Research map / Large language models / Language, safety and interpretability
Exploring Transfer Learning For End-to-End Spoken Language Understanding
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 |
|---|---|---|---|
| pretrain with ASR data, then finetune on the 10% SLU data | AT-AT on internal music SLU · Internal music SLU, 10% data | SemER relative to full-data E2E: +8.63 -> +1.45 | Helped |
| pretrain with ASR data, then finetune on the 10% SLU data | AT-AT on internal music SLU · Internal music SLU, 10% data | EM Accuracy relative to full-data E2E: -11.82 -> -1.49 | Helped |
| add LibriSpeech ASR pretraining | AT-AT on full-data internal music SLU · Internal music SLU, full data | SemER relative to E2E: +0.61 -> -1.16 | Helped |
| add MLM to SLU + ASR pretraining (ablation) | AT-AT on full-data internal music SLU · Internal music SLU, full data | SemER relative to E2E: -1.16 -> -1.01 | Hurt |
| add MLM to SLU + ASR pretraining (ablation) | AT-AT on full-data internal music SLU · Internal music SLU, full data | EM Accuracy relative to E2E: +1.39 -> +1.23 | Hurt |
| finetune the jointly pretrained model rather than train E2E from scratch | AT-AT · FluentSpeech | HypER: 8.3 -> 0.5 | Helped |
| use full AT-AT rather than only its pretrained audio encoder (ablation) | AT-AT on SNIPS Audio · SNIPS Audio close field | Full EM Accuracy: 53.90 -> 66.51 | Helped |
| use full AT-AT rather than only its pretrained audio encoder (ablation) | AT-AT on SNIPS Audio · SNIPS Audio far field | Full EM Accuracy: 38.92 -> 53.25 | Helped |
| add synthetic speech to text-only AT-AT finetuning | AT-AT zeroshot · Facebook TOP audio test | EM Accuracy: 51.54 -> 70.60 | Helped |
| use text-only zero-shot finetuning instead of synthetic-audio E2E training | AT-AT zeroshot · Facebook TOP audio test | EM Accuracy: 69.19 -> 51.54 | Hurt |
| add synthetic speech to text-only AT-AT finetuning | AT-AT zeroshot · Internal books SLU | SemER relative to real-audio E2E: +11.90 -> +3.31 | Helped |
Closest papers by meaning
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.86)
- Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models (NeurIPS 2025 · similarity 0.85)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.85)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.85)
- MATS: An Audio Language Model under Text-only Supervision (ICML 2025 · similarity 0.85)
- MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix (NeurIPS 2025 · similarity 0.85)
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.