Research map / Architectures and efficiency / RNNs and normalization
Word-Level Speech Recognition With a Letter to Word Encoder
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 |
|---|---|---|---|
| Predict words with a letter-to-word encoder instead of word pieces | LibriSpeech 960h seq2seq, without LM · LibriSpeech 960h test-clean | WER: 3.05 -> 2.87 | Helped |
| Predict words with a letter-to-word encoder instead of word pieces | LibriSpeech 960h seq2seq, without LM · LibriSpeech 960h test-other | WER: 7.01 -> 6.69 | Helped |
| Use word-level rather than word-piece outputs with an external LM | LibriSpeech 960h seq2seq · LibriSpeech 960h test-clean | WER: 2.58 -> 2.78 | No clear effect |
| Use word-level rather than word-piece outputs with an external LM | LibriSpeech 960h seq2seq · LibriSpeech 960h test-other | WER: 6.15 -> 6.18 | No clear effect |
| Use word-level rather than word-piece outputs | LibriSpeech train-clean-100 seq2seq, without LM · LibriSpeech train-clean-100 test-clean | WER: 9.5 -> 8.6 | Helped |
| Use word-level rather than word-piece outputs | LibriSpeech train-clean-100 seq2seq, without LM · LibriSpeech train-clean-100 test-other | WER: 23.3 -> 21.9 | Helped |
| Use word-level CTC instead of word-piece CTC with an external LM | LibriSpeech train-clean-100 · LibriSpeech train-clean-100 test-clean | WER: 10.3 -> 6.8 | Helped |
| Use word-level CTC instead of word-piece CTC with an external LM | LibriSpeech train-clean-100 · LibriSpeech train-clean-100 test-other | WER: 24.0 -> 19.4 | Helped |
| Increase acoustic-model stride from 8 to 16 (ablation) | LibriSpeech 960h word-level CTC without LM · LibriSpeech 960h test-clean | WER: 3.15 -> 3.24 | Hurt |
| Increase acoustic-model stride from 8 to 16 (ablation) | LibriSpeech 960h word-level CTC without LM · LibriSpeech 960h test-other | WER: 7.45 -> 7.73 | Hurt |
| Increase acoustic-model stride from 8 to 16 (ablation) | LibriSpeech 960h word-level seq2seq with 4-gram LM · LibriSpeech 960h test-clean | WER: 2.78 -> 2.93 | Hurt |
| Increase acoustic-model stride from 8 to 16 (ablation) | LibriSpeech 960h word-level seq2seq with 4-gram LM · LibriSpeech 960h test-other | WER: 6.18 -> 6.90 | Hurt |
Closest papers by meaning
- UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation (ICLR 2025 · similarity 0.85)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.84)
- SyllableLM: Learning Coarse Semantic Units for Speech Language Models (ICLR 2025 · similarity 0.84)
- Language Model Memory and Memory Models for Language (NeurIPS 2026 · similarity 0.84)
- Training Neural Networks as Recognizers of Formal Languages (ICLR 2025 · similarity 0.84)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (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.