Velvet Feedback Matrix in FDNs

• VFM is a specialized feedback matrix in FDNs that uses velvet-noise-like sparse FIR filters to disperse impulses and achieve dense, lossless echo responses.
• Its design employs cascading stages with normalized Hadamard matrices and delayed vectors to maintain paraunitary properties and computational efficiency.
• The VFM enables rapid, decorrelated reverberation in artificial reverb systems, offering high echo density with significantly reduced processing load.

The Velvet Feedback Matrix (VFM) is a specialized class of filter feedback matrices within feedback delay networks (FDNs) designed to significantly increase short-term echo density while maintaining computational efficiency and losslessness. In contrast to the standard scalar feedback matrices of conventional FDNs, the VFM leverages paraunitary FIR filtering with "velvet-noise"-like characteristics, spreading each echo path impulse over a short burst and thereby emulating the temporal diffusion produced by non-specular acoustic scattering. This construction enables the generation of dense, decorrelated reverberation responses at a fraction of the cost associated with traditional high-dimensional FDNs (Schlecht et al., 2019).

1. Theoretical Foundations and Definitions

FDNs are recursive filter structures featuring $M$ delay lines of integer lengths $D = [d_1, \dots, d_M]$ coupled by a feedback matrix $F$. The VFM replaces the traditional scalar, constant, unitary $M \times M$ feedback matrix with a paraunitary filter matrix $A(z)$, each entry $a_{ij}(z)$ being a sparse FIR filter. The resulting transfer function is

$H(z) = c^\top [D(z^{-1}) - A(z)]^{-1} b + d$

where $D(z^{-1}) = \text{diag}(z^{-d_1}, \dots, z^{-d_M})$, $b$, $c$ are input/output vectors, and $d$ is a direct path. The VFM is characterized by $A(z)$ with entries composed of "velvet noise" sequences—FIR sequences with randomly located $\pm 1$ pulses of controlled sparsity.

The paraunitary (lossless) property is crucial: $A(1/z^*)^H\, A(z) = I_M,$ which ensures energy preservation and thus system stability.

2. Mathematical Construction of the Velvet Feedback Matrix

The construction of the VFM proceeds by cascading $K$ stages, each formed by a fast orthogonal transform and a sparse delay matrix:

• Begin with a normalized Hadamard matrix $H$ of size $M \times M$ ($HH^\top=I$, entries $\pm M^{-1/2}$).
• Choose a velvet density $\rho$ (expected pulses per sample per filter), typically $\rho = 1/30$.
• For each stage $k=1,\dots,K$, select a unique delay vector $\Delta_k \in \mathbb{N}^M$, with delays distributed over $[0, L_{k-1}/\rho]$ (where $L_{k-1} = \lceil M^{k-1} / \rho \rceil$).
• Set filter order $L_k = \lceil M^k/\rho \rceil$.
• Iteratively define:

$$A_0(z) = H,\quad A_k(z) = H \cdot \text{diag}(z^{-\Delta_k}) \cdot A_{k-1}(z)$$

• After $K$ stages, $A(z) = A_K(z)$ has FIR order $L_K$ and each entry forms a sparse, velvet-like noise sequence of $\sim M^K$ pulses spread within $L_K$ samples.

The final VFM is thus: $$A(z) = \prod_{k=1}^K [ H\, \text{diag}(z^{-\Delta_k}) ] \cdot H$$ This design ensures that each input impulse is temporally diffused into a mini-impulse train, greatly enhancing short-term echo density (Schlecht et al., 2019).

3. Echo Density and Physical Interpretation

In a conventional FDN, each echo path of length $\ell$ manifests as a solitary sample at the time corresponding to the sum of the delays along the path. The VFM mechanism replaces this point-wise response with a temporally spread burst of up to $L_K$ samples. After $\ell$ traversals through the VFM, the impulse response has been convolved $\ell$ times by the velvet filter, yielding a temporal spread of approximately $\ell \cdot L_K$ samples.

