1Optimizing Streaming Speech Recognition on Arm® Edge Devices with ExecuTorchpytorch.orgHigh10·26 18:27Z
PyTorch published a case study running NVIDIA's multilingual Nemotron 3.5 streaming speech-recognition model on a Google Pixel 11 through ExecuTorch 1.3.1, profiling Arm SME2 kernel performance1. Enabling SME2 cut total processing time for a 24.9-second recording from 9.79 to 4.61 seconds at FP16 and from 5.20 to 3.24 seconds at QC8W, yielding real-time factors of 0.19 and 0.13 respectively. QC8W added only 0.21–0.31 percentage points of word error rate relative to FP32, while FP16 matched FP32 within 0.03 points at every chunk size tested.