Question

I’m currently implementing a beat detection algorithm with python and numpy/scipy.
I basically need to read a .wav file and process it. Here is the code:

sampling_rate, wave_data = scipy.io.wavfile.read(argv[1])

wave_data is a 1-D numpy array with about 441 000 elements (10 seconds of sound with 44.1 kHz sampling rate). Now, I need to do some basic math on every two elements in this array. This is how I do it now:

wave_data = [sampling_rate * (wave_data[i+1] - wave_data[i]) 
             for i in xrange(len(wave_data)-1)]

This opreation takes too much time (noticeable without profiling). I need to map the array pairwise “in-place”, without creating a new python list. I know there is numpy.vectorize, but I don’t know how can I do the mapping pairwise (map every two elements of the array).

Answer 1

Either of the following will do it:

wave_date = sampling_rate * np.diff(wave_data)

or

wave_date = sampling_rate * (wave_data[1:] - wave_data[:-1])

For example:

In [7]: sampling_rate = 2

In [8]: wave_data = np.array([1, 3, 5, 2, 8, 10])

In [9]: sampling_rate * (wave_data[1:] - wave_data[:-1])
Out[9]: array([ 4,  4, -6, 12,  4])

As far as performance is concerned, both of these approaches are about 500x faster than the list comprehension:

In [16]: wave_data = np.array([1., 3, 5, 2, 8, 10, 5, 2, 4, 7] * 44100)

In [17]: %timeit sampling_rate * np.diff(wave_data)
100 loops, best of 3: 2.2 ms per loop

In [18]: %timeit sampling_rate * (wave_data[1:] - wave_data[:-1])
100 loops, best of 3: 2.15 ms per loop

In [19]: %timeit [sampling_rate * (wave_data[i+1] - wave_data[i]) for i in xrange(len(wave_data)-1)]
1 loops, best of 3: 970 ms per loop