asr-internal

ASR checkpoints, mirrored for the live-translation app

Streaming Zipformer2 transducers for sherpa-onnx, laid out exactly as the app expects them: kroko/<language>/{encoder,decoder,joiner}.onnx + tokens.txt. Nothing here was retrained, re-quantised or otherwise modified — these are the publishers' bytes, re-hosted so the app controls its own download source.

nemotron/ — the multilingual fallback

NVIDIA Nemotron 3.5 ASR streaming 0.6b, the 1120 ms chunk export, int8, from https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b via the sherpa-onnx export above. Cache-aware FastConformer-RNNT; takes a per-stream language prompt ("de-DE", "auto", …). Note the prompt table spells Arabic ar / ar-AR and has no ar-SA.

Both decoders are here. decoder.onnx is the float export the app loads today, because ONNX Runtime was reported to mis-fuse the int8 decoder's LSTM slice. decoder.int8.onnx is a quarter of the size and, measured on 47 FLEURS German clips through sherpa-onnx 1.13.6, transcribes at 11.20 WER against the float decoder's 11.47 — a paired bootstrap of [-0.09, +0.69], i.e. no difference. Whether it loads on a given Android runtime still has to be checked on the device; if it does, it saves 44.8 MB.

Attribution and licensing

Nemotron

Copyright NVIDIA. Governed by OpenMDW-1.1 — a copy is included as LICENSE-OpenMDW-1.1.txt, as that licence requires of anyone distributing any portion of the materials. It grants permission "to deal in the Model Materials without restriction", so this mirror is permitted; the obligation is to retain the agreement and the notices of origin, which is what this section and that file are.

Kroko

The three kroko/ checkpoints are Kroko ASR models by Banafo EOOD — https://kroko.ai, https://huggingface.co/Banafo/Kroko-ASR. Redistributed unmodified.

Turkish comes from the .data community generation, which Banafo describes as CC BY-SA ("use our CC-BY-SA licensed community models"). Unpacking the container reproduces the four files byte-for-byte; the SHA-256 of every one matches the copy already shipping. Banafo states no CC version and its LICENSE file is empty, so the version is taken as unspecified.

German and French come from the earlier loose-.onnx generation. Banafo's model card of 2025-09-25 said of that generation: "the new ones (with .data extension) are CC-BY-SA licensed, the .onnx are still non-commercial only." Those files were later removed from Banafo's own repository. Commercial use of kroko/de and kroko/fr therefore needs written permission from [email protected] — this mirror does not grant it and does not imply Banafo endorses anything here.

No warranty is given or implied. See the upstream repositories for the authoritative terms.