| # Streamalign-SLM (R16) | |
| StreamAlign speech language model (Llama-3.2-1B backbone, R16 hier | |
| durfirst-durreg, text-KL recipe). Scores speech via the paired R16 RVQ | |
| teacher `Streamalign-R16`. | |
| - Original checkpoint: `R16_step180k_lossWtext5_kl05_210k/step_00185000.pt` | |
| - Speech tokenizer / teacher: `Streamalign-R16` | |
| - Reported metrics: SALMon 69.1 (4-task, no pitch), StoryCloze 72.1 (sSC 59.0 / tSC 85.1) | |