Classification problem

Issue with Neural networks - Storage issue with TFlite model, 256kb RAM in the nano BLE but the model itself is 6800kb. Post-training quantization does not help (model size at 1700kb).

Number of parameters in the model = (1088 + 8) + (4*4000) = 16,652

Memory for 16,652 parameters is 70kb and post-training quantization the size is 17kb

Moreover, the inference time is super fast

Metrics for Keyword Spotting

1. Accuracy: Reduce false positives and false negatives using a FAR vs FRR curve. Find an operating point based on the environment and fine-tune sensitivity.

1. Latency: Should be minimal
2. Memory usage: Pruning and quantization.

1. Quality of experience: Account for diversity in terms of environment and user.

Streaming audio

Single shot vs. Streaming audio.