The front end determines model input
Echo cancellation requires a useful reference and sufficient timing alignment. Beamforming relies on array geometry, calibration and the sound field. Evaluate double-talk, wind and nearby playback before attributing recognition errors to the language model.
Interpret TOPS in context
The supplied 0.8 TOPS INT8 target describes arithmetic throughput, not application throughput. Operator support, tensor shape, memory traffic and scheduling affect realized performance. Report a named model and measured latency, together with accuracy after quantization.
Account for every active state
An always-listening workload is different from active streaming and playback. Microphone bias, DSP, NPU, memory and radio power belong in a system budget. The 8.2 mW preliminary figure cannot be treated as total device power across all modes.
