Research map / Large language models / Language, safety and interpretability
Textually Pretrained Speech Language Models
Michael Hassid, Tal Remez, Tu Anh Nguyen, Itai Gat, Alexis Conneau, Felix Kreuk, Jade Copet, Alexandre Défossez, Gabriel Synnaeve, Emmanuel Dupoux, Roy Schwartz, Yossi Adi
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 |
|---|---|---|---|
| Warm-initialize the speech model from pretrained OPT-350M rather than random initialization | SpeechLM · Zero-shot speech modeling, 100/200/500 tokens at 50Hz and 500 tokens at 25Hz | sWUGGY: 68.91 -> 71.30; 69.85 -> 72.92; 66.65 -> 70.69; 79.44 -> 81.42 | Improved |
| Use a lower speech-token frequency, 25Hz instead of 50Hz (ablation) | OPT-350M SpeechLM · Zero-shot speech modeling | sWUGGY: 70.69 -> 81.42 | Improved |
| Increase the speech tokenizer vocabulary from 100 to 200 tokens at 50Hz (ablation) | OPT-350M SpeechLM · Zero-shot speech modeling | sWUGGY: 71.30 -> 72.92 | Improved |
| Increase the speech tokenizer vocabulary from 200 to 500 tokens at 50Hz (ablation) | OPT-350M SpeechLM · Zero-shot speech modeling | sWUGGY: 72.92 -> 70.69 | Hurt |
| Warm-initialize using pretrained textual OPT rather than random initialization | SpeechLM · Speech generation | text-PPL: 88.17 -> 74.95; 117.17 -> 92.69; 67.68 -> 47.8 | Improved |
| Increase model size from 1.3B to 13B parameters (ablation) | TWIST · sStoryCloze | accuracy: 52.4 -> 55.4 | Improved |
| Increase model size from 1.3B to 13B parameters (ablation) | TWIST · tStoryCloze | accuracy: 70.6 -> 76.4 | Improved |
| Increase model size from 1.3B to 7B parameters (ablation) | TWIST · LibriSpeech-clean | MMOS: 3.51($\pm$0.07) -> 3.79($\pm$0.06) | Improved |
| Initialize speech model from image-pretrained ImageGPT instead of random initialization (ablation) | SpeechLM · ImageGPT initialization, 200 tokens, 50Hz | PPL: 5.22 -> 8.21 | Hurt |
| Initialize speech model from image-pretrained ImageGPT instead of random initialization (ablation) | SpeechLM · ImageGPT initialization, 200 tokens, 50Hz | sWUGGY: 72.22 -> 54.26 | Hurt |
| Initialize speech model from image-pretrained ImageGPT instead of random initialization (ablation) | SpeechLM · ImageGPT initialization, 500 tokens, 25Hz | sBLIMP: 56.39 -> 51.84 | Hurt |
| Use pretrained textual BLOOM or Pythia weights for warm initialization | SpeechLM · Zero-shot speech modeling | sBLIMP: BLOOM: 53.03 -> 55.79; Pythia: 53.07 -> 56.00 | Improved |
Closest papers by meaning
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.90)
- SyllableLM: Learning Coarse Semantic Units for Speech Language Models (ICLR 2025 · similarity 0.88)
- EmergentTTS-Eval: Evaluating TTS Models on Complex Prosodic, Expressiveness, and Linguistic Challenges Using Model-as-a-Judge (NeurIPS 2025 · similarity 0.87)
- Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning (ICLR 2024 · similarity 0.86)
- DataComp-LM: In search of the next generation of training sets for language models (NeurIPS 2024 · similarity 0.86)
- Paralinguistics-Aware Speech-Empowered Large Language Models for Natural Conversation (NeurIPS 2024 · similarity 0.86)
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.
