I am working on audio dataset. The dataset contains train and test audio files. The train audio file contains crowd environment where only human are murmmering and chattering. The test data contains these chattering voices of human alongwith gunshot, car passing, and accidents.
My question is which features are good for extracting information from this kind of training data who has no specific sound, just the chattering voices?
Regards
Tariq, it is not clear from your question whether you are treating the training clips at a perceptual level as background noise or a masking signal or at a cognitive level as distracting and/or intelligible information. The relevance of audio features depends heavily on this and on the research question you are trying to answer.
For example, acoustic standards on background noise measurement require the type of noise to be classified as broadband, tonal or impulsive, as these are perceived differently by the listener. People are generally much more tolerant of broadband noise than of tonal or impulsive noise. The last two contain intelligible information, of sorts, while broadband noise does not. The auditory system tends to focus much more on intelligible sounds than on unintelligible ones as we are programmed to respond to auditory threats and opportunities, just as we are programmed to respond to threats and opportunities in other forms.
At a very basic level, it would probably be useful to know the spectrum of your clips. It may also be useful to know the amount of fluctuation in each band, which could be measured as L10 and L90 or something similar. If you are interested in masking qualities however, these measures are only very crude and you would need to analyse critical bands related to the spectrum of the sound being masked, along with the time envelopes of the masked and masking sounds in those bands. If there is intelligible speech in your training clips, you may need to assess the level of intelligibility via a random word recognition test, as the level of intelligibility may affect the degree of distraction presented.
If you can give us some more information on the aims of your experiment we may be able to help a little more.
Ian Dash Thanks for your detail explanation.
At training level, I have human chattering voices. And I am considering it as useful information not the background noise. Because my test data have human chattering voice, car noise, bicycle noises and I am considering these car noise, bicycle noise as anomaly.
My dataset do not contain any intelligent speech.
Below is the file attach of spectrogram of my audio file. The audio file has sampling frequency=44kHz, and duration 10s.
Thanks for the spectrogram, Tariq. It looks quite uniformly random, so I can't see any obvious features from it. There is a pale stripe across the bottom below the 64 Hz band which looks more constant than everything else - is there some low frequency hum in your sample?
And can you explain a bit more what you mean by "useful information"? What is the aim of your study?
- Badges
- Science topic
- Similar topics
- Filing