Spaces:
Sleeping
Sleeping
| import numpy | |
| #@njit SLOWER | |
| def detect_bpm(audio, samplerate, bpm_low=40, bpm_high=300, bpm_step=0.1, mode=1, shift_step=10): | |
| """A very slow and inefficient algorithm!""" | |
| audio = numpy.asarray(audio) | |
| audio = (audio[0] + audio[1]).astype(numpy.float32) | |
| length=len(audio) | |
| mlength=length- int( 1 / ((bpm_low / 60) / samplerate) ) # to make sure those parts do not affect the calculation as they will be cut sometimes | |
| #audio[:int(spb_low)]=0 # to make sure those parts do not affect the calculation as they will be cut sometimes | |
| bpmdiffs=[] | |
| bpmdiffsi=[] | |
| minimum=100000000 | |
| for i in range(int((bpm_high-bpm_low)/bpm_step)): | |
| spb=int(round(1/(((bpm_low + i*bpm_step) / 60) / samplerate))) | |
| # audio is reshaped into a 2d array with bpm | |
| end=-int(length % spb) | |
| if end == 0: end = length | |
| image = audio[:end].reshape(-1, spb) | |
| if mode == 1: image=image.T | |
| # diff21, diff22, diff41, diff42 = image[:-2].flatten(), image[2:].flatten(), image[:-4].flatten(), image[4:].flatten() | |
| # difference=abs( numpy.dot(diff21, diff22)/(numpy.linalg.norm(diff21)*numpy.linalg.norm(diff22)) + numpy.dot(diff41, diff42)/(numpy.linalg.norm(diff41)*numpy.linalg.norm(diff42)) ) | |
| diff2=numpy.abs ( (image[:-2] - image[2:]).flatten()[:mlength] ) | |
| diff4=numpy.abs ( (image[:-4] - image[4:]).flatten()[:mlength] ) | |
| difference=numpy.sum(diff2*diff2*diff2*diff2) + numpy.sum(diff4*diff4*diff4*diff4) | |
| # for i in range(len(image)-1): | |
| # difference.append(numpy.sum(image[i]-image[i]+1)) | |
| if mode == 3: | |
| image=image.T | |
| diff2=numpy.abs ( (image[:-2] - image[2:]).flatten()[:mlength] ) | |
| diff4=numpy.abs ( (image[:-4] - image[4:]).flatten()[:mlength] ) | |
| difference=numpy.sum(diff2*diff2*diff2*diff2) + numpy.sum(diff4*diff4*diff4*diff4) | |
| bpmdiffs.append(spb) | |
| bpmdiffsi.append(difference) | |
| if difference<minimum: | |
| #print(f'{spb}: testing BPM = {(1/spb)*60*samplerate}; value = {difference}') | |
| print(i) | |
| minimum=difference | |
| spb = bpmdiffs[numpy.argmin(numpy.asarray(bpmdiffsi))] | |
| #print(f'BPM = {(1/spb)*60*samplerate}') | |
| bpmdiffs=[] | |
| bpmdiffsi=[] | |
| #audio[int(spb):]=0 | |
| print(spb) | |
| for shift in range(0, spb, shift_step): | |
| #print(shift) | |
| end=-int(length % spb) | |
| if end == 0: end = length+shift | |
| image = audio[shift:end].reshape(-1, spb) | |
| length-=shift_step | |
| if mode == 1: image=image.T | |
| diff = numpy.abs ( (image[:-1] - image[1:]).flatten()[:mlength] ) | |
| difference=numpy.sum(diff*diff) | |
| if mode == 3: | |
| image=image.T | |
| diff = numpy.abs ( (image[:-1] - image[1:]).flatten()[:mlength] ) | |
| difference += numpy.sum(diff*diff) | |
| bpmdiffs.append(shift) | |
| bpmdiffsi.append(difference) | |
| #if shift%1000==0: print(f'testing shift = {shift}; value = {difference}') | |
| shift = bpmdiffs[numpy.argmin(numpy.asarray(bpmdiffsi))] | |
| #print(f'BPM = {(1/spb)*60*samplerate}; shift = {shift/samplerate} sec.') | |
| return numpy.arange(shift, length, spb) |