because I didn't own the software; I only had access to it via a friend's laptop. This meant I couldn't iterate in real-time. I began studying Signalsmith and various audio engineering blogs to develop the intuition needed to measure the effect accurately.

The Workflow

---

📐 The Assembly of Tests

The core challenge was designing a single test file that could capture all the necessary characteristics of the filter without requiring multiple recording sessions.

1. Impulse Response (IR) Measurement

By passing a single nonzero sample (an impulse) through the filter, any subsequent audio in the recording is a direct result of the filter's processing. This reveals:

• The shape and extent of the decay.
• Associated phase delays.

If we define the response to an impulse as $h$, then the response to an input $x$ at position $k$ is simply $h$ shifted to $k$ and scaled by $x[k]$. The total output is the convolution:

$y[n]=\sum_k x[k]h[n-k]=(x*h)[n]$

2. Verifying LTI Properties

I needed to confirm if the filter was a Linear Time-Invariant (LTI) system. I tested for:

• Homogeneity: $F(ax)=aF(x)$ (Louder inputs produce proportionally louder outputs).
• Additivity: $F(x_1+x_2)=F(x_1)+F(x_2)$ (Combined inputs equal the sum of individual responses).
• Time-Invariance: The response remains the same regardless of when the input is played.

For a stereo system, the relationship is represented by the matrix:

$\begin{bmatrix} y_L\\ y_R \end{bmatrix} {}={} \begin{bmatrix} h_{LL} & h_{LR}\\ h_{RL} & h_{RR} \end{bmatrix} * \begin{bmatrix} x_L\\ x_R \end{bmatrix}$

To test this, I created a file with impulses at varying amplitudes, played them in the left channel, the right channel, both simultaneously, and with the right channel phase-inverted.

3. Sine Sweeps vs. Impulses

While impulses are useful, I noticed that professional literature favors sine sweeps. While an impulse contains all frequencies, it has very little energy per frequency, making it susceptible to noise. A sweep increases the Signal-to-Noise Ratio (SNR).

The transfer function $H(f)$ is calculated as: $Y(f)=H(f)X(f), \qquad H(f)=\frac{Y(f)}{X(f)}$

To prevent division by zero (which would amplify recording noise), I used a regularization term $\lambda$:

H(f) \approx \frac{Y(f)X^*(f)}{|X(f)|^2 + \lambda}

Pros and Cons of Sweeps:

• Pro: Better noise rejection and phase retention.
• Con: Longer duration, which might trigger adaptive behavior in the commercial program.

Finally, I added steady tones and amplitude-modulated tones (quiet $\rightarrow$ loud $\rightarrow$ quiet) to analyze gain changes and the system's temporal response.