Skip to main content

audio-wake-word

[UNTESTED] — v0.5 paper-correct

Builds clean on native_sim/native/64. Real PDM mic capture + Ethos-U55 dispatch + Vela-compiled model land in v0.6 AEN HiL.

Always-on "Hey Alp" keyword spotting on the E1M-AEN family's low-power AI subsystem. Targets the Cortex-M55 HE ("High Efficiency") core at ~50 MHz with the on-die Ethos-U55 NPU bursting convolutions on demand.

Source: examples/audio-wake-word/.

The AEN low-power story

The M55 HE core stays awake 24/7 at low clock running the keyword-spotter front-end; each 50 ms inference window dispatches conv layers to the Ethos-U55 NPU and parks the M55 HE in WFI for the rest of the window. Duty-cycle math: average system power < 1 mW for continuous listening.

board.yaml

schema_version: 2

som:
sku: E1M-AEN701

carrier:
name: E1M-EVK

cores:
m55_he:
os: zephyr
app: ./src
peripherals: [i2s]
inference: { backend: ethos_u, tensor_arena_kb: 128 }

diagnostics:
log_level: info

PDM mic capture rides Zephyr's audio subsystem (CONFIG_AUDIO_DMIC=y); the application uses the unified <alp/audio.h> surface.

See also

Questions about this page? Discuss in Community Forum