Research map / Label-efficient and robust learning / Self-supervised and contrastive
Losses Can Be Blessings: Routing Self-Supervised Speech Representations Towards Efficient Multilingual and Multitask Speech Processing
Yonggan Fu, Yang Zhang, Kaizhi Qian, Zhifan Ye, Zhongzhi Yu, Cheng-I Lai, Yingyan Lin
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 |
|---|---|---|---|
| ORI mask initialization instead of weight-magnitude (WMI) (ablation) | S3-Router · LibriSpeech 10m | WER test-clean: 40.09 -> 36.68 | Improved |
| random (RI) instead of ORI mask initialization (ablation) | S3-Router · LibriSpeech 10m | WER test-clean: 36.68 -> 36.98 | Hurt |
| random mask initialization compared with weight finetuning (ablation) | S3-Router · LibriSpeech 10h | WER test-clean: 9.7 -> 9.78 | No clear effect |
| finetune connections with S3-Router instead of weights | wav2vec2-base · LibriSpeech 10m | WER test-clean: 38.40 -> 36.68 | Improved |
| add a 4-gram LM decoder | S3-Router vs weight finetuning · LibriSpeech 10m | WER test-clean: 24.83 -> 23.17 | Improved |
| S3-Router instead of weight finetuning on data2vec | data2vec · LibriSpeech 10m | WER test-clean: 30.75 -> 29.86 | Improved |
| S3-Router sparsity 0.10 instead of 0.09 on data2vec (ablation) | S3-Router · LibriSpeech 10m | WER test-clean: 29.86 -> 31.30 | Hurt |
| S3-Router instead of weight finetuning on 8 SUPERB tasks | wav2vec2-base · SUPERB speaker identification | Acc: 66.8 -> 71.07 | Improved |
| S3-Router instead of adaptor tuning | wav2vec2-base · LibriSpeech 10m | WER test-clean: 41.82 -> 36.68 | Improved |
| tune only FFN connections instead of self-attention only (ablation) | S3-Router · LibriSpeech 1h | WER test-clean: 24.39 -> 18.62 | Improved |
| tune both FFN and SA instead of FFN only (ablation) | S3-Router · LibriSpeech 1h | WER test-clean: 18.62 -> 19.26 | No clear effect |
| tune only the 3rd block group instead of the 1st (ablation) | S3-Router · LibriSpeech 1h | WER test-clean: 85.68 -> 24.57 | Improved |
Closest papers by meaning
- CA-SSLR: Condition-Aware Self-Supervised Learning Representation for Generalized Speech Processing (NeurIPS 2024 · similarity 0.87)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.87)
- LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration (AAAI 2025 · similarity 0.86)
- Multi-resolution HuBERT: Multi-resolution Speech Self-Supervised Learning with Masked Unit Prediction (ICLR 2024 · similarity 0.86)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.86)
- Speech Self-Supervised Learning Using Diffusion Model Synthetic Data (ICML 2024 · similarity 0.86)
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.
