Research map / Large language models / Language, safety and interpretability
Timers and Such: A Practical Benchmark for Spoken Language Understanding with Numbers
Loren Lugosch, Piyush Papreja, Mirco Ravanelli, Abdelwahab HEBA, Titouan Parcollet
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 |
|---|---|---|---|
| Use the Timers and Such transcript language model instead of the LibriSpeech language model | ASR-based SLU models · test-real | accuracy: 55.5% ± 3.4% -> 64.0% ± 3.3% | Improved |
| Train the decoupled model with the Timers and Such transcript language model instead of the LibriSpeech language model | Decoupled · test-real | accuracy: 24.1% ± 1.1% -> 43.5% ± 2.0% | Improved |
| Add synthetic training data to real training data | Multistage (TAS LM) · test-synth | accuracy: 51.5% ± 2.9% -> 85.4% ± 0.2% | Improved |
| Train only on synthetic data rather than real plus synthetic data | Direct · test-real | accuracy: 77.5% ± 1.6% -> 68.0% ± 5.5% | Hurt |
| Train directly from audio instead of using an ASR transcript as the intermediate input | SLU models trained on train-real · test-real | accuracy: 64.0% ± 3.3% -> 81.6% ± 5.4% | Improved |
| Train directly from audio instead of using an ASR transcript as the intermediate input | SLU models trained on train-real + train-synth · test-synth | accuracy: 85.4% ± 0.2% -> 96.7% ± 0.3% | Improved |
| Train the direct model on synthetic speech, which permits learning synthesizer-specific pronunciation patterns | Direct · test-synth | accuracy: 70.0% ± 5.7% -> 96.7% ± 0.3% | Improved |
| Use synthetic-only rather than real-only training data | Direct · test-real | accuracy: 81.6% ± 5.4% -> 68.0% ± 5.5% | Hurt |
Closest papers by meaning
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.84)
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.84)
- BIGOS V2 Benchmark for Polish ASR: Curated Datasets and Tools for Reproducible Evaluation (NeurIPS 2024 · similarity 0.84)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.83)
- Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks (ICLR 2025 · similarity 0.83)
- WorldSpeech: A Multilingual Speech Corpus from Around the World (NeurIPS 2026 · similarity 0.83)
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.
