input shape to the LSTM net when doing inference for VAD tasks

Hi, I am following this article to train a LSTM network for VAD tasks: https://www.mathworks.com/help/deeplearning/ug/voice-activity-detection-in-noise-using-deep-learning.html

My question is, when testing a trained LSTM network, as in the article did, the input data is not shaped as the training input as (#frames, #time_steps, #features), does this mean, when doing inference, the trained LSTM network will take each frame as a input independetly, and classify if this frame is noise or voice, so basically there is no hidden states used when doing inference, am I right?

Thank you in advance!

Answer 1

[Brian Hemmat]

I did not look at the dimensions you're discussing, but I can say that you are correct that the "streaming" code in the example classifies chunks independently. Note that it is calling classify and not classifyAndUpdateState.

Stay tuned for the R2023a release, where we have updated the example to maintain state (should be coming in the next few weeks).