Murmur translation models (CTranslate2, int8)
OPUS-MT tc-big checkpoints converted to CTranslate2 and quantised to int8,
for on-device use in Murmur. CPU only:
the app's GPU is held by speech recognition.
Measured on FLORES+ devtest ru->en, 1012 rows, beam 1, one thread, Apple M1 Max:
| model | chrF++ | BLEU | p50 | on disk |
|---|---|---|---|---|
| bergamot student (realtime tier) | 56.79 | 29.55 | 12 ms | 22.5 MB |
this, ru-en |
60.40 | 35.50 | 378 ms | 253 MB |
int8 quantisation costs -0.06 chrF++ against float32 - within noise - for a 3.9x reduction in size.
en-ru is a group checkpoint (eng->zle) and carries target_tag.txt; the
tag token selects between Russian, Ukrainian and Belarusian and must be
prepended to the source. Without it the model guesses, which happens to be
right for Russian and wrong for the others.
Attribution
Source models by the Helsinki-NLP OPUS-MT / Tatoeba-MT project, licensed CC BY 4.0. Conversion to CTranslate2 int8 is the only change; weights are otherwise unmodified.
- Upstream: https://github.com/Helsinki-NLP/Tatoeba-Challenge
- Converter:
ct2-opus-mt-converter --quantization int8(CTranslate2 4.8.2)
European language catalog
The app catalog now covers all 24 official EU languages, plus Russian and Ukrainian. English-centered OPUS routes provide both directions for every language; Russian–Finnish also has direct models. Other combinations use two OPUS passes through English. Models are downloaded on demand.
Languages: Bulgarian (bg), Croatian (hr), Czech (cs), Danish (da), Dutch (nl), English (en), Estonian (et), Finnish (fi), French (fr), German (de), Greek (el), Hungarian (hu), Irish (ga), Italian (it), Latvian (lv), Lithuanian (lt), Maltese (mt), Polish (pl), Portuguese (pt), Romanian (ro), Slovak (sk), Slovenian (sl), Spanish (es), Swedish (sv), Russian (ru), Ukrainian (uk).
This is a mixture of OPUS-MT tc-big and standard OPUS-MT checkpoints, not a claim that every direction is tc-big or equally accurate. Each new folder carries provenance.json with the exact upstream model, source hash/revision, original model-card license, converter version, tokenizers, target-language tag and a smoke translation. Helsinki-NLP / OPUS-MT provides the source models. The Irish and selected South Slavic copies use pinned Helsinki-NLP Hugging Face checkpoints where original archive delivery or conversion is unavailable.
For native callers that supply raw SentencePiece pieces, the converted Transformers models explicitly append the source EOS through config.json, matching MarianTokenizer preprocessing. Preserve all files including target_tag.txt when present. en-pt selects European Portuguese with >>por<<. Translation support is separate from speech-recognition support; this repository contains translation models.