Research map / Large language models / Language, safety and interpretability
Toward a Better Understanding of Deep Neural Network Based Acoustic Modelling: An Empirical Investigation
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 |
|---|---|---|---|
| increase DNN size from 3.6M to 50M parameters (ablation) | 5-layer DNN · Switchboard 33h dev frames | frame accuracy: 66.20 -> 86.06 | Helped |
| increase DNN size from 3.6M to 50M parameters (ablation) | 5-layer DNN · Switchboard 33h HUB5 EV | WER: 21.2 -> 20.9 | No clear effect |
| increase DNN size from 3.6M to 50M parameters (ablation) | 5-layer DNN · Switchboard 33h training subset | WER: 17.52 -> 11.56 | Helped |
| widen context window from ±10 to ±20 frames (ablation) | 50M DNN · Switchboard 33h HUB5 EV | WER: 20.9 -> 22.0 | Hurt |
| widen context window from ±10 to ±20 frames (ablation) | 50M DNN · Switchboard 33h dev frames | frame accuracy: 86.06 -> 93.05 | Helped |
| add dropout (ablation) | DNNs of 3.6M, 10M, 50M · Switchboard 33h HUB5 EV | WER: 21.2 -> 20.8 | Helped |
| add dropout (ablation) | 10M DNN · Switchboard 33h HUB5 EV | WER: 20.8 -> 20.5 | Helped |
| add dropout (ablation) | 50M DNN · Switchboard 33h HUB5 EV | WER: 20.9 -> 20.7 | Helped |
| early stopping at best test WER (ablation) | 50M DNN · Switchboard 33h HUB5 EV | WER: 20.9 -> 20.7 | Helped |
| raise max momentum from 0.90 to 0.99 (ablation) | classical momentum 3.6M DNN · Fisher 210h HUB5 EV | WER: 22.8 -> 20.6 | Helped |
| use NAG instead of classical momentum at max momentum 0.99 (ablation) | 3.6M DNN · Fisher 210h frame accuracy | frame accuracy: 55.26 -> 55.39 | Helped |
| lower initial learning rate from 0.01 to 0.001 (ablation) | 3.6M DNN · Fisher 210h HUB5 EV | WER: more than 1% absolute higher | Hurt |
Closest papers by meaning
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.90)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.87)
- AND: Audio Network Dissection for Interpreting Deep Acoustic Models (ICML 2024 · similarity 0.87)
- Zipformer: A faster and better encoder for automatic speech recognition (ICLR 2024 · similarity 0.87)
- The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning (ICML 2025 · similarity 0.86)
- Speech Robust Bench: A Robustness Benchmark For Speech Recognition (ICLR 2025 · 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.