Feasibility of this research for real time deployment

[saikirandingu]

Dear Author,
first of all I would like to say that this work is a very good solution for low data resource kws. Thank you for sharing the e2e training recipe of this novel idea.
I am interested to know if this is a reliable solution for real time Speech commands recognition on low power embedded devices.
Another question I would like to ask is:
Could we use the Embedding extractor model which was trained for 10-way 15-shot experiment for deploying into a 6-way keyword spotting pipeline?

Thanks,
Saikiran

[ArchitParnami]

Hello @saikirandingu ,

Thank you for you comments.

I am interested to know if this is a reliable solution for real time Speech commands recognition on low power embedded devices.

Currently, we are working on deploying a follow-up work on this research to test its effectiveness in real time. So please stay tuned. But, as a comparison to encoders mentioned in the paper (those which have been used in low power devices eg. TC-ResNet) our proposed encoder has less number of parameters and better Few-Shot Classification performance. So we are hopeful on the results.

Encoder	# Parameters
TDResNet7 (Ours)	100,832
C64 (Prototypical Networks)	111,936
TCResNet8	127,200
cnn-trad-fpool3	555,552

Could we use the Embedding extractor model which was trained for 10-way 15-shot experiment for deploying into a 6-way keyword spotting pipeline?

I believe yes. Snell et. al 2017 suggested that harder training (in terms of number of ways) results in better classification performance at test time. They trained on 20-way 5-shot and tested the encoder on 5-way 5-shot and observed that it has better classification performance when compared to 5-way training (section 3.1 & 3.2 of the paper).

[saikirandingu]

Fantastic Archit! I am very thankful to you for sharing your answers. Keep up your good work!

…

On Tue, 4 May 2021, 5:45 am Archit Parnami, ***@***.***> wrote: Hello @saikirandingu <https://github.com/saikirandingu> , Thank you for you comments. I am interested to know if this is a reliable solution for real time Speech commands recognition on low power embedded devices. Currently, we are working on deploying a follow-up work on this research to test its effectiveness in real time. So please stay tuned. But, as a comparison to encoders mentioned in the paper (those which have been used in low power devices eg. TC-ResNet <https://github.com/hyperconnect/TC-ResNet>) our proposed encoder has less number of parameters and better Few-Shot Classification performance. So we are hopeful on the results. Encoder # Parameters TDResNet7 (Ours) 100,832 C64 (Prototypical Networks) 111,936 TCResNet8 127,200 cnn-trad-fpool3 555,552 Could we use the Embedding extractor model which was trained for 10-way 15-shot experiment for deploying into a 6-way keyword spotting pipeline? I believe yes. Snell et. al 2017 <https://arxiv.org/pdf/1703.05175.pdf> suggested that harder training (in terms of number of ways) results in better classification performance at test time. They trained on 20-way 5-shot and tested the encoder on 5-way 5-shot and observed that it has better classification performance when compared to 5-way training (section 3.1 & 3.2 of the paper). — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEX2GN2M43ALQVVOMFKFGOLTL44ANANCNFSM4352NPUA> .

[saikirandingu]

Hi Archit,
would you like to suggest/recommend some strategy for preparing the Unknown class data, in order to make the system more robust to the real world false positive triggers like some unrelated speech uttered, or some hesitation sounds or unrelated words or phrases?