Related to MFCCs - Getting better results by choosing wrong sampling rate.

I have an audio recorded at 48 kHz.

I read this audio file by selecting a wrong sampling rate of 22050 Hz and calculated MFCCs by taking wrong sampling rate 44100 Hz.

Though both the sampling rates are incorrect but I am getting better classification results compare to the true sampling rate i.e. 48 kHz.

I want to know theoretically what is going on through which I am getting better result by implicating both the mistakes.

Answer 1

[Drew]

What is the task? What type of classification is being done? Are you performing training and testing of your model, or just testing an existing model?

You mention that one classification result is "better" than the other. Is the difference statistically significant by a mcnemar test (https://www.mathworks.com/help/stats/testcholdout.html)? Or, is the difference just normal statistical variation? This is the first question to answer.

[D S Parihar]

Hi Drew,

The task is to classify nornal cough vs abnormal cough signal using MFCCs and Logistic regression. For this, We are performing training to build a model and using k-fold cross validation predicting the output.

Again I repeat the question.

Case 1: When I read the audio file using sr = 48 kHz and calculating MFCCs at the same sampling rate.

Case 2: I read the audio file at sr = 22050 Hz and calculating MFFCs at 44100 Hz. Though both the sampling rate are wrong. This happened by mistake while analysing classification performance.

Other parameters are, mel filter = 128, nfft = 2048, win_size = 2048, win = hamming

Now, I am getting AUC of 86% in the case 2 while it is 80% in the case 1.

I want to understand the retionale of what exactly going on in both the cases.