一般的な問題は発症検出と呼ばれ、試すことができる多くの方法があります。おそらくあなたのユースケースでは機能しない、非常に単純なソリューションを提供します。
from scipy.io import wavfile
from scipy.signal import argrelmax
from matplotlib.mlab import specgram
sr, x = wavