mlsanityData to 14 Oct 2026

A Study of the Scale Invariant Signal to Distortion Ratio in Speech Separation with Noisy References

arXiv 2508.14623 ↗First posted 20 Aug 2025eess.AS cs.AI cs.SDcited by 4

Not on the map. Why?

Abstract (arXiv)

This paper examines the implications of using the Scale-Invariant Signal-to-Distortion Ratio (SI-SDR) as both evaluation and training objective in supervised speech separation, when the training references contain noise, as is the case with the de facto benchmark WSJ0-2Mix. A derivation of the SI-SDR with noisy references reveals that noise limits the achievable SI-SDR, or leads to undesired noise in the separated outputs. To address this, a method is proposed to enhance references and augment the mixtures with WHAM!, aiming to train models that avoid learning noisy references. Two models trained on these enhanced datasets are evaluated with the non-intrusive NISQA.v2 metric. Results show reduced noise in separated speech but suggest that processing references may introduce artefacts, limiting overall quality gains. Negative correlation is found between SI-SDR and perceived noisiness across models on the WSJ0-2Mix and Libri2Mix test sets, underlining the conclusion from the derivation.

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Builds on 37 of its 59 references that are in the map. First posted on arXiv 20 Aug 2025.

Builds on

37 of 59 references.

References · 37

  1. 2018SDR - half-baked or well done?cited by 423
  2. 2020DNSMOS: A Non-Intrusive Perceptual Objective Speech Quality metric to evaluate Noise Suppressorscited by 241
  3. 2017Supervised Speech Separation Based on Deep Learning: An Overviewcited by 216
  4. 2015Deep clustering: Discriminative embeddings for segmentation and separationcited by 212
  5. 2021NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasetscited by 192
  6. 2016Permutation Invariant Training of Deep Models for Speaker-Independent Multi-talker Speech Separationcited by 191
  7. 2020Attention is All You Need in Speech Separationcited by 185
  8. 2019Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separationcited by 182
  9. 2019WHAM!: Extending Speech Separation to Noisy Environmentscited by 156
  10. 2020LibriMix: An Open-Source Dataset for Generalizable Speech Separationcited by 155
  11. 2017TasNet: time-domain audio separation network for real-time, single-channel speech separationcited by 115
  12. 2021MetricGAN+: An Improved Version of MetricGAN for Speech Enhancementcited by 114
25 more
  1. 2020Wavesplit: End-to-End Speech Separation by Speaker Clusteringcited by 75
  2. 2020Dual-Path Transformer Network: Direct Context-Aware Modeling for End-to-End Monaural Speech Separationcited by 68
  3. 2024EARS: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberationcited by 64
  4. 2019WHAMR!: Noisy and Reverberant Single-Channel Speech Separationcited by 61
  5. 2024Open-Source Conversational AI with SpeechBrain 1.0cited by 56
  6. 2020Sudo rm -rf: Efficient Networks for Universal Audio Source Separationcited by 51
  7. 2022TF-GridNet: Making Time-Frequency Domain Models Great Again for Monaural Speaker Separationcited by 48
  8. 2016Single-Channel Multi-Speaker Separation using Deep Clusteringcited by 45
  9. 2023MossFormer: Pushing the Performance Limit of Monaural Speech Separation using Gated Single-Head Transformer with Convolution-Augmented Joint Self-Attentionscited by 27
  10. 2019Divide and Conquer: A Deep CASA Approach to Talker-independent Monaural Speaker Separationcited by 20
  11. 2017Speaker-independent Speech Separation with Deep Attractor Networkcited by 17
  12. 2019FurcaNeXt: End-to-end monaural speech separation with dynamic gated dilated temporal convolutional networkscited by 15
  13. 2020An empirical study of Conv-TasNetcited by 15
  14. 2020On Cross-Corpus Generalization of Deep Learning Based Speech Enhancementcited by 11
  15. 2024TF-Locoformer: Transformer with Local Modeling by Convolution for Speech Separation and Enhancementcited by 11
  16. 2018End-to-End Speech Separation with Unfolded Iterative Phase Reconstructioncited by 10
  17. 2019Analysis of Deep Clustering as Preprocessing for Automatic Speech Recognition of Sparsely Overlapping Speechcited by 8
  18. 2019Mixup-breakdown: a consistency training method for improving generalization of speech separation modelscited by 8
  19. 202530+ Years of Source Separation Research: Achievements and Future Challengescited by 7
  20. 2024Boosting Unknown-number Speaker Separation with Transformer Decoder-based Attractorcited by 6
  21. 2024Rethinking Processing Distortions: Disentangling the Impact of Speech Enhancement Errors on Speech Recognition Performancecited by 6
  22. 2022Low-Latency Speech Separation Guided Diarization for Telephone Conversationscited by 5
  23. 2022Deep neural network techniques for monaural speech enhancement: state of the art analysiscited by 3
  24. 2023Frequency bin-wise single channel speech presence probability estimation using multiple DNNscited by 1
  25. 2024Robustness of Speech Separation Models for Similar-pitch Speakerscited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 1; 2026: 3.

4 papers.

  1. 2025Exploring Resolution-Wise Shared Attention in Hybrid Mamba-U-Nets for Improved Cross-Corpus Speech Enhancementcited by 0
  2. 2026DynFOA: Generating First-Order Ambisonics with Conditional Diffusion for Dynamic and Acoustically Complex 360-Degree Videoscited by 0
  3. 2026VorTEX: Various overlap ratio for Target speech EXtractioncited by 0
  4. 2026DynFOA: Generating First-Order Ambisonics with Conditional Diffusion for Dynamic and Acoustically Complex 360-Degree Videoscited by 0