Research map / Label-efficient and robust learning / Semi-supervised learning
Int*-Match: Balancing Intra-Class Compactness and Inter-Class Discrepancy for Semi-Supervised Speaker Recognition
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 Int*-Match, a semi-supervised speaker recognition method that selects pseudo labels using an adaptive inter-class threshold from labeled data and an intra-class threshold that tracks pseudo-label compactness.
Problem
Semi-supervised speaker recognition with few labels: threshold-based SSL methods from image classification select pseudo labels only by inter-class discrepancy, ignoring intra-class compactness that matters for open-set speaker verification.
Approach
Derive an inter-class threshold from correctly predicted labeled data and an intra-class threshold from labeled-data compactness; adaptively lower the inter-class threshold and raise the intra-class threshold depending on the compactness of selected pseudo labels, balancing pseudo-label quality and quantity.
Key results
- Int*-Match achieves 1.00% EER on VoxCeleb1-O with 20-50% labeled data, only 0.06% behind fully supervised learning (0.94%) per the paper's text.
- With 20 labels per class, Int*-Match reaches 1.28% EER on VoxCeleb1-O versus 1.30% for GCN, which uses more labeled data.
- Int*-Match keeps pseudo-label quality near 100% during training in the 20-label setting, per the paper's analysis.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Int*-Match instead of FlexMatch | ECAPA-TDNN semi-supervised speaker recognition · VoxCeleb1-O (40% labels) | EER (%): 1.21 -> 1.00 | Helped |
| use Int*-Match instead of FlexMatch | ECAPA-TDNN semi-supervised speaker recognition · VoxCeleb1-E (40% labels) | EER (%): 1.39 -> 1.33 | Helped |
| use Int*-Match instead of FlexMatch | ECAPA-TDNN semi-supervised speaker recognition · VoxCeleb1-H (40% labels) | EER (%): 2.58 -> 2.44 | Helped |
| use Int*-Match instead of FlexMatch | ECAPA-TDNN semi-supervised speaker recognition · VoxCeleb1-O (40% labels) | minDCF: 0.080 -> 0.075 | Helped |
| use Int*-Match instead of FlexMatch | ECAPA-TDNN semi-supervised speaker recognition · VoxCeleb1-E (40% labels) | minDCF: 0.091 -> 0.085 | Helped |
| use Int*-Match instead of FlexMatch | ECAPA-TDNN semi-supervised speaker recognition · VoxCeleb1-H (40% labels) | minDCF: 0.157 -> 0.150 | Helped |
| set initial intra-class threshold τ to 0.6 rather than 0.65 (ablation) | Int*-Match · VoxCeleb1-O (2 labels/class) | EER (%): 1.45 -> 1.55 | Hurt |
| set initial intra-class threshold τ to 0.625 rather than 0.65 (ablation) | Int*-Match · VoxCeleb1-O (2 labels/class) | EER (%): 1.45 -> 1.56 | Hurt |
| set initial intra-class threshold τ to 0.675 rather than 0.65 (ablation) | Int*-Match · VoxCeleb1-O (2 labels/class) | EER (%): 1.45 -> 1.55 | Hurt |
| set initial intra-class threshold τ to 0.7 rather than 0.65 (ablation) | Int*-Match · VoxCeleb1-O (2 labels/class) | EER (%): 1.45 -> 1.71 | Hurt |
| lower the initial intra-class threshold (ablation) | Int*-Match · pseudo-label selection | pseudo-label quality | Hurt |
Surprising findings
- Pseudo-label baseline and several threshold-based SSL methods (FixMatch, Dash) perform worse than or similar to supervised learning in the speaker setting, showing that image-classification SSL does not transfer directly.
Limitations
- Threshold initialization tau (0.65) is set by hyperparameter search and performance varies with it
- Comparisons with GCL and GCN use different labeled-data settings, so direct comparison is not fully controlled
Experimental setup
Datasets: VoxCeleb2 MUSAN RIR
Benchmarks: VoxCeleb1-O VoxCeleb1-E VoxCeleb1-H
Metrics: EER minDCF
Baselines: Pseudo label FixMatch FlexMatch Dash FreeMatch SoftMatch GCL GCN GLL Supervised learning
Builds on: FixMatch FlexMatch AAM-softmax ECAPA-TDNN
Closest papers by meaning
- Leveraging Local Variance for Pseudo-Label Selection in Semi-supervised Learning (AAAI 2024 · similarity 0.88)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.86)
- Barely-Supervised Learning: Semi-supervised Learning with Very Few Labeled Images (AAAI 2022 · similarity 0.85)
- Improving Barely Supervised Learning by Discriminating Unlabeled Samples with Super-Class (NeurIPS 2022 · similarity 0.84)
- On Non-Random Missing Labels in Semi-Supervised Learning (ICLR 2022 · similarity 0.84)
- SpeakerLM: End-to-End Versatile Speaker Diarization and Recognition with Multimodal Large Language Models (AAAI 2026 · similarity 0.84)
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.