Long duration causing memory leaks

[eypklyc]

Hi,

I'm using Silero VAD to catch speeches in a long audio records. When using get_speech_timestamps function it's causing memory leak. Eventually the process consuming all memory and starting to write swap. Any advice to prevent this apart from cutting the samples multiple pieces?

silero-vad/utils_vad.py

Line 267 in 563106e

speech_prob = model(chunk, sampling_rate).item()

Just for information using reset states solving the problem but making predictions less accurate...

[snakers4]

Does this help?
#2 (comment)

long audio record

How long? How much memory is used for audios of which size?
You can consider trying the streaming interface if your audio is very long or "eternal".

Any advice to prevent this apart from cutting the samples multiple pieces?
Just for information using reset states solving the problem but making predictions less accurate ...

Cutting audio into 5-15m long segments should not harm the overall quality very much.
Resetting states forgets some context, but it should be negligible compared to the overall length of audio.

[eypklyc]

Sorry for the late reply. Yes that helped very well, disabling the gradients solved the leakage

torch.set_grad_enabled(False)

Thank you for the support..