i want to detect baby cry in real-time and i am using windows system. I had extracted the mfcc features of baby cry and trained a neural network model.
i got model accuracy 99%. This model i export to raspberry pi for real-time testing and got good results. But the problem is lot of false positives.
the model is detecting every loud sounds it shows lot of false positives.
my aim is to detect only baby cry.
i also used pitch features but no changes in reducing false positives.
please help me to overcome this problem.
Dear Prashanth,
Did you try to include into training period of your classifier those false sounds? It will be deciding how to change the sensitivity and specificity of your classifier after including noise or background sounds.
I do recommend to create a database of sounds and work above it.
You might try to change your classification method when you are seeing a decline of sensitivity and specificity after the proposed strategy.
There are existing tricks how to work with biosignals and the volume of the papers dealing with it is increasing.
When interested, do bit hesitate to ask for more.
Jiri
Adding to Jiří Kroc , also look at your feature extraction process. Make sure you extract features from both time and frequency domain. In frequency domain, use windowing to extract features at different frequency bands. This might help in further reducing false positives.
Perform PCA to your data and look at the scatter for PC1 and PC2. You should see some groups in the scatter. If you see one big scatter then, you might have to consider collecting more data.
How many data you have?
Do you control overfitting on training the model?
How do you choose your hyper-parameters?
The questions above are simple ways to say you that on neural networks if you have a lot of data and control overfitting, the model becomes resilient to false positives.
Your hight model accuracy and the loud sound detection behavior looks like an overfitting effect to me.
You can use by Combining the red flag approach with process mining.
Article Reducing false positives in fraud detection: Combining the r...
my data include 108 baby cry, 90 non-baby cry(baby laugh, baby noise, baby silence). i had trained KNN model with MFCC 13 features got model accuracy 90% after cross validation and grid search. when i test this model in real-time i got many false positives, like (its is giving false predictions).
the good thing is 100% accurate prediction for baby cry, but for non-baby cry the model predicting as baby cry.
There is one trick that comes immediately to my mind when you are speaking with troubles related to false positives. It would be great to enlarge testing database for mixed sounds. Create new items and add them to the databases to all sounds of a child in the following way. Make mixed sounds by contaminating pure sounds with sounds of noise, ratio, speak, music, noisy doors, barking dog, you see the point.
When you compare spectra of crying children with those contaminated, you must find a combination of features in the frequency domain that separate cry from the rest.
When you will be very initiative and clever you can add self-learning mode and allow users to classify alarms for false positives and negatives and hence allow the device to improve and adapt to given home environments.
sir could you please suggest me a research paper on classification task using ML to built robust system.
thanks in advance@ Jiří Kroc
I searched for papers, one interesting about MFCC follows
https://ieeexplore.ieee.org/document/1415167
It is directly related to your problem with noise. Personally, I never solved such problems. My field is to deal with biosignals.
Another avenue to your goal. You are using Python, hence, it would be to try more methods simultaneously on the same data and use ensemble learning composed from different methods. This gives almost always more robust data but is more computationally expensive. Using this method you can select the best classifying methods and decrease their number to a reasonable minimum.
Is this procedure clear?
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training