Streamalign (R16)
StreamAlign RVQ teacher / speech tokenizer (R16, C512). Tokenizes speech into text-aligned RVQ units (16 quantizers, codebook size 512) consumed by Streamalign-SLM (R16).
- Original checkpoint:
260507_RVQ_R16_C512_nodistill_codelr10x_from_nofsq_e13/epoch_16.pt - Paired SLM:
Streamalign-SLM-R16 - Reported metrics (paired): SALMon 69.1, StoryCloze 72.1