Research map / Learning theory and algorithms / Kernels and regression theory
Kernel Approximation Methods for Speech Recognition
Avner May, Alireza Bagheri Garakani, Zhiyun Lu, Dong Guo, Kuan Liu, Aurélien Bellet, Linxi Fan, Michael Collins, Daniel Hsu, Brian Kingsbury, Michael Picheny, Fei Sha
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 |
|---|---|---|---|
| add a linear bottleneck (ablation) | Sparse Gaussian kernel model · Bengali dev | token error rate (%): 73.0 -> 71.5 | Improved |
| add ERLL-based learning rate decay together with bottleneck (ablation) | Sparse Gaussian kernel model · Bengali dev | token error rate (%): 71.5 -> 70.9 | Improved |
| add ERLL-based learning rate decay alone (ablation) | Sparse Gaussian kernel model · Bengali dev | token error rate (%): 73.0 -> 73.0 | No clear effect |
| add random feature selection (ablation) | Laplacian kernel model · Cantonese dev | token error rate (%): 69.9 -> 68.4 | Improved |
| add random feature selection (ablation) | Sparse Gaussian kernel model · Bengali dev | token error rate (%): 73.0 -> 72.9 | Improved |
| add random feature selection (ablation) | Gaussian kernel model · Bengali dev | token error rate (%): 72.6 -> 74.1 | Hurt |
| add linear bottleneck and ERLL decay (ablation) | DNN with 4000 hidden units · Bengali dev | token error rate (%): 71.1 -> 70.2 | Improved |
| add ERLL-based learning rate decay (ablation) | DNN with 2000 hidden units · TIMIT dev | token error rate (%): 19.0 -> 19.2 | Hurt |
| use heldout ERLL instead of cross-entropy for stopping/decay (ablation) | kernel and DNN acoustic models · correlation with TER on Bengali | token error rate (%): 0.31 -> 0.91 | Improved |
| use heldout ERLL instead of cross-entropy for stopping/decay (ablation) | kernel and DNN acoustic models · correlation with TER on TIMIT | token error rate (%): -0.21 -> 0.95 | Improved |
| increase number of random features from 5000 to 100000 (ablation) | kernel models · BN-50 | cross-entropy and TER | Improved |
| compare best kernel model to best DNN | kernel vs DNN · TIMIT test | token error rate (%): 20.5 -> 20.4 | Mixed |
Closest papers by meaning
- Toward a Better Understanding of Deep Neural Network Based Acoustic Modelling: An Empirical Investigation (AAAI 2016 · similarity 0.90)
- Gaussian Quadrature for Kernel Features (NeurIPS 2017 · similarity 0.87)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.86)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.86)
- Zipformer: A faster and better encoder for automatic speech recognition (ICLR 2024 · similarity 0.85)
- Stochastic Kernel Regularisation Improves Generalisation in Deep Kernel Machines (NeurIPS 2024 · similarity 0.84)
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.
