Add NAS KWS model (trained using dynamic augmentation)

[alicangok]

The trained networks in this PR achieve the following accuracies for the current version of our KWS20 testing set¹

• v2: 92.0%
• v3: 90.8%
• NAS: 93.8%

Footnotes

1. This testing set contains all 11005 test examples from the original Google Speech Commands dataset without any "background/others" class balancing, as opposed to many benchmark settings. For our dataset, the "others" class makes up 25.7% of the testing set, making it more challenging than the aforementioned benchmarks which feature a much smaller "others" class (10%), as well as an easily identifiable "silence" class (10%). ↩

[alicangok]

PR changed to draft mode, awaiting training completion for v2 & v3 models with dynamic augmentation.

[ermanok]

Is it possible to rename the KWS20_v3 and KWS20_nas models as kws_light and kws? We can also remove the v1 & v2 models...

[alicangok]

The demo, the KAT example, and arch fields in the model files and trained checkpoints) refers to the model as "v3" in multiple places, so we would need make the changes to all at the same time.

I would suggest holding off on this until for a few weeks until we finalize the frequency-domain model, then decide on long-term names.

[ermanok]

That makes sense. Thanks...

[aniktash]

Looks good

[alicangok]

Final changes:

• updated v2 & v3 checkpoints. They are now trained using dynamic augmentation, achieving better accuracies.
• updated the NAS checkpoint: It's now trained with samples not expanded to the full range of [-128, 127]. This method was scrapped as it significantly reduced robustness.
• modified copyright texts for files generated by ai8xize.py (namely cnn.c, cnn.h, main.c, Makefile, and softmax.c), following @sihyung-maxim's comments in feat(Examples): Add Action Recognition Demo for MAX78000 and Known Answer Test for MAX78002 msdk#851 .

Please see the edited PR description accuracy details.

The PR is now marked as "Ready".

[MaximGorkem]

Looks good