This article explores five different methods for computing the Short-Time Fourier Transform (STFT) on a Qualcomm System on a Chip (SoC), specifically the Snapdragon SM8650. It details implementations using the CPU, a dedicated digital signal processor (cDSP), and the Neural Processing Unit (NPU), measuring accuracy, latency, and power consumption for each. The findings indicate that the optimal method depends on the specific requirements of the downstream audio-AI pipeline, rather than a single universally best approach. AI
IMPACT Provides data for optimizing audio AI pipelines on edge devices by comparing compute paths for STFT.
RANK_REASON The item details a technical comparison of different computational paths for a specific signal processing algorithm on hardware, including methodology and measured results. [lever_c_demoted from research: ic=1 ai=0.7]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
