Support: Memory Efficiency for long recordings

[Evan8456]

Issue Description

Had trouble performing feature extraction locally on audio recordings longer than a minute.

Environment Details

Ram: 16 Gb
Os: Ubuntu 20.04.3 LTS
Python: 3.10.9

Attempted Solutions

No response

Reproduction Steps

No response

Additional Notes

No response

[github-actions]

👋 Welcome to Senselab!

Thank you for your interest and contribution. Senselab is a comprehensive Python package designed to process behavioral data, including voice and speech patterns, with a focus on reproducibility and robust methodologies. Your issue will be reviewed soon. Stay tuned!

[fabiocat93]

Hi @Evan8456, thank you for reporting the issue! Could you please provide more details on the problem you're experiencing? Are you encountering any specific error messages? Any unusual behavior during the process? Does it just stick forever?

Additionally, could you share the steps or code to help us reproduce the issue on our end? This will allow us to understand the problem better and assist you more effectively

[Evan8456]

Hi, you can see the context of my issue here. The bug seems to occur when I attempt to run extract_features_from_audios on audio recordings that are 30+ seconds long.

[ibevers]

We've agreed it should be a lazy object.

[ibevers]

Laziness possibly optional

[ibevers]

Is there a version of torchaudio load that is lazy? Possibly, similar to: librosa, cv2

[fabiocat93]

Is there a version of torchaudio load that is lazy? Possibly, similar to: librosa, cv2

I have checked and unfortunately librosa doesn't seem to have lazy load 👎

[900miles]

I have checked and unfortunately librosa doesn't seem to have lazy load 👎

librosa.stream seems to be fairly similar -- it returns a generator of audio chunks rather than loading the whole audio into memory at once.

The library I was thinking of during the meeting was soundfile (which it looks like librosa uses underneath). The SoundFile class opens an audio file, then you can use various methods to seek around in that audio file and read/write a given amount of frames.

[fabiocat93]

I have checked and unfortunately librosa doesn't seem to have lazy load 👎

librosa.stream seems to be fairly similar -- it returns a generator of audio chunks rather than loading the whole audio into memory at once.

The library I was thinking of during the meeting was soundfile (which it looks like librosa uses underneath). The SoundFile class opens an audio file, then you can use various methods to seek around in that audio file and read/write a given amount of frames.

maybe StreamReader is the corresponding tool in torchaudio: https://pytorch.org/audio/main/tutorials/streamreader_basic_tutorial.html