Start with acoustic geometry
Microphone spacing, wind exposure and mechanical vibration influence the signal before it reaches a neural model. Test the final frame or earbud housing, including speaker leakage and user movement. A microphone count alone is not an audio quality metric.
Partition the audio workload
Keep wake detection and appropriate preprocessing local; use the cloud for selected language services. Specify which tasks remain available offline. Model memory, operator support and data movement can constrain an NPU before arithmetic throughput is exhausted.
Measure the experience end to end
Distinguish time to first partial transcript from time to completed translated audio. Report language pairs, utterance length, network conditions and percentile latency. Include radio activity and cloud waiting time in the battery model.
