Research map / Label-efficient and robust learning / Self-supervised and contrastive
Variable-rate hierarchical CPC leads to acoustic unit discovery in speech
Santiago Cuervo, Adrian Lancucki, Ricard Marxer, Paweł Rychlikowski, Jan K Chorowski
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 variable-rate downsampling learned from the high-level CPC objective instead of no downsampling | Two-level CPC · LibriSpeech clean 100 test | Frame accuracy: 67.49 -> 72.57 | Improved |
| Use variable-rate downsampling instead of fixed-rate downsampling | Two-level CPC · LibriSpeech clean 100 test | Frame accuracy: 67.92 -> 72.57 | Improved |
| Add high-level CPC supervision with learned variable-rate segmentation | Frame-level CPC representations · LibriSpeech clean 100 test | Frame accuracy: 68.60 -> 72.57 | Improved |
| Use variable-rate segmentation rather than phone-label-supervised downsampling | Frame-level CPC representations · LibriSpeech clean 100 test | Frame accuracy: 71.01 -> 72.57 | Improved |
| Use variable-rate learned downsampling rather than no downsampling | High-level CPC · LibriSpeech clean 100 test | Phone accuracy: 83.41 -> 78.93 | Hurt |
| Use phone-label-supervised downsampling rather than learned variable-rate downsampling | High-level CPC · LibriSpeech clean 100 test | Phone accuracy: 78.93 -> 85.74 | Improved |
| Use the proposed segmentation instead of SCPC segmentation | Phone boundary detection · LibriSpeech clean 100 test | Precision: 64.05 -> 79.94 | Improved |
| Use the proposed segmentation instead of SCPC segmentation | Phone boundary detection · TIMIT test (non-speech removed) | F1: 85.31 -> 80.73 | Hurt |
| Add adjacent negative sampling to enforce distinguishability of successive units (ablation) | Full hierarchical CPC · LibriSpeech clean 100 test | Frame accuracy: 64.13 -> 72.52 | Improved |
| Add adjacent negative sampling to enforce distinguishability of successive units (ablation) | Full hierarchical CPC · LibriSpeech clean 100 test | R-val: 21.10 -> 81.09 | Improved |
| Add average segment-length policy regularization (ablation) | Hierarchical CPC with adjacent negative sampling · LibriSpeech clean 100 test | Frame accuracy: 66.87 -> 72.52 | Improved |
| Add average segment-length policy regularization (ablation) | Hierarchical CPC with adjacent negative sampling · LibriSpeech clean 100 test | R-val: 0.0 -> 81.09 | Improved |
Closest papers by meaning
- Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data (NeurIPS 2017 · similarity 0.86)
- Self-Supervised Disentangled Representation Learning for Robust Target Speech Extraction (AAAI 2024 · similarity 0.85)
- Multi-resolution HuBERT: Multi-resolution Speech Self-Supervised Learning with Masked Unit Prediction (ICLR 2024 · similarity 0.85)
- CA-SSLR: Condition-Aware Self-Supervised Learning Representation for Generalized Speech Processing (NeurIPS 2024 · similarity 0.84)
- UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation (ICLR 2025 · similarity 0.84)
- SECodec: Structural Entropy-based Compressive Speech Representation Codec for Speech Language Models (AAAI 2025 · 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.
