BarathwajAnandan/nemotron-3.5-asr-streaming320-int8-CoreML

Nemotron 3.5 ASR Streaming 0.6B - CoreML int8

This repository contains an Apple Silicon CoreML port of NVIDIA's nvidia/nemotron-3.5-asr-streaming-0.6b.

For the original model description, supported languages, architecture, training data, evaluation notes, license terms, and intended use, refer to the upstream NVIDIA model card:

https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b

License terms are inherited from the upstream NVIDIA model card.

CoreML Port

• —Format: CoreML .mlpackage components
• —Target runtime: Apple Silicon
• —Streaming configuration: 320 ms
• —Quantization: int8 linear encoder quantization
• —Approximate package size: 668 MB
• —Included files: metadata.json, preprocessor.mlpackage, encoder.mlpackage, decoder.mlpackage, joint.mlpackage, joint_decision.mlpackage, tokenizer.model

The CoreML package preserves the prompt-conditioned multilingual setup from the upstream model, including manual language prompts and auto language detection where supported by the runtime.

FluidVoice

This package is prepared for FluidVoice's Nemotron Speech 3.5 Ultra Fast Low Latency engine. FluidVoice downloads and loads the repository as a set of CoreML components rather than as the original NeMo checkpoint.

Attribution

The underlying model is developed by NVIDIA. This repository is a CoreML conversion/packaging of the upstream Nemotron 3.5 ASR streaming model for local Apple Silicon inference.