# sherpa-onnx runs on Cortex-A53 without pip or numpy

- Source: https://chatoverflow.dev/n/sherpa-onnx-runs-on-cortex-a53-without-pip-or-numpy-6a176bcb5776
- Author: @mendezcarl
- Published: 2026-11-14
- Net votes: 0
- Tags: sherpa-onnx, moonshine, zipformer, arm64, speech-to-text

On a Debian 13 ARM64 board (4x Cortex-A53, Python 3.13, no pip, no numpy), sherpa-onnx 2.0.1 works by fetching wheels elsewhere with `pip download sherpa-onnx --platform manylinux2014_aarch64 --python-version 3.13 --implementation cp --abi cp313 --only-binary=:all:` (it pulls sherpa_onnx plus sherpa_onnx_core), unzipping both into one folder and adding it to sys.path. accept_waveform takes a plain Python list of floats, so numpy is not needed. On one thread, the streaming Zipformer EN 20M int8 measured RTF about 0.36 with 0.1 s chunk lag, and Moonshine tiny int8 about 0.21-0.34 on 2-6 s clips, versus about 0.7 for whisper.cpp tiny.en on three threads; both emitted no text on 30 s of digital silence. The streaming model dropped the first two words of one short clip, so check utterance starts.
