Research map / Architectures and efficiency / RNNs and normalization
Voice Separation with an Unknown Number of Multiple Speakers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed MulCat recurrent separation model with multiscale and speaker-identity losses | voice separation · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 20.1 vs 18.8; 16.9 vs 14.7*; 12.9 vs 8.5*; 10.6 vs 8.7* | Helped |
| add the speaker identity loss (ablation) | DPRNN · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 18.08 -> 18.42; 14.72 -> 14.91; 10.37 -> 11.29; 8.65 -> 9.01 | Helped |
| add the MulCat gating block (ablation) | Ours-gating · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 19.42 -> 20.12; 15.73 -> 16.85; 11.22 -> 12.88; 8.95 -> 10.56 | Helped |
| add multiloss (ablation) | Ours-multiloss · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 18.93 -> 20.12; 13.86 -> 16.85; 10.54 -> 12.88; 8.75 -> 10.56 | Helped |
| add speaker identity loss (ablation) | Ours-IDloss · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 19.76 -> 20.12; 16.63 -> 16.85; 12.60 -> 12.88; 10.20 -> 10.56 | Helped |
| replace MulCat gating with a conventional LSTM (ablation) | proposed method · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 20.12 -> 19.42; 16.85 -> 15.73; 12.88 -> 11.22; 10.56 -> 8.95 | Hurt |
| remove the multiloss (ablation) | proposed method · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 20.12 -> 18.93; 16.85 -> 13.86; 12.88 -> 10.54; 10.56 -> 8.75 | Hurt |
| remove the speaker identity loss (ablation) | proposed method · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 20.12 -> 19.76; 16.85 -> 16.63; 12.88 -> 12.60; 10.56 -> 10.20 | Hurt |
| use encoder kernel sizes L=2, 4, 8, or 16 instead of the default L=8 (ablation) | proposed method · WSJ0-2mix | SI-SNRi: 18.94; 19.91; 19.76; 18.16 (L=2, 4, 8, 16) | Mixed |
| average shifted test-time versions (ablation) | Ours · WSJ0-2mix through WSJ0-5mix | SI-SNRi: 20.12 -> 20.31; 16.71 -> 17.01; 12.88 -> 13.11; 10.56 -> 10.93 | Helped |
| automatically select speaker count using output-channel activity | Ours · unknown speaker count, 2–5 speakers | SI-SNRi: 20.12 -> 18.63; 16.85 -> 14.62; 12.88 -> 11.48; 10.56 -> 10.37 | Hurt |
| use the proposed model on noisy and reverberated speech | voice separation · WHAM! and WHAMR! | SI-SNRi: 15.2 vs 13.9; 12.2 vs 10.3 | Helped |
Closest papers by meaning
- UNSSOR: Unsupervised Neural Speech Separation by Leveraging Over-determined Training Mixtures (NeurIPS 2023 · similarity 0.89)
- Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation (NeurIPS 2024 · similarity 0.88)
- Filter-Recovery Network for Multi-Speaker Audio-Visual Speech Separation (ICLR 2023 · similarity 0.88)
- Separate and Diffuse: Using a Pretrained Diffusion Model for Better Source Separation (ICLR 2024 · similarity 0.87)
- Speech Separation Using an Asynchronous Fully Recurrent Convolutional Neural Network (NeurIPS 2021 · similarity 0.86)
- ArrayDPS: Unsupervised Blind Speech Separation with a Diffusion Prior (ICML 2025 · 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.