Research map / Label-efficient and robust learning / Recognition and long-tail
Analyzing Hidden Representations in End-to-End Automatic Speech Recognition Systems
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 |
|---|---|---|---|
| Use first convolutional-layer representations instead of input spectrograms | DeepSpeech2 · TIMIT phone frame classification | accuracy | Helped |
| Add a second convolution after the first convolution | DeepSpeech2 · TIMIT phone frame classification | accuracy | Hurt |
| Add recurrent layers up to rnn5 | DeepSpeech2 · TIMIT phone frame classification | accuracy | Helped |
| Add recurrent layers beyond rnn5 | DeepSpeech2 · TIMIT phone frame classification | accuracy | Hurt |
| Use DeepSpeech2-light instead of DeepSpeech2 (ablation) | End-to-end ASR model · LibriSpeech test-clean | WER: 12 -> 15 | Hurt |
| Use DeepSpeech2-light instead of DeepSpeech2 for phone representations (ablation) | End-to-end ASR model · TIMIT phone frame classification | accuracy: 33.67% -> 37.77% | Helped |
| Use DeepSpeech2 rather than DeepSpeech2-light convolutional representations (ablation) | End-to-end ASR model · TIMIT phone frame classification | accuracy: +1% and +4% | Helped |
| Remove time strides from the convolutional layers during feature generation (ablation) | DeepSpeech2 · TIMIT phone frame classification | accuracy | Helped |
| Use a window of features around the current frame instead of a single frame (ablation) | DeepSpeech2 without convolutional strides · TIMIT test phone frame classification | accuracy: +10-15% | Helped |
| Increase the feature-window width (ablation) | DeepSpeech2 without convolutional strides · TIMIT test phone frame classification | accuracy | Helped |
| Use a reduced 48-phone label set instead of the full 60-phone set (ablation) | DeepSpeech2 representations · TIMIT phone frame classification | accuracy | No clear effect |
| Replace the one-hidden-layer classifier with a linear classifier (ablation) | Frame representation evaluation · TIMIT phone frame classification | accuracy | Hurt |
Closest papers by meaning
- Untangling in Invariant Speech Recognition (NeurIPS 2019 · similarity 0.89)
- Understanding the Representation and Computation of Multilayer Perceptrons: A Case Study in Speech Recognition (ICML 2017 · similarity 0.89)
- Toward a Better Understanding of Deep Neural Network Based Acoustic Modelling: An Empirical Investigation (AAAI 2016 · similarity 0.88)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.85)
- AND: Audio Network Dissection for Interpreting Deep Acoustic Models (ICML 2024 · similarity 0.84)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 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.