Analyzing stereo files takes twice the time of mono files

[fegue]

The code clearly loads audio files as mono in audio.py Do you have any idea why BirdNET still takes twice the time?

This happens on a RaspberryPi 3A+ Memory usage is similar when analyzing mono files. The function librosa.load(..., mono = True) executes almost instantaneously

[kahst]

Hmmm, weird, do you have a sample file that I can use to try to replicate this behavior?

[fegue]

Well ok, so its not always double the time but it always takes significantly longer. The audio files were recorded in stereo and then converted to mono using ffmpeg (10min & Sample rate 48 kHz). 1 File (10 min): Mono 70s ; Stereo 90s 2 Files (à 10 min): Mono 110s; Stereo 260s 3 Files: Mono 190s; Stereo 460s

CPU_THREADS = 1 TFLITE_THREADS = 4

Example files are here https://oekofor.ddns.net/index.php/s/7mMXFWGNTeCTYkp

[kahst]

I wasn't able to reproduce the issue, which machine and model did you use to run the test?

This is the log from my workstation PC using only one CPU thread:

Analyzing example/mono/SN99_2022-04-07T084531+0000.wav
Finished example/mono/SN99_2022-04-07T084531+0000.wav in 18.88 seconds
Analyzing example/mono/SN99_2022-04-07T084531+0000_2.wav
Finished example/mono/SN99_2022-04-07T084531+0000_2.wav in 18.30 seconds
Analyzing example/mono/SN99_2022-04-07T084531+0000_2_mono.wav
Finished example/mono/SN99_2022-04-07T084531+0000_2_mono.wav in 17.58 seconds
Analyzing example/mono/SN99_2022-04-07T084531+0000_3.wav
Finished example/mono/SN99_2022-04-07T084531+0000_3.wav in 18.23 seconds
Analyzing example/mono/SN99_2022-04-07T084531+0000_3_mono.wav
Finished example/mono/SN99_2022-04-07T084531+0000_3_mono.wav in 17.43 seconds
Analyzing example/mono/SN99_2022-04-07T084531+0000_mono.wav
Finished example/mono/SN99_2022-04-07T084531+0000_mono.wav in 17.43 seconds

[fegue]

I ran the tests on a Raspberry Pi 3A+, 512mb Ram and a 2Gb Swapfile which fills up quite a bit. The model version is 1.4 and it runs on tflite-runtime without the full tensorflow installation.

[kahst]

Did you find a solution or workaround for that issue? Haven't been able to reproduce the behavior so far.

[mzakharo]

@fegue -> try placing the audio files into /tmp ramdisk, and analyzing from there, trying to rule out the SD card reads. Also, if you are using client/server.py, then server saves temporary file to local folder, again, you can edit the file to move it to /tmp.