so far i started with downloaded audio cry data and convert to list of arrays. after i need to do feature extraction.
please suggest me some feature extraction techniques to mark important segments in the signal.
my goal is to predict a baby cry in real time
[i have one query for building the model we should do feature extraction and concatenation of all features to feed to the model. ]
how to proceed.
Do you mean "predict" or "detect" a baby cry?
If "predict", then there is no feature in the baby cry from which to predict that a baby is going to cry. All of your predictive training information is coming from the cry itself, which is too late, the cry is already underway. The predictive information must come when the baby is not crying, presumably in the relatively quiet moments before crying begins.
There is no need to extract features if you are going to use deep learning for the prediction/detection. The sole purpose of using deep learning is to learn the feature itself and classify, hence the input to the DNN is just raw audio signal. Of course there are some possibility to preprocess the data to make the learning quick and reliable. The learned features are known to work better than handcrafted features most of the time.
If you are going to use a classical neural network for the classification then you will need to extract some hand crafted features. In that case you can start searching from "Mel-frequency cepstral coefficients (MFCCs)" to many other types of feature which will be suitable for your problem domain.
Do you mean "predict" or "detect" a baby cry?
If "predict", then there is no feature in the baby cry from which to predict that a baby is going to cry. All of your predictive training information is coming from the cry itself, which is too late, the cry is already underway. The predictive information must come when the baby is not crying, presumably in the relatively quiet moments before crying begins.
Also, what would be a useful prediction? Is knowing 2-3 seconds in advance that a baby will cry really useful? If so, adding this rational to your question would greatly motivate people to help you more!
@Ronald Kessel, @Bruno Martin
According to the private discussion held with Prashanth (the writer of this question), It seems what he actually meant is not the prediction or detection but the interpretation. The use of term"predict" must be a mistake.
sorry@Ronald kessel,@Bruno Martin
the goal is to detect baby cry when starts crying
If the goal is to "detect baby cry when starts crying", the simple difference in volume from the base noise level should do the trick... If that is not enough, frequency analysis should provide a way to discriminate cry/non-cry far more understandably than an obscure deep learning model (which is indeed a black-box as far as understanding the problem goes).
Good luck!
It seems to me that a baby's cry is "designed" by evolution to be especially irritating in order to convey displeasure and motivate quick maternal response. I would suggest from my limited experience with babies that a rough first-order crude approximation of a baby cry is:
That should plausibly discriminate the cry from the short happy laugh and clicking vocalizations (both of short impulsive transient nature), and from humming and singing (both more narrowband and not as loud as a cry).
Note that in practice, you have to assess the tolerance one has for missed cries and for false alarms (detection of non-cries). Very low tolerance for such errors generally requires a very good detector/discriminator across a wide range of babies and operating conditions. It will not suffice to (A) merely demonstrate that cries can be detected/recognized. One must also (B) demonstrate that it is done with such a reliability that reliance on the detector is justified and that its risks/costs of errors (missed cries and false alarms) are acceptably low.
(A) will be easy to do. (B) will be much more difficult.
- Badges
- Science topic
- Similar topics
- Analytical Chemistry
- Extraction