Lorqa/nemotron-3.5-asr-streaming-multilingual-0.6b-320ms-coreml

Lorqa transcription — Fast (320 ms)

Byte-identical mirror of aufklarer/Nemotron-3.5-ASR-Streaming-0.6B-CoreML-INT8 at commit 447095fe87b480b5e6a15367f135303d479de8ac. Based on NVIDIA Nemotron 3.5 ASR streaming 0.6B. Core ML INT8 encoder, FP16 decoder and joint; 16 kHz mono PCM. All weights and language/tokenizer/configuration files are unchanged.

Download: 642,196,943 bytes of model/runtime data (642 MB decimal / 612 MiB). Runtime memory is higher than download size. 320 ms is chunk duration, not guaranteed end-to-end latency. Chinese accuracy needs evaluation on the intended recordings.

Languages and prompt slots: see languages.json. Auto detection and explicit BCP-47 language hints are supported by the companion runtime. Do not use this bundle with FluidAudio's 2240 ms manager; the graph and cache layouts differ.

Runtime reference: soniqo/speech-swift (Apache-2.0). Model weights retain OpenMDW 1.1, with upstream attribution retained. This repository is a distribution mirror, not a retraining or accuracy claim.

The Balanced (2240 ms) model remains in Lorqa/nemotron-3.5-asr-streaming-multilingual-0.6b-coreml, under multilingual/2240ms.