Research map / Image generation / Autoregressive and latent generation
Latent Speech-Text Transformer
Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srini Iyer, Duc Le
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 |
|---|---|---|---|
| add speech patching (curriculum) to the speech-text LM | Base SpeechLLM · HellaSwag S->S (compute-controlled) | accuracy: 39.0 -> 45.5 | Improved |
| add speech patching (curriculum) to the speech-text LM | Base SpeechLLM · HellaSwag T->T (compute-controlled) | accuracy: 47.0 -> 52.2 | Improved |
| add speech patching (curriculum) with data fixed | Base SpeechLLM · HellaSwag (data-controlled) | accuracy: S->S 40.2 -> 45.5; T->T 49.6 -> 52.2 | Improved |
| replace patching with BPE on speech tokens | Base SpeechLLM · HellaSwag S->S (compute-controlled) | accuracy: 39.0 -> 38.0 | Hurt |
| use curriculum rather than aligned-only patching (ablation) | LST · HellaSwag S->S (compute-controlled) | accuracy: 42.7 -> 45.5 | Improved |
| use aligned patching instead of static at matched size 6 (ablation) | LST · StoryCloze S->S | accuracy: 58.7 -> 60.3 | Improved |
| use aligned (sil merged) instead of static at matched size 9 (ablation) | LST · HellaSwag S->S | accuracy: 37.2 -> 38.5 | Improved |
| add curriculum patching vs static size 4 (ablation) | LST · HellaSwag S->S | accuracy: 40.5 -> 41.3 | Improved |
| scale to 1.8B with LST (compute-optimal) | Baseline · HellaSwag S->S | accuracy: 35.3 -> 39.0 | Improved |
| scale to 420M with LST (compute-optimal) | Baseline · HellaSwag S->S | accuracy: 28.4 -> 29.2 | Improved |
| add LST at 7B fixed tokens | Baseline · HellaSwag S->S | accuracy: 42.0 -> 44.2 | Improved |
| add LST for ASR adaptation | Baseline · LibriSpeech clean ASR, 1k iters | WER %: 140 -> 6.8 | Improved |
Closest papers by meaning
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.90)
- DiTTo-TTS: Diffusion Transformers for Scalable Text-to-Speech without Domain-Specific Factors (ICLR 2025 · similarity 0.89)
- DELTA-TTS: Adapting Autoregressive Model into a Diffusion Language Model for Text-to-Speech (NeurIPS 2026 · similarity 0.88)
- TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling (NeurIPS 2025 · similarity 0.88)
- NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.88)
- MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer (ICLR 2025 · similarity 0.88)
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.
