Research map / Architectures and efficiency / RNNs and normalization
Fully Neural Network Based Speech Recognition on Mobile and Embedded Devices
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 input gate (i-SRU) to SRU, with 1-D conv (ablation) | 6x700 i-SRU vs 6x800 SRU · WSJ eval92 | WER (HCLM): 4.90 -> 4.90 | No clear effect |
| add depth-wise 1-D conv at input of each recurrent layer (ablation) | 6x700 i-SRU · WSJ eval92 | WER (greedy): 45.22 -> 19.07 | Helped |
| add depth-wise 1-D conv at input of each recurrent layer (ablation) | 6x800 SRU · WSJ eval92 | WER (HCLM): 15.41 -> 4.90 | Helped |
| add depth-wise 1-D conv at input | 4x600 LSTM · WSJ eval92 | WER (HCLM): 8.75 -> 7.01 | Helped |
| use hierarchical character LM instead of character LM in decoding | 6x700 i-SRU, 1-D conv · WSJ eval92 | WER: 7.30 -> 4.90 | Helped |
| restrict 1-D conv to causal (past-only, 14 steps) instead of symmetric (ablation) | 6x700 i-SRU, 1-D conv · WSJ eval92 | WER (HCLM): 4.90 -> 5.32 | Hurt |
| use causal 1-D conv with 7 past steps instead of symmetric (-7,7) (ablation) | 6x700 i-SRU, 1-D conv · WSJ eval92 | WER (HCLM): 4.90 -> 5.55 | Hurt |
| train on additional WSJ data (167 h vs 81 h) | 6x700 i-SRU, 1-D conv · WSJ eval92 | WER (HCLM): 4.90 -> 3.73 | Helped |
| use more WSJ training data for word piece model | 6x700 i-SRU, 1-D conv word piece · WSJ eval92 | WER (WPLM): 10.50 -> 8.28 | Helped |
| increase word piece downsampling to x4 in conv layer (from x2) (ablation) | word piece AM · WSJ eval92 | WER (WPLM): 10.93 -> 11.83 | Hurt |
| add x2 recurrent-layer downsampling to x2 conv downsampling (ablation) | word piece AM · WSJ eval92 | WER (WPLM): 10.93 -> 10.50 | Helped |
| use word piece units instead of characters | 6x700 i-SRU, 1-D conv · LibriSpeech test-clean | WER (RNN LM): 13.78 -> 9.61 | Helped |
Closest papers by meaning
- Shallow RNN: Accurate Time-series Classification on Resource Constrained Devices (NeurIPS 2019 · similarity 0.88)
- Persistent RNNs: Stashing Recurrent Weights On-Chip (ICML 2016 · similarity 0.86)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.85)
- Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.84)
- HAINAN: Fast and Accurate Transducer for Hybrid-Autoregressive ASR (ICLR 2025 · similarity 0.84)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.83)
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.