1
$\begingroup$

We want to provide an ML model that recognises diseases from the voice. The feature extraction is based on a proprietary algorithm. Normally we have always used wav files. We keep asking ourselves if we could also use something like mp3? Since we probably won't be able to collect the data we collect again and we can't foresee what kind of information a further development of our proprietary algorithm will need at some point, we are concerned that something like mp3 would lead to too much loss of information. What do you think?

Examples of diseases: ADHD, depression, diabetes, Parkinson's disease, suicide.

Examples of categories from which the algorithm generates features: melody, rhythm, pauses, speaking speed, dynamics...

I am less interested in a concrete answer, as I am aware that it is not possible to give a concrete answer due to the open questions on certain details. It's more about sharing experiences and exchanging ideas. It could be, for example, that someone has experienced in the field of machine learning that the areas that are cut away in MP3, for example, usually have little or no effect on ML algorithms. Which is just one example. I haven't really found anything on this topic on the internet.

Even if this doesn't have so much to do with an "either/or" regarding the decision in favour of a format, one article, for example, recommends the following: "Data augmentation is a well known practice in machine learning. We can simply treat the coded version of our audio data as an augmentation of the data like we would by adding traffic noise or simulating the echoes of a room on clean speech. For example, augmenting your training data with just a few different Opus quality levels will improve the classification of all Opus test samples." ...and that is very interesting.

$\endgroup$
2
  • $\begingroup$ what are we supposed to think? we don't know your proprietary algorith, hence we can't tell you how sensitive it is to compression losses! We don't know how diverse your data set is, so that we don't know how much stress you can put on the MP3 encoder to get artifacts. We also don't know what exactly you're learning, and whether that would at all care about such losses. $\endgroup$ Commented Jan 18 at 11:54
  • $\begingroup$ I am less interested in a concrete answer, as I am aware that it is not possible to give a concrete answer due to the open questions on certain details. It's more about sharing experiences and exchanging ideas. It could be, for example, that someone has experienced in the field of machine learning that the areas that are cut away in MP3, for example, usually have little or no effect on ML algorithms. Which is just one example. I haven't really found anything on this topic on the internet. $\endgroup$
    – Tütü
    Commented Jan 18 at 12:09

2 Answers 2

3
$\begingroup$

MP3 (and AAC, Ogg, Opus, Dolby, DTS, etc) are PERCEPTUAL encoders. They are based on specific properties of the human auditory system, the primary one being auditory masking (see for example https://en.wikipedia.org/wiki/Auditory_masking)

areas that are cut away in MP3

That's the wrong way to think about it. The only thing that MP3 cuts aways is any content above 16 kHz. Perceptual coders ADD stuff: specifically they add quantization noise. They just shape it so that the noise is under the masking threshold and and hence inaudible.

something like mp3

That' too broad a question. There a probably a dozen or so perceptual encoders and they come in lots of different flavors and each of those covers a wide range of bit rates.

What do you think?

Given the lack of specific information, I can't even guess. What I think is that you will just have to do the research work for your specific field of application. That shouldn't be all that hard

  1. Familiarize with the inner workings of perceptual audio encoders.
  2. Relate how they work to the specific features your model is trying to learn. Chances are the closer these features are to human perception, the less impact the encoders will have.
  3. Pick the most promising encoder (which may be also the most practical in terms of license, open source, commercial entanglements, etc.)
  4. Pick a bunch of relevant use cases.
  5. Encode your training and test corpus with the chosen encoder at a few different bit rates.
  6. Train with wave and the different bit rates, compare performance and whatever other metrics you can get from the training.
  7. Analyze the results, adjust as necessary and repeat until you have a satisfactory answer.
$\endgroup$
1
  • $\begingroup$ Thank you very much. That gives some hints and directions. $\endgroup$
    – Tütü
    Commented Jan 18 at 12:49
0
$\begingroup$

You provide very few details about the "disease" and the "proprietary algorithm" which leaves few space for comments.

But I can provide some ideas:

  • if the disease affects voice acoustics (such as Parkinson's) and you are using acoustic features, then it can be risky to use lossy compressed records.
  • if the disease does has few or no impact on acoustics (e.g. Alzheimer's) then there are no risks on using lossy formats.
$\endgroup$
1
  • $\begingroup$ Thank you. I have added an update $\endgroup$
    – Tütü
    Commented Jan 18 at 14:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.