Not entirely sure if this Stack or a ML-related stack would be a fit for this question, but I focus on the signal processing here on this question, so I decided this one

I'm developing a speech classification system. Imagine, that you have a dataset of happy and sad speech, and you want to classify them. I'm looking for a minimum set of standard pre-processing steps that I have to do to compare happy and sad speech and build a machine learning-based classifier out if it. This is something, that I believe is called normalisation in the literature, it is definitely expected to be done, and I find it confusing.

To contrast, in image-based problems, I find that most "normalisation" can be safely neglected. In the case of audio, I'm not sure if it can be neglected, and what is a proper way to do normalisation. I tried to read more about this, but I find even the definition of normalisation ambigous in audio, i.e:

  1. Some people refer to normalisation as preprocessing the two classes of data so they are comparable. Sometimes this is straightforward, but in speech classification problems, I certainly don't think it is.
  2. Normalisation has a different meaning when we refer to audio normalisation. Wikipedia link
  3. Normalisation as standardisation (MeanStd scaling) or normalisation (MinMax scaling) is also a predominant definition.

There are many techniques that fit the normalisation bill then. I'm not aiming to give a complete list, but here is a few

  1. Channel normalisation.
  2. Loudness normalisation
  3. Voice activity detection (cutting out silence)
  4. Speech pseudonymisation
  5. Cepstral mean and variance normalisation

What is the minimum I'm expected to do when I'm building a classifier for happy/sad speech detection?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.