nemotron-3.5-asr-streaming-0.6b-8bit

Nemotron 3.5 ASR Streaming 0.6B — MLX 8-bit

This is Spoke's pinned distribution of the 8-bit MLX conversion of nvidia/nemotron-3.5-asr-streaming-0.6b. It is a 600M-parameter, cache-aware streaming FastConformer-RNNT ASR model with language-ID prompt conditioning, punctuation, and capitalization.

The model contains 8-bit quantized Linear and Embedding weights with group size 64. Convolution and normalization layers remain in bfloat16.

Provenance

The files are mirrored from mlx-community/nemotron-3.5-asr-streaming-0.6b-8bit at revision 7279359e4481b5e9e185a318bd618e429c6d86cd. The MLX conversion was made by @ARahim3 with mlx-audio.

Use

from mlx_audio.stt import load

model = load("spokedotso/nemotron-3.5-asr-streaming-0.6b-8bit")
print(model.generate("speech.wav").text)

The trained attention profiles are [56,3], [56,0], [56,6], and [56,13]. The [56,13] profile gives the lowest offline WER. Spoke uses the streaming profile selected by its local runtime.

License

The model materials are from NVIDIA Corporation and are distributed under the OpenMDW License Agreement 1.1. See LICENSE and NOTICE.