vrstt-models

vrSTT models

Model weights for vrSTT — local speech-to-text for VR and streaming. Everything vrSTT runs is here so the app can fetch the one model a user chose, verify it, and never touch the network again.

This repository redistributes third-party models. It trains nothing and modifies nothing except file format. Licences differ per directory and are stated below; the upstream terms travel with the weights.

Layout

manifest.json                       # what the fetcher reads first
whisper/ggml-<name>.bin             # whisper.cpp ggml format, one file each
translation/<name>/                 # CTranslate2 format, a directory each

manifest.json carries a sha256 and a byte count for every file. The fetcher verifies both before handing a path to whisper.cpp or CTranslate2 — a truncated ggml file otherwise fails at load with an error about magic bytes, which sends you looking at the wrong thing.

What is here

Speech recognition — whisper.cpp ggml conversions, byte-identical to ggerganov/whisper.cpp. Mirrored rather than linked so the bytes vrSTT downloads are pinned and hashed.

tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large-v1, large-v2, large-v3, large-v3-turbo

vrSTT defaults to small.en for commits and base.en for live previews — the split is deliberate, previews are allowed to be wrong and are replaced within a second, so they should not spend the larger model.

Translation — NLLB-200 converted to CTranslate2 with ct2-transformers-converter --quantization float16, from facebook/nllb-200-*.

nllb-200-distilled-600M (default), nllb-200-distilled-1.3B, nllb-200-1.3B

float16 rather than int8 because CTranslate2 can cast down at load time but not up — one artifact serves both, and converting to int8 here would cap quality for everyone. 600M is the default because vrSTT's stated minimum-VRAM target is 8 GB, and a 1.3B model beside the ASR stages and a game does not fit comfortably there.

Licences — read this before redistributing further

NLLB-200 is non-commercial. Redistribution is allowed with attribution and the licence must travel with it — which is what this section is — but neither Meta's weights nor anything derived from them may be used commercially. If vrSTT ever becomes a paid product, the translation models have to be replaced rather than relicensed.

NLLB-200 also carries Meta's own caveats, which apply unchanged here: it is research work, not certified for production translation, it was trained on Wikimedia-domain text and does not target long-form or domain-specific input, and translation quality varies widely by language pair. Mistranslations can carry real consequences and vrSTT displays them live on a stream — the include_original setting exists so the source text stays visible beside the translation for exactly this reason, and it is on by default.

Citation

Whisper — Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision, OpenAI, 2022.

NLLB-200 — NLLB Team et al., No Language Left Behind: Scaling Human-Centered Machine Translation, Meta AI, 2022.