Physically, this emulates the effect of multiple incoherent scatterings, analogous to sound undergoing a series of reflections on rough surfaces with each reflection inducing temporal and amplitude dispersion. The result is drastically increased short-term echo density, achievable with fewer delay lines (lower $M$) or lower computational burden compared to scalar-matrix FDNs.

4. Implementation and Computational Considerations

Efficient implementation of the VFM exploits the fast Hadamard transform (WHT) and circular buffer operations. The main loop consists of:

• Input injection.
• For $k=1\dots K$: delay updates via $M$ circular-buffer accesses; Hadamard matrix multiplication per stage.
• Final Hadamard stage.
• Output computation.

Per time step, each stage involves $M$ buffer reads/writes and $O(M \log M)$ arithmetic operations. For $K$ stages, the total per-sample complexity is $O(K(M \log M + 2M))$. Alternatively, block-based fast convolution (FFT) is available for larger $L_K$.

Parameter guidelines:

• $M$: Higher $M$ increases modal density; VFM allows small $M$ (as low as 4) with maintained echo richness.
• $K$: Higher $K$ (generally 1–3) increases pulse count and filter order, with denser echo distributions.
• $\rho$: Controls sparsity/density; typically $1/30$–$1/10$.
• $\Delta_k$: Chosen for column-distinctness; e.g., $$\Delta_1 = [0,\, \lfloor L_0/\rho \rfloor,\, 2\lfloor L_0/\rho \rfloor,\, \ldots]$$.

Lossy FDNs can be realized by inserting a diagonal gain matrix $G(z)$ post-delay/gain to achieve the desired reverberation time, compensating for the group delay imparted by $A(z)$.

5. Empirical Performance and Comparative Assessment

The VFM demonstrates substantial improvements in both mixing time and computational efficiency:

• Achieves the target echo density $\sim$10x faster than a $4 \times 4$ scalar-matrix FDN and with $\sim$3x lower cost than a $16 \times 16$ scalar-matrix FDN.
• For a $4 \times 4$ VFM versus $4 \times 4$ and $16 \times 16$ standard FDNs, the VFM achieves high echo density within 40–80 ms compared to 200–300 ms for the scalar $4 \times 4$, with far fewer arithmetic operations than the scalar $16 \times 16$ (Schlecht et al., 2019).
• Modal decay distribution: raising $K$ increases the spread of per-mode RTs, but for $K \leq 2$ the maximum error remains $<$8%, below the just-noticeable difference for most listeners.

A summary table of computational costs across designs is presented in Table I of the source, demonstrating the operation count advantages of the VFM structure.

6. Stability, Robustness, and Design Trade-offs

Stability of the VFM is ensured by the paraunitary construction of $A(z)$, such that all system poles remain on the unit circle. Memory requirements scale with $M$ and $K$ due to ring buffers for each channel-stage combination. Group delay introduced by $A(z)$ must be compensated within the attenuation design to control reverberation time accuracy.

The trade-off between the parameters $K$ and $\rho$ directly impacts computational load, echo density, and RT variability. Higher values yield denser, more diffuse reverberation at increased processing costs and larger reverberation time jitter. Residual RT variation is generally small ($<$8% for practical settings), and can be tuned to remain below perceptual thresholds.

7. Practical Applications and Limitations

The primary application domain for the VFM is artificial reverberation and decorrelation systems, particularly where low-latency, computationally efficient, yet perceptually dense reverberation is required. A practical advantage is the ability to achieve high echo and modal densities with small matrices ($M$), bypassing the memory and processing overhead of larger scalar FDNs.

A plausible implication is that the VFM’s ability to mimic acoustic scattering with minimal computation enables its use in embedded audio devices, interactive environments, and high-fidelity rendering on resource-constrained platforms.

Physical interpretation aligns with cumulative temporal scattering; thus, VFM-based FDNs provide a valuable modeling tool for simulating rough-room acoustic phenomena in a real-time and lossless manner.

References:

• "Scattering in Feedback Delay Networks" (Schlecht et al., 2019).