Research map / Architectures and efficiency / RNNs and normalization
Untangling in Invariant Speech Recognition
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 |
|---|---|---|---|
| Train on word recognition rather than use initial random weights | CNN · WSJ word manifolds | manifold capacity | Helped |
| Train on character-level ASR rather than use initial random weights | Deep Speech 2 · LibriSpeech word manifolds | manifold capacity | Helped |
| Train the CNN for word recognition instead of speaker recognition (ablation) | CNN · WSJ word manifolds | manifold capacity | Helped |
| Train the CNN for speaker recognition instead of word recognition (ablation) | CNN · CNN speaker manifolds | manifold capacity | Helped |
| Train the CNN for word recognition rather than speaker recognition (ablation) | CNN · CNN speaker manifolds | manifold capacity | Hurt |
| Train the CNN jointly on word and speaker recognition instead of a single task (ablation) | CNN · Word and speaker manifolds | manifold capacity | Mixed |
| Train the CNN jointly on word and speaker recognition instead of a single task (ablation) | CNN · Word and speaker manifolds | manifold capacity | Hurt |
| Train Deep Speech 2 on ASR rather than use initial weights | Deep Speech 2 · Speaker manifolds | manifold capacity | Hurt |
| Measure word manifolds at each recurrent time step rather than using a projection across all features | Deep Speech 2 · LibriSpeech word manifolds | relative capacity: ~1.4 -> ~3.6 | Helped |
| Reduce the context window around the word (ablation) | Deep Speech 2 · LibriSpeech word manifolds | manifold capacity | No clear effect |
| Permute the category labels instead of preserving true manifold labels (ablation) | CNN and Deep Speech 2 · Word, speaker, character, phoneme, and part-of-speech manifolds | manifold capacity | Hurt |
| Use the untrained Deep Speech 2 model rather than the trained model (ablation) | Deep Speech 2 · Word manifolds across input time steps | manifold capacity | Hurt |
Closest papers by meaning
- Understanding the Representation and Computation of Multilayer Perceptrons: A Case Study in Speech Recognition (ICML 2017 · similarity 0.90)
- Analyzing Hidden Representations in End-to-End Automatic Speech Recognition Systems (NeurIPS 2017 · similarity 0.89)
- AND: Audio Network Dissection for Interpreting Deep Acoustic Models (ICML 2024 · similarity 0.88)
- Toward a Better Understanding of Deep Neural Network Based Acoustic Modelling: An Empirical Investigation (AAAI 2016 · similarity 0.88)
- Understanding Adaptive, Multiscale Temporal Integration In Deep Speech Recognition Systems (NeurIPS 2021 · similarity 0.87)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (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.