Research map / Image generation / Image restoration
ArrayDPS: Unsupervised Blind Speech Separation with a Diffusion Prior
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ArrayDPS, an unsupervised, array-agnostic diffusion posterior sampling method for blind speech separation that approximates the intractable multichannel likelihood via FCP relative-RIR estimation and IVA initialization, using a single-speaker speech diffusion prior.
Problem
Blind speech separation from microphone-array mixtures where the array geometry, room impulse responses, and sources are all unknown, so the likelihood for diffusion posterior sampling is intractable.
Approach
Extends diffusion posterior sampling to multichannel blind separation by approximating the likelihood at each diffusion step: relative RIRs are estimated by maximum likelihood (Forward Convolutive Prediction) from the current denoised source estimates, and the resulting multichannel reconstruction error gives the likelihood score; IVA provides initialization and early-step filter guidance.
Key results
- On 3-channel SMS-WSJ, ArrayDPS-A reaches 15.8 dB SDR and 15.0 dB SI-SDR, beating the unsupervised UNSSOR (15.4 / 14.4) and comparable to supervised TF-GridNet-SMS (16.2 SDR).
- ArrayDPS-A-ML (selecting the max-likelihood sample of 5) achieves 16.9 dB SDR, slightly above supervised TF-GridNet-SMS on SDR.
- On 4-channel ad-hoc Spatialized WSJ0-2Mix, ArrayDPS-A has 16.1 dB SDR vs. 15.2 for UNSSOR and 15.8 for supervised TF-GridNet-Spatial.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use an anechoic clean-speech prior instead of a reverberant-speech prior (ablation) | ArrayDPS · SMS-WSJ 3-channel | SDR (dB): 15.1$\pm$1.1 -> 15.8$\pm$1.2 | Helped |
| Remove reference-channel guidance (ablation) | ArrayDPS-A · SMS-WSJ 3-channel | SDR (dB): 15.8$\pm$1.2 -> 14.9$\pm$1.3 | Hurt |
| Remove IVA initialization (ablation) | ArrayDPS-A · SMS-WSJ 3-channel | SDR (dB): 15.8$\pm$1.2 -> 8.5$\pm$4.6 | Hurt |
| Remove both IVA initialization and reference-channel guidance (ablation) | ArrayDPS-A · SMS-WSJ 3-channel | SDR (dB): 8.5$\pm$4.6 -> 0.9$\pm$1.6 | Hurt |
| Select the maximum metric-scoring sample among 5 samples | ArrayDPS-A · SMS-WSJ 3-channel | SDR (dB): 15.8$\pm$1.2 -> {17.2} | Helped |
| Select the maximum-likelihood sample from 5 samples | ArrayDPS-A · SMS-WSJ 3-channel | SDR (dB): 15.8$\pm$1.2 -> 16.9 | Helped |
| Select the maximum-likelihood sample from 5 samples | ArrayDPS-A · Spatialized WSJ0-2Mix 4-channel | SDR: 16.1\n$\pm$0.6 -> 16.6 | Helped |
| Remove IVA initialization (ablation) | ArrayDPS-A · Spatialized WSJ0-2Mix 4-channel | SDR: 16.1\n$\pm$0.6 -> 6.7\\$\pm$4.5 | Hurt |
Surprising findings
- Removing IVA initialization causes a severe drop (SDR 15.8 to 8.5 dB with high variance), so the method is sensitive to initialization.
- Using an anechoic clean-speech prior works better than a reverberant-speech prior, attributed to the behavior of the FCP estimator.
Limitations
- Generative sampling is slow and needs multiple samples; the best-of-5 (Max) result is impractical, and the ML selection only partly closes the gap to supervised methods.
- Performance depends heavily on IVA initialization and reference-channel guidance; without them results are unstable.
- Evaluated only on 2-3 speaker simulated mixtures plus a small set of 15 real recordings.
Experimental setup
Datasets: LibriTTS SMS-WSJ Spatialized WSJ0-2Mix Spatialized WSJ0-3Mix
Benchmarks: SMS-WSJ Spatialized WSJ0-2Mix Spatialized WSJ0-3Mix
Metrics: SDR SI-SDR PESQ eSTOI
Baselines: Spatial Clustering IVA UNSSOR TF-GridNet-SMS TF-GridNet-Spatial
Builds on: diffusion posterior sampling EDM Forward Convolutive Prediction Independent Vector Analysis UNSSOR
Closest papers by meaning
- Unsupervised Single-Channel Audio Separation with Diffusion Source Priors (AAAI 2026 · similarity 0.90)
- Separate and Diffuse: Using a Pretrained Diffusion Model for Better Source Separation (ICLR 2024 · similarity 0.87)
- ZeroSep: Separate Anything in Audio with Zero Training (NeurIPS 2025 · similarity 0.86)
- Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation (NeurIPS 2024 · similarity 0.83)
- Shallow Diffusion for Fast Speech Enhancement (Student Abstract) (AAAI 2024 · similarity 0.83)
- DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.