A proposed array-agnostic Ambisonics encoder called ADEPS led the reported comparisons on SISDR, a measure of signal distortion, across the microphone layouts tested. In the ideal setup, it also had the lowest interaural-coherence (IC) error in every listed array column, although rankings for spectral error, coherence and interaural level-difference (ILD) error varied. Under order mismatch, ADEPS again recorded the highest SISDR in each listed column, but it was not best on every spectral-error comparison.
The study frames ADEPS as a response to array-dependent Ambisonics artifacts and the limits of systems built for fixed microphone arrangements. Its proposal is zero-shot encoding across arbitrary microphone-array topologies, meaning different physical layouts.
How the test was built
The reported dataset used VCTK speech for training and WSJ0 speech for evaluation, with 20,000 training scenes and 1,000 evaluation scenes. Evaluation used 13 test arrays: Project Aria plus four random irregular realizations for each microphone count of four, five and six. Signals were corrupted at 50 dB signal-to-noise ratio, producing 13,000 test signals.
At the design level, the numerical study set up an ideal case with Neff = 5 and an order-mismatch case with Neff = 15. ADEPS used an Ambisonics prior of order Np = 5, while the evaluation covered coefficients up to Nenc = 1. The main arbitrary-array comparators were linear encoding and parametric synthesis.
The generative prior in ADEPS was trained in an unsupervised manner solely on target Ambisonic representations. The scorecard included SISDR, mean log-magnitude spectrum error and magnitude-square coherence. It also measured binaural interaural level-difference (ILD) and interaural-coherence (IC) cue errors.
Where the signal score pulled ahead
In the ideal comparison, ADEPS's SISDR values were 11.66 dB with four microphones, 14.18 dB with five, 16.20 dB with six and 10.00 dB for Project Aria. Those were the highest SISDR values in every listed array column. Its IC errors were also the lowest in every column. The broader scorecard was less one-sided: spectral error, coherence and ILD-error rankings varied.
The mismatch case lowered ADEPS's reported SISDR values to 9.85, 10.57, 12.67 and 4.94 dB across the four-mic, five-mic, six-mic and Aria columns. Even there, ADEPS had the highest SISDR in every listed column, and the supplied analysis reports a coherence drop relative to linear encoding. The method was not best on every spectral-error comparison.
The authors also report reduced low-frequency artifacts and suppressed high-frequency spatial aliasing across the test signals.
A narrower neural comparison
The comparison with U-Net used one 4-mic ideal-array configuration. ADEPS reported 11.67 dB SISDR versus 8.03 dB for U-Net; mean spectral error was 5.31 versus 5.79 dB, ILD error was 0.92 versus 2.10 dB, and IC error was 0.06 versus 0.11. U-Net, however, had higher coherence: 0.85 versus ADEPS's 0.78. On the reported metrics, ADEPS was better on four measures, while U-Net led on coherence.
The boundary of the result
The paper's clearest technical boundary is order. The framework is restricted to spatially resolvable orders, and underdetermined higher-order coefficients are identified as future work. The numerical setup used Np = 5 and evaluated up to Nenc = 1, with the mismatch scenario at Neff = 15; performance under other mismatch patterns is not established.
This is a numerical benchmark built from the reported speech scenes and array configurations. It covers 13 test arrays and 13,000 test signals under the reported 50 dB signal-to-noise condition. The supplied analysis reports no confidence intervals or significance tests, so the tables provide descriptive comparisons for those tested conditions.
The document identifies itself as arXiv:2608.24558v1, dated 25 Aug 2026, and states that source code is available at https://github.com/Amitmils/AmbiDiffEnc.
Paper data and sources
Original title: Array-Agnostic Ambisonics Encoding via Diffusion Posterior Sampling
Authors: Amit Milstein, Nir Shlezinger, Boaz Rafaely
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text