unreadable noise signals in the fma_small subset

[ChangLee0903]

Dear MultiSV members,

I just noticed that there are some unused noise signals that cannot be loaded by torchaudio and librosa, such as 'fma_small/108/108925', 'fma_small/099/099134', and 'fma_small/133/133297'. I think there might be some broken files in the noise dataset?

best regards,
Chi-Chang Lee

[Lamomal]

Dear Chi-Chang,

Thank you for bringing this up. Indeed, the noise files you mentioned do not contain any audio. It is a known issue of the FMA dataset. For reference, please see FMA wiki and a related issue. In line with your comment, MultiSV users might encounter warnings during data preparation related to these files (which can be safely ignored). Even if the FMA small dataset is downloaded correctly (as indicated by the create_training_data.sh script, which compares checksums), the erroneous files will cause the following warnings during the conversion to wav:
WARNING: conversion failed for: <output_dir>/noises_training/fma_small/108/108925.mp3
WARNING: conversion failed for: <output_dir>/noises_training/fma_small/099/099134.mp3
WARNING: conversion failed for: <output_dir>/noises_training/fma_small/133/133297.mp3

Since MultiSV does not use these three files, they do not pose an issue for the corpus.

Best,
Ladislav