The feature set for my project is pitch values in the form
features = (p1,n1,p2,n2,p3,n3.....pt,nt)
where p1 is the pitch value and n1 is the duration (number of times that pitch value is repeated)
I am currently using FDTW to compare two signals after remaking the inputs in the form and normalizing them by their mean and standard deviation.
(p1,p1,p1,p1..... ,,, p2,p2,p2,p2 .... and so on)
I have sampled each the hum and original tune to compare with and used the algorithm to determine similarity. Which makes the comparisons very slow. Currently the time complexity for a single hum to song comparison is
O(n) = n3log(n
FDTW has given me an accuracy of 60-90% based on the correctness of hummed tune.
I am unclear about how to use Machine Learning for my project. Is it even required and will it improve the time complexity or accuracy?
You can use machine learning algorithms if you have large amount of samples whose classes are known.(supervised learning)
It is not necessary to implementing other machine learning algorithms directly since DTW is usually used in certain application. Perhaps combination of them is a promising approach as
http://pages.bangor.ac.uk/~mas00a/papers/jrlkkdd07.pdf
Your features seem to be quite simple. Do you have enough data for each query?
Basically for an ML system (supervised training) you need a couple of hummed queries per song to train a classifier.
ANN and SVM might not be as useful, because they are designed for static data, but in your case the temporal variations would be difficult to model. They can be used in combination with a sequence model, such as HMM, but this can get complicated...
One possibility (provided that you have enough data) is to train a recurrent neural network (state-of-the-art model is called LSTM: long short-term memory RNN)
Have a look at this paper: http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=1223846&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D1223846
If you are familiar with python, there is quite a handy library called keras. There is an example of IMDB sentiment classifier that can be adapted for your problem: https://github.com/fchollet/keras/blob/master/examples/imdb_lstm.py
This example of using LSTM in keras can also be useful http://stats.stackexchange.com/questions/192014/how-to-implement-a-lstm-based-classifier-to-classify-speech-files-using-keras
It depends on what you mean by machine learning.
In 2003-5, Marc Bohlen made UVM
http://www.realtechsupport.org/repository/uwm.html
Its an artwork in which the subject whistles a tune and the system imrpovises (whistles) a related tune back. Such projects are not uncommon in the history of computer art and computer music. Sadly, few computer scientists know anything about the field.Those who do have often plundered artists work for commercial applications. One could write a book on the subject.
A possible approach is to use a Nearest Neighbor algorithm because it looks like what you need is to compute a similarity value between songs in a database and the song being hummed (the FDTW you are using is a possible measure for that purpose). Obviously you need to represent in the same way the songs in the database. NN are what is known as lazy learners because do not learn a general model based on the training examples but in your case I do not think you need to learn a model. There is an extensive literature in NN algorithms.
- Badges
- Science topic