0
$\begingroup$

I have a dataset of quarter second 16 bit mono PCM audio clips, each clip labelled as speech or nonspeech. I am trying to estimate the quality of the labelling by randomly sampling a few of them and listening to the clips. How do I detect if the samples are labelled correctly by hearing them?

$\endgroup$
11
  • 1
    $\begingroup$ Hello! I'm sorry, I'm not quite sure what you're asking here. You listen to them – if you hear someone speaking, then someone is speaking, if I'm understanding you correctly. So, what is then the question? $\endgroup$ – Marcus Müller Jun 3 at 10:14
  • 1
    $\begingroup$ @MarcusMüller I think the "quarter second" here is the tricky part. $\endgroup$ – Hilmar Jun 3 at 12:14
  • 2
    $\begingroup$ ah! So, the question is kind of "I want to use my human perception to verify presence of speech in a snippet, but the snippet is too short. How can I preprocess it to increase my own accuracy?", if I understand correctly? $\endgroup$ – Marcus Müller Jun 3 at 12:19
  • 2
    $\begingroup$ I would redo your experiment entirely as otherwise I would claim your results are invalid given it was a judgement call if speech was present or not. Instead take known longer samples of “speech” that meet whatever criteria you place on it being “speech” and known longer samples of “no speech” and from each of those pull out .25 second samples and use that to test your algorithm $\endgroup$ – Dan Boschen Jun 3 at 18:49
  • 1
    $\begingroup$ It would be the thorough verification so I would go as far to suggest if you can’t get the originals recreate new ones. $\endgroup$ – Dan Boschen Jun 3 at 22:21
0
$\begingroup$

That's really NOT a signal processing question, so I will interpret it as "what signal properties can be used to perceptually identify speech in very short recordings?".

A quarter second is quite short. It accommodates on average maybe 2 syllables or 3-4 phonemes. And that's only on average: one of the most common sounds in speech is silence which can easily last for more than a quarter second.

You first need to define precisely what do you mean by "speech". Let's say you take a recording of an audio book and chop it up into 1/4 second long snippets. Are all of those snippets supposed to be speech (since they came from a real of speech recording)? Even the silence ones ? How about the ones that contain 0.2s of silence and 0.05s of a partial phoneme? How about the ones that are half silence half speech ?

Next I would take a look at the non-speech samples. If it's not speech then what is it? Noise, environmental sounds, traffic, music or movies, silence, typical residential background, all of the above, etc? Please note that many of these can also contain human voice. If the snippet is from a song with someone actively singing: is that speech or not?

Once you have a crisp definition of speech and non-speech, you can manually create a representative training set that represents accurately your definitions. And then you need to listen to them and see if you can train yourself to distinguish them reliably.

Depending on how exactly you define speech and non-speech, that may be rather easier or pretty much impossible. For example: You can't distinguish "silence in speech" from "other silence".

$\endgroup$
1
  • $\begingroup$ Please elaborate on the 'easy' end - assume silence is treated as non-speech. For example, is there anything at all that indicates that a quarter second clip contains human speech? Is there anything that indicates that a quarter second clip that is not entirely silent does not contain human speech - i.e., is only noise? $\endgroup$ – Anand Jun 3 at 17:09

Your Answer

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

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