The Stack Overflow podcast is back! Listen to an interview with our new CEO.
10

The way MFCC's are always used is by feeding them into the classifier. This can be done on a frame-by-frame basis (12x1 vector), or by concatenating (12xN) - same as a spectrogram. Thus for DTW, you must perform the classification by calculating the distance between 12D vectors. It's a Dynamic Time Warping, so the difference must be calculated between ...


5

If you are looking for a good explanation of some of the methods used in projects like GRAIL look here: Back to the Future of Hand Writing Recognition


3

Surprised? DARPA (GRAIL project?, et.al.) supported handwriting recognition research circa 3 decades earlier, and on mainframes less powerful than a PalmPilot 1000. Hawkins talks a bit about implementing Graffiti in a Computer History Museum oral history.


3

p and q - order of the moments, it is the analogue of the mechanical moments (i.e. of force or inertia). Image moments are very basic properties of image, invariant to rotation, could be used as simple descriptor. Example, Moment of inertia Ix= I2 = mass *x_coordinate^2. Image moment Ix2 = intensity * *x_coordinate^2. http://en.wikipedia.org/wiki/Moment_(...


2

Because wo want to get the centroid of the image(a block/patch) by the intensity. m00:p = q = 0,sum the intensity matrix. m10:p =1,q = 0,sum of the x-direction. m01:p = 0,q = 1,sum of the y-direction. (m10/m00,m01/m00) is the centroid.


2

Have a good read of various features that can be extracted from popular audio analysis libraries like librosa: https://librosa.github.io/librosa/feature.html Fundamental frequency (F0) can definitely be extracted. This itself a whole subfield, so google and previous SO questions are your best starting point after reading through what a pitch tracker looks ...


2

Try one of these algorithms for feature selection: BLogReg CFS Chi Square FCBF Fisher Score Gini Index Information Gain Kruskal-Wallis mRMR Relief-F SBMLR T-test SPEC


2

Your understanding of the convolution process is correct. However, note that in the convolution the kernel is first mirrored before the dot product. Though, for a symmetric kernel this does not matter. Understanding the convolution in your way, of finding specific features, is a similar interpretation as that of a Matched filter. Furthermore, using the ...


2

Well, I don't know whether it'll actually help you – you just said it would! Now, in any case, using an algorithm to extract features from a signal that mimics or resembles human perception should inherently give those feature vectors a higher "mathematical" resemblance when a human would find the original signals similar, too. This is obviously what you ...


2

It seems most of the data is in Graffiti (Palm OS) - Wikipedia. It seems the technology is related to Unistrokes technology (U.S. Patent 5,596,656, granted in 1997). Since the model relies on single stroke and using symbols that are exceptionally well separated from each other graphically I would be assume parameterizing the stroke would be a great way to ...


2

This project Speech Signal Processing Toolkit (SPTK) provides several features you are looking for. Here is a good wrapper around it https://github.com/r9y9/pysptk This wrapper is using a slightly different version of SPTK, but installation is straightforward. pip install pysptk 1) I would use kaldi or another speech recognition package on phones output ...


2

I think a high-level explanation is the right thing to give here: For recognition of a producing process (in this case: speech) from a signal (in this case: audio), you need to find a mapping from signal to a set of features. The closer (correlated) these features are to the parameters of the producing process (here: voicing utterances), the more ...


2

Well the state of the art performance on such tasks is achieved by deep neural networks, and especially, the convolutional ones (CNNs) set you free of extracting hand crafted features. The network learns those features as well as the weights. This way you avoid the hustle or the rots o'ruck. For sequential data input, such as audio signals, recurrent neural ...


1

Assuming you have epochs/segments of data. For this kind of signals, it is a safe approach to extract features using wavelet representations. Using FFT might work as well, but I dont know how problematic would be the stationarity assumption in this kind of applications. Besides, FFT estimates for this kind of signals is sometimes very noisy. If you insist ...


1

VTLN is pretty old technology. There is a fast linear VTLN or basis fMLLR which is about the same thing. These days everyone is doing neural networks though, they provide more advanced speaker adaptation, at the same time they could be pretty efficient with quantization and proper weight pruning. For example you can check DOMAIN AND SPEAKER ADAPTATION FOR ...


1

TL;DR the x-axis is time and the y-axis is coefficients which means they have no units. Let's go through the process of manufacturing the MFCC features while emphasizing measure units tracking. We start from a sampled audio signal $s(n)$ with amplitude units on the y-axis and time units on the x-axis. First step is to frame the audio signal. An audio ...


1

How this 36*1 came and how we calculated it? HOG is an algorithm which: works on a portion of the image, called "detection window"; divides the "detection window" in a certain number of cells; associates an histogram (of oriented gradients) to each cell. Each histogram has N orientation bins. WLOG, here, we'll consider N=9); "Normalization step" (to make ...


1

In the case of switched analog POTS phone lines, the audio response is shaped by the limited frequency response of the carbon microphone and the speaker in the handset.


1

A possible way to approach the problem would be to extract the maximum values of the discrete circular convolution of each pre-normalized sequence with an NXN matrix formed by the elements of an orthonormal base of the Nx1 sequence space. The output of this step would be a Nx1 vector that would contain a measure of the similitude of each sequence with the ...


1

if you can get into the IEEE Explore Library Pull up D. J. Burr, "Designing a Handwriting Reader," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-5, no. 5, pp. 554-559, Sept. 1983. and then look at the citations, which include both the Palm and Xerox Patents. Xerox claimed that Palm infringed so you might pick up some clues.


1

The openSMILE audio feature extraction toolkit may be able to provide the functionality you desire, where the input is a .wav file and the output extracted audio features. See: http://audeering.com/technology/opensmile/ openSMILE provides a command line executable that is coupled with a configuration file that defines the features to be extracted. The ...


1

I have a solution for your problem other than wavelet transforms that could be worth trying. Here are the steps : Apply AGC & normalise the time series data. This will ensure that even in different lighting conditions etc and you will have relative similar peaks and troughs amplitudes. Choose a wavelet that best represents your individual event. Eg : ...


1

A wavelet analysis or wavelet transform will give you scale(frequency) vs time vs amplitude. I am not sure how that will help you with your problem. Here is an algorithm for wavelet transform : Choose your wavelet : Haar , Molet etc ..say $ \psi (n) $ Choose Scale : the scale should be a multiple of smallest time interval dt. Calculate the FFT of wavelet ...


1

As long as you train and evaluate the same features then you will be perfectly fine with this delay. In practice first two frames of the feature vector for both $\Delta$ and $\Delta\Delta$'s are zeros. So in fact the coefficients are delayed by two frames. These can be calculated according to the formulas: $$\Delta[t] = c[t]-c[t-2]$$ $$\Delta\Delta[t] = c[...


1

Looking at these references, this is what I would understand: The Haralick descriptors only create a single value for each GLCM (as you also thought). The Rampun paper creates 32 images, describing 32 different features for each input image. The crucial point is this sentence: On the other hand, we used a small window size of 5 x 5 throughout the process ...


1

I dont think there is a common framework for defining all those operators. Each of them were developed by different people, at several moments of the history. They are being integrated day by day onto a common discipline called DSP. As clarification, an operator (like a moment or any other) and the "redundancy" of a signal are very different and unrelated ...


Only top voted, non community-wiki answers of a minimum length are eligible