Abstract
Artificial reverberation (AR) models are foundational to modern audio engineering, yet estimating their parameters (ARPs) to match specific acoustic environments remains a complex task. This paper introduces Differentiable Artificial Reverberation (DAR) models, which bridge the gap between classical digital signal processing (DSP) and deep learning. By enabling end-to-end training through differentiable reverb layers, the proposed method allows neural networks to optimize parameters based on the actual sound produced rather than just numerical targets. To solve the computational bottleneck of recursive filters on GPUs, the authors employ a frequency-sampling method to create efficient FIR approximations. Results demonstrate that DAR models significantly outperform traditional non-differentiable baselines in both objective accuracy and subjective listening quality.
The Challenge: Non-Differentiable DSP
Traditionally, tuning an AR model to match a reference room impulse response (RIR) or reverberant speech was treated as a "parameter matching" problem. A neural network would attempt to predict the exact numerical settings of a reverb plugin. However, this approach has two critical limitations:
- Gradient Disconnect: Standard reverb plugins are non-differentiable. During training, the network cannot "hear" the result of its predictions; it only knows if its predicted numbers match the ground truth, not if the resulting audio sounds correct.
- GPU Inefficiency: AR models frequently rely on Infinite Impulse Response (IIR) filters. These are recursive, meaning each output sample depends on the previous one, making them difficult to parallelize on GPUs.
Methodology: Differentiable Reverb Layers
The authors propose a framework where the reverb model itself is a differentiable layer within the neural network. The architecture consists of an encoder that processes reference audio into a latent representation, which is then projected into the specific parameters required by the AR model.
Frequency-Sampling for GPU Acceleration
To overcome the IIR bottleneck, the authors introduce the Frequency-Sampling Method. Instead of implementing recursive difference equations, they sample the filter's frequency response and convert it into a high-order Finite Impulse Response (FIR) filter. This allows the reverberation to be generated using Fast Fourier Transforms (FFTs), which are highly optimized for parallel processing on GPUs.
As shown in the figure above, as the number of frequency samples ($N$) increases, the FIR approximation (red) converges to the original filter's response (blue), providing a high-fidelity differentiable proxy for the original DSP component.
Key Insights and Results
- The "DSP Prior" Advantage: Unlike "black-box" neural decoders that can produce audible artifacts, DAR models act as a structural prior. They constrain the network's output to sound like a real room, leading to much higher subjective quality scores.
- Objective Performance: The DAR approach achieved dramatic improvements in accuracy. For instance, the error in reverberation time ($\\Delta T_{30}$) was reduced from over 20% in traditional baselines to under 5%.
- Subjective Quality: In MUSHRA listening tests, the end-to-end DAR method scored between 70-80 on a 100-point scale, indicating that the difference from the reference was "noticeable but not much," whereas non-differentiable baselines scored below 60.
- Efficiency: Despite the complexity of the training process, the resulting models are efficient enough for real-time use on a CPU, with the Advanced Filtered Velvet Noise (AFVN) model requiring only ~1511 FLOPs per sample.
Conclusion
Differentiable Artificial Reverberation represents a significant step forward in integrating classical DSP with modern machine learning. By making the reverb generation process part of the computational graph, researchers can leverage the strengths of both worlds: the interpretability and efficiency of DSP models and the powerful optimization capabilities of deep neural networks.
Summary updated