Lorqa/nemotron-3.5-asr-streaming-multilingual-0.6b-coreml
Nemotron 3.5 ASR Streaming Multilingual 0.6B — CoreML (2240 ms tier)
CoreML build of NVIDIA's Nemotron 3.5 streaming ASR model, used by Lorqa for on-device live captions and audio transcripts on macOS and iOS. Runs on the Apple Neural Engine through the FluidAudio SDK.
- 40+ languages with automatic language detection (no per-language download)
- Cache-aware streaming encoder: state is carried across chunks, so decoding is incremental rather than re-running on a growing buffer
- Fully on-device — audio never leaves the machine
What's in here
Only the `multilingual/2240 ms` tier — 22 files, ~634 MB.
multilingual/2240ms/
├── encoder.mlmodelc/ # 539 MB
├── decoder.mlmodelc/
├── decoder_joint.mlmodelc/
├── joint.mlmodelc/
├── preprocessor.mlmodelc/
├── metadata.json # prompt_dictionary, lang tag token ids
└── tokenizer.jsonUpstream ships 8 bundles (latin / multilingual × 560/1120/2240/4480 ms). The other seven are deliberately not mirrored here — see below.
Why the 2240 ms tier
The four tiers are not quality levels or versions. They are streaming chunk lengths: how much audio goes into one forward pass (chunk_mel_frames × 10 ms hop). All four share the same att_context_size of [42, 13].
Measured on the multilingual bundle (RTFx = audio-seconds per wall-second, higher is better; parenthesised figure is WER, or CER for zh/ja — lower is better):
2240 ms is the upstream-recommended default and the only tier that is strong on both axes for CJK. 560 ms falls off the trained attention tiling, so it is worse on accuracy and throughput — it is not a lower-latency trade, it is dominated. Figures are from the upstream model card, not re-measured here.
Provenance
NVIDIA nemotron-3.5-asr-streaming-0.6b → CoreML export by FluidInference → this mirror. Weights are unmodified; this repo exists so Lorqa pins a stable source and users in mainland China get a fast one. The same content is mirrored on HuggingFace and ModelScope.
Licensed under OpenMDW 1.1, following upstream.