OpenVoiceOS/mms-1b-all-onnx

MMS-1B-all ONNX — one shared base, one small pack for each of 1198 languages

ONNX export of Meta's `facebook/mms-1b-all` for onnx-asr. A single 1B-parameter base graph serves every language. Each language adds only a small adapter pack.

This release holds every adapter in the checkpoint: 1198 languages.

How it works

MMS is one wav2vec2 base with a small attention adapter and a CTC head per language. A merged export costs 3.6 GB for every language. Here the adapter layers and the CTC head are graph inputs, not baked-in weights:

One session holds 3.75 GB and each extra pack adds about 7 MB. Feeding the adapter tensors costs about 0.02 ms per call, so switching language is free.

model.onnx            shared base graph
model.onnx_data       base weights (fp32)
model.int8.onnx       shared base graph, dynamic int8 weights
model.int8.onnx_data  base weights (int8)
config.json           model_type, language list, aliases
manifest.json         per-language verification record
adapters/<iso>.npz    adapter layers + CTC head + true vocab size
vocabs/<iso>.txt      character vocabulary for that language

Usage

The whole repository is about 15 GB, so onnx-asr downloads one language at a time. That needs the feat/adapter-lazy-download branch of the TigreGotico fork:

pip install "onnx-asr[cpu] @ git+https://github.com/TigreGotico/onnx-asr@feat/adapter-lazy-download"

import onnx_asr

model = onnx_asr.load_model("OpenVoiceOS/mms-1b-all-onnx")
print(len(model.asr.languages))

print(model.recognize("clip.wav", language="swh"))
print(model.recognize("clip.wav", language="sw-KE"))

The first load fetches the base graph, config.json and the pack of default_language (swh). Every other pack is fetched the first time its language is used and then cached. To pay that cost up front:

model.asr.preload("yor", "ces", "khm")

Add quantization="int8" to load_model for the int8 base.

language takes the pack name (swh), a BCP-47 tag whose primary subtag maps to a pack (sw-KE, cs), or one of the 133 aliases in config.json. Pack names with a script or dialect suffix (azj-script_latin, cak-dialect_central) must be given in full.

Verification

manifest.json records what was checked for each language.

97 languages were checked on a real recording; 29 of those were also compared against the native transformers model. The remaining 1101 languages have no public recording, so they have the graph and vocabulary check only. Every pack was also checked structurally at export time: tensor names, tensor shapes and vocabulary size against the vocabulary file.

A character error rate in the table is against one FLEURS sentence. It says the model produces the right script and roughly the right words; it is not a benchmark.

Languages

<details> <summary>All 1198 languages</summary>