Kakao published weights for its LM-SPT speech tokenizer to Hugging Face on 2 September at 05:32 UTC, with the initial release commit at 05:49 UTC and the paired code repository pushed at 05:10 UTC. The model is a 12.5 Hz neural speech tokenizer over 24 kHz audio with 8 codebooks — one semantic codebook of 16,384 entries plus seven acoustic residual-VQ codebooks of 4,096, at a 1920× downsample.
What the common framing gets wrong
“Kakao open-sources speech tokenizer” implies usable. The code is Apache-2.0; the weights are CC BY-NC 4.0 — no commercial use. And Kakao did not pick that licence. The model card says the weights carry “the same license as the Emilia dataset used for training”. The restriction flows from the data, through the checkpoint, to anyone who downloads it.
The trap for anyone checking quickly
Go upstream to verify and the picture gets worse. The Emilia repository on Hugging Face is tagged cc-by-4.0, but its own terms split the corpus: Emilia is CC-BY-NC, and only Emilia-YODAS is CC-BY. The machine-readable licence tag on the dataset is more permissive than the dataset's actual text, and Kakao's card is the stricter, correct reading. A team doing licence diligence by reading metadata rather than terms would clear this and be wrong.
Two smaller corrections
The weights ship without model code — the lmspt package must be installed separately. And this is not a new research result: the paper is an arXiv preprint from June 2025, now published in IEEE TASLP 2026, vol. 34, pp. 3714–3727. The code repository's first commit predates the weights by twelve days. Only the weights are new. Card languages are English, Chinese, Korean, German, French and Japanese; the auxiliary semantic decoder runs at 16 kHz, not 24.
Why this generalises
Almost every open speech-LLM stack trains on the same handful of scraped corpora. If the corpus is non-commercial, the checkpoint is too, regardless of how permissive the code repository looks — and the platform's licence field will not warn you.