plox99/llm-translate-models

LLM-Translate model mirror

Byte-for-byte mirror of every downloadable model used by LLM-Translate, an offline Android translator (RU / EN / NE). This repository exists as a fallback source: if the upstream hosts below are unreachable, the same files can be downloaded from here without any conversion.

Folder names match the app's model IDs from models_catalog.json. Every file is identical to its upstream original (same name, same bytes — see CHECKSUMS.sha256 for SHA-256 of each file).

For stable downloads pin a commit SHA instead of main: https://huggingface.co/plox99/llm-translate-models/resolve/<commit-sha>/<folder>/<file>

Contents and licenses

Folder	Upstream source	Model license	Notes
firefox-bergamot-en-ru, firefox-bergamot-ru-en	Mozilla Firefox Translations GCS (storage.googleapis.com/moz-fx-translations-data-…)	MPL 2.0 — model files, per the mozilla/translations README	Redistribution allowed; keep MPL notice when modifying model files
indictrans2-indic-en-int8, indictrans2-en-indic-int8	hari31416/indictrans2-*-ONNX-int8 (ONNX export + INT8 quant of ai4bharat/indictrans2-*-dist-200M)	MIT (preserved from AI4Bharat)
stt-ru-gigaam-v3-e2e-ctc	k2-fsa/sherpa-onnx asr-models release, export of ai-sage/GigaAM-v3	MIT
stt-en-parakeet-110m	k2-fsa/sherpa-onnx asr-models release, export of nvidia/parakeet-tdt_ctc-110m	CC-BY-4.0	Attribution: “Parakeet TDT-CTC 110M © NVIDIA / Suno, CC-BY-4.0”
stt-ne-indicconformer	parismitaglobalsolutions/indicconformer-sherpa-onnx, export of AI4Bharat IndicConformer	MIT (AI4Bharat)
stt-whisper-small-multilingual	same repo — re-export of OpenAI Whisper small (multilingual)	MIT / Apache-2.0 (permissive)	Export/quantization by parismitaglobalsolutions
tts-ru-piper-ruslan, tts-en-piper-ryan, tts-ne-piper-google	k2-fsa/sherpa-onnx tts-models release, voices from rhasspy/piper-voices	see caveat below

Piper voices license caveat

Piper (MIT on code) deliberately does not license voice weights; users must judge from the upstream training history (see rhasspy/piper#271):

• ru_RU ruslan — training dataset (ruslan-corpus) licensed CC BY-NC-SA 4.0
• en_US ryan — training dataset (RyanSpeech) licensed CC BY-NC-SA 4.0
• ne_NP google — training dataset (National Speech Corpus, openslr.org/43) licensed CC BY-SA 4.0
• all three medium-quality voices are fine-tuned from the lessac voice, trained on the Blizzard Challenge 2013 corpus whose terms restrict use to research purposes

Mirror is uploaded unmodified for app fallback use in a free, non-commercial open-source app. Evaluate the above before any commercial use.

Provenance

This mirror was produced by tools/hf_mirror.py in the LLM-Translate repository; file set follows app/src/main/assets/models_catalog.json.