Curation of a dataset of audio files in different formats

Question

I have a large dataset of audio files (around 80k files of around 45 minutes long each) that are related to my work. Some of them are in format ".m4a" and others in format ".mp3". I am not a computer scientist, but I understand that these are two different formats.

I will use the audios later on for a variety of analyses. My main concern for now is to compile a homogenous dataset. It is bugging me that the audio files are encoded in two different formats. I could convert one into another, but I am afraid I will lose quality on the converted files.

What would be a good strategy for me? Should I just keep the files as they are, and only convert local versions for analysis?

Answer 1

Keep the files the way they are for long term storage.

For experimenting, it might be beneficial to deal with just one format.

Depending on what resources you have available issues may be:

• supported formats
• disk space/speed
• decompression speed
• random access

Depending on the requirements of your analyses, you might want to reduce the quality of your tracks to gain speed advantages.

Supported Formats

If the software you use for analysis supports both MP3 and M4A, you're set. If not, you'll need to convert using tools like ffmpeg or sox.

Disk Space/Speed

If you were to convert your files to uncompressed WAV, this will eat up a lot of disk space. If you don't have this space, WAV is not a good format. Instead convert to another lossless format like FLAC or Apple Lossless. You can do so without loosing any quality.

Besides disk space, you need to pay attention to disk speed. For some deep learning tasks you will have to provide your network with lots and lots of different samples, which you need to read from disk. With WAV being roughly 10 times as large on disk, disk I/O may become a bottleneck, so minimizing file size on disk may become desirable.

Decompression Speed

Using a compressed format like FLAC or MP3 comes at a (small) price. You will have to decompress your audio files on-the-fly before analysis.

Random Access

If you need to quickly access audio starting at a certain timestamp, say at 10.455s, formats like WAV have a distinct advantage, as they allow random access of audio very easily. This is usually not possible (as quickly) with compressed formats.

So what format should you choose for your experiments?

It all depends on your requirements.

Do you need quick random access to the audio? Do you really need stereo quality at 44,100 Hz? You can reduce the problem size a lot, by converting to mono 22,050 Hz. If that's still good enough for your purposes, go for it and save yourself a lot of trouble (but keep your original files around!).

Also, to get more technical, if you are going to analyze your data with something like Python and you know you need spectrograms, there's nothing wrong with preprocessing your MP3/M4A files with a library like librosa and simply store numpy arrays, if that's what you ultimately need.

Answer 2

If you have enough storage, covert all these to some common format like 320kbps MP3 or WAV.

This will save some issue with analysis pipeline (and move these issues to ETL pipeline).

Issues are :

1. Compatibility issues with Audio processing libraries (Such as a specific library might only accept raw audio, another might accept mp3)
2. Compatibility issues with Metadata tagging and extraction (Some libraries will not read tags from wav, others wont read them from m4a)
3. Artifacts introduced by format conversion at runtime (Offline conversion can afford to spend more CPU cycles)

This will need ~10TB or 40TB of storage.