Research map / Label-efficient and robust learning / Knowledge distillation
Revisiting the Entropy Semiring for Neural Speech Recognition
Oscar Chang, Dongseong Hwang, Olivier Siohan
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 |
|---|---|---|---|
| Add alignment entropy regularization | CTC LSTM, 22M · Librispeech Test-Clean | WER: 7.7 ± 0.3 -> 7.2 ± 0.3 | Improved |
| Add alignment entropy regularization | CTC Conformer, 9M · Librispeech Test-Clean | WER: 4.1 ± 0.2 -> 4.1 ± 0.2 | No clear effect |
| Add alignment entropy regularization | RNN-T LSTM, 25M · Librispeech Test-Clean | WER: 7.4 ± 0.3 -> 7.2 ± 0.3 | Improved |
| Add alignment entropy regularization | RNN-T Conformer, 10M · Librispeech Test-Clean | WER: 2.8 ± 0.2 -> 2.7 ± 0.2 | Improved |
| Add alignment entropy regularization | CTC LSTM, 22M · Librispeech Test-Other | WER: 21.8 ± 0.6 -> 20.8 ± 0.6 | Improved |
| Add alignment entropy regularization | CTC Conformer, 9M · Librispeech Test-Other | WER: 10.2 ± 0.4 -> 9.9 ± 0.4 | Improved |
| Add alignment entropy regularization | RNN-T LSTM, 25M · Librispeech Test-Other | WER: 24.0 ± 0.8 -> 23.1 ± 0.7 | Improved |
| Add alignment entropy regularization | RNN-T Conformer, 10M · Librispeech Test-Other | WER: 6.7 ± 0.3 -> 6.5 ± 0.3 | Improved |
| Add token-space state-wise distillation | 0.6B causal Conformer RNN-T · Librispeech Test-Clean | WER: 2.7 ± 0.2 -> 2.1 ± 0.2 | Improved |
| Add token-space state-wise distillation | 0.6B causal Conformer RNN-T · Relative latency | latency (ms): 0.0 -> +2.7 | Hurt |
| Remove token-space state-wise distillation, keeping alignment distillation (ablation) | Semiring distillation, 0.6B causal Conformer RNN-T · Relative latency | latency (ms): -5.2 -> -65.7 | Improved |
| Remove token-space state-wise distillation, keeping alignment distillation (ablation) | Semiring distillation, 0.6B causal Conformer RNN-T · Librispeech Test-Other | WER: 5.2 ± 0.3 -> 5.8 ± 0.3 | Hurt |
Closest papers by meaning
- Aligner-Encoders: Self-Attention Transformers Can Be Self-Transducers (NeurIPS 2024 · similarity 0.88)
- SECodec: Structural Entropy-based Compressive Speech Representation Codec for Speech Language Models (AAAI 2025 · similarity 0.87)
- Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs (NeurIPS 2026 · similarity 0.85)
- Heuristic-free Knowledge Distillation for Streaming ASR via Multi-modal Training (AAAI 2025 · similarity 0.85)
- Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework (ICLR 2024 · similarity 0.85)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · 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.
