How to pass in 8k data for VAD

[RuABraun]

Hi,
I have a question about using 8k with the silero_vad model.

Is the right way to pass in 8k data to the model, or should I upsample the 8k data to 16k and then pass that to the model (or is the latter method wrong)?

Thank you in advance

[snakers4]

Hi,

Is the right way to pass in 8k data to the model, or should I upsample the 8k data to 16k and then pass that to the model (or is the latter method wrong)?

The jit model natively supports 8 kHz, just check the torch module docstring / and or see how the model is invoked in the utils.

As for the ONNX model, you have to upsample, because the ONNX model now cannot consume 8 kHz natively.

[RuABraun]

So the ONNX model will work with upsampled 8k data?

[snakers4]

The ONNX model just works only with 16k data

How its performance degrades if the data is upsampled we did not investigate, but I assume that it will be more pronounced that this - https://github.com/snakers4/silero-vad/wiki/Quality-Metrics#sample-rate-comparison

Also may depend on the upsampling method

[RuABraun]

Okay thank you

[zachllee]

@RuABraun i know this was a few years ago, how did the performance do when you upsampled?