Research map / Architectures and efficiency / Attention and transformers
BiCycle: Group-wise Recursive Transformer Based on ASR Mechanism
Min Ho Jang, Eun Seo Seo, Jin Young Kim, Hyeongsoo Lim, Ji Won Yoon
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 BiCycle group-wise recursion instead of CYCLE | Conformer-CTC · LibriSpeech test-other (depth ~16-18) | WER: 9.30% -> 9.11% | Improved |
| increase recursion steps (depth 16 to 20) in BiCycle (ablation) | BiCycle · LibriSpeech test-other | WER: 9.11% -> 8.93% | Improved |
| increase recursion depth for conventional RT (CYCLE 18 to 27) (ablation) | CYCLE · LibriSpeech test-other | WER: 9.30% -> 10.23% | Hurt |
| use SEQUENCE structure with depth 27 (ablation) | SEQUENCE · LibriSpeech | WER: did not converge | Hurt |
| add recursion to BiCycle (vs w/o recursion) (ablation) | BiCycle · Common Voice 7.0 French test | WER: 13.16% -> 12.38% | Improved |
| use Relaxed Recursive Transformer with LoRA-style relaxation | Relaxed Recursive Transformer · LibriSpeech test-other | WER: 11.44% -> 16.55% | Hurt |
| add logit KD to BiCycle (ablation) | BiCycle · LibriSpeech test-other | WER: 9.11% -> 8.64% | Improved |
| use group-wise feature distillation instead of FitNets final-layer distillation (ablation) | BiCycle · LibriSpeech test-other | WER: 8.51% -> 8.43% (depth 20); 8.53% -> 8.14% (depth 23) | Improved |
| use GFD instead of FitNets (ablation) | BiCycle depth 23 · LibriSpeech test-clean | WER: 3.50% -> 3.53% | Mixed |
| single recursion step in each group (RP:1 RK:1) (ablation) | BiCycle · LibriSpeech test-other | WER: 11.44% -> 9.88% | Improved |
Closest papers by meaning
- Depth through Recurrence: Towards Ultra-Efficient On-Device ASR (NeurIPS 2026 · similarity 0.89)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.89)
- Zipformer: A faster and better encoder for automatic speech recognition (ICLR 2024 · similarity 0.87)
- LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration (AAAI 2025 · similarity 0.86)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.85)
- Aligner-Encoders: Self-Attention Transformers Can Be Self-Transducers (NeurIPS 2024 · similarity 0.85)
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.
