KleidiAI

Headlines · 1

1. 5d
   Nemotron 3.5 Streaming ASR Benchmarked on Pixel 11

   PyTorch profiled multilingual Nemotron 3.5 streaming ASR through ExecuTorch on a Google Pixel 11, testing chunk sizes, FP16 and QC8W precision, and Arm SME2 acceleration. With SME2 enabled, processing a 24.9-second recording fell from 9.79 to 4.61 seconds for FP16 and from 5.20 to 3.24 seconds for QC8W. The study also shows the accuracy-latency tradeoff: longer audio chunks improved word error rate but increased estimated time to first token.

   PyTorch Blog · 🔥 0

Experience and discussion from the community

Share my KleidiAI experienceAsk about KleidiAI