Program to find occurrences of one audio clip inside another audio clip.

versus.py

# Program to find occurrences of one audio clip inside another audio clip.

# you need to youtube-dl the video with id mPdcedZ9GxE for the example
# WARNING! example requires 3GB of free RAM

from moviepy.editor import *

from numpy import *
from scipy.signal import *



# poppy bros:  8200 10000 68
def occurences(haystack, needle, threshold, peak_dist, needle_seconds):
    '''Finds all occurences of needle in haystack (both are audio encoded as 1xn-array)
    Magical constants: threshold and peak_dist. Different every time, adjust by hand!
    needle_seconds: length of the needle audio, in seconds
    >>Returns: array of occurences (in seconds)'''

    corr = abs(fftconvolve(haystack, needle, mode='same'))

    peaks = where(corr > threshold)[0]

    summary_peaks = []
    last_peak = -inf
    for peak in peaks:
        if peak - last_peak > peak_dist:
            summary_peaks.append(peak)
            last_peak = peak

    return (array(summary_peaks) / len(corr) * needle_seconds, corr[summary_peaks])



if __name__ == '__main__':
    '''Example: Find all >shine collect<s in a Super Mario Sunshine Longplay'''

    s = "[Longplay] Super Mario Sunshine - Part 1-mPdcedZ9GxE.mp4"
    clip = VideoFileClip(s).subclip((0,0), (12,0))

    t = "shine_collect.mp4"
    clip.subclip((3,18),(3,22)).write_videofile(t)  # we only need to generate it once
    template = VideoFileClip(t)
    print('step 1/3 (loading haystack and needle) done')

    ts = template.audio.to_soundarray()[::-1]  # convolution is REVERSED correlation
    # i do not know if this is ok, but i have a 2xn-array and need a 1xn-array...
    ts = ts[:,0] + ts[:,1]
    cs = clip.audio.to_soundarray()
    cs = cs[:,0] + cs[:,1]

    print('step 2/3 (converting audio to_soundarray) done')
    set_printoptions(threshold=inf)
    print(occurences(cs, ts, 3500, 400000, 720))
    print('step 3/3 (finding all needles in haystack) done')

    # expected
    #  3:18   198
    # 11:11   671
    # 15:41   941
    # 20:45  1245

    # result (first 30 mins, using 12 GB of mem LOL) (12 min works in 3 ;))
    # array([  198.94752835,   671.8446712 ,   941.03247165,  1245.3403855 ])