Research map / Large language models / Instruction and fine-tuning
Unlocking Speech–Text Compositional Powers: Instruction-Following Speech Language Models without Instruction Tuning
Congrui Du, Yang Zhang, Kaizhi Qian, Shiyu Chang
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 |
|---|---|---|---|
| combine speech-adapted weights from base model with instruct weight difference instead of continuing pre-training the instruct model | SpeechCombine (Qwen3-8B) · OpenbookQA | accuracy: 78.46 -> 86.59 | Improved |
| weight combination instead of continuous pre-training plus SFT | SpeechCombine · EmphAssess emphasis detection | F1: 2.61 -> 60.84 | Improved |
| remove speech captions [cap] from pre-training data (ablation) | SpeechCombine · EmphAssess emphasis detection | F1: 60.84 -> 0.39 | Hurt |
| add [cap] speech captions to pre-training data (ablation) | SpeechCombine · GenEmotion-en | score: 27.18 -> 45.42 | Improved |
| add [text] transcription before each speech section (ablation) | SpeechCombine · EmphAssess emphasis detection | F1: 0.40 -> 60.84 | Improved |
| add [text] transcription before each speech section (ablation) | SpeechCombine · GenEmotion-en | score: 35.69 -> 45.42 | Improved |
| enable long thinking via answer template (ablation) | SpeechCombine · OpenbookQA | accuracy: 64.83 -> 86.59 | Improved |
| enable long thinking (ablation) | SpeechCombine · GenEmotion-en | score: 36.37 -> 45.42 | Improved |
| add instruction weight difference delta theta_inst instead of in-context examples (ablation) | SpeechCombine · GenEmotion-en | score: 15.34 -> 45.42 | Improved |
| in-context examples replace delta theta_inst vs no-thinking SpeechCombine (ablation) | SpeechCombine · OpenbookQA | accuracy: 64.83 -> 71.68 | Improved |
| increase combination weight lambda (ablation) | SpeechCombine · QA vs speech tasks | task performance | Mixed |
| force 'Based on the prosody' at start of thinking (ablation) | SpeechCombine · EmphAssess emphasis detection | answer correctness | Improved |
Closest papers by meaning
- InstructSpeech: Following Speech Editing Instructions via Large Language Models (ICML 2024 · similarity 0.88)
- Instruction Tuning With Loss Over Instructions (NeurIPS 2024 · similarity 0.87)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.87)
- Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs (NeurIPS 2026 · similarity 0.87)
- Joint Fine-tuning and Conversion of Pretrained Speech and Language Models towards Linear Complexity (ICLR 2025 · similarity 0.87)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.87)
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.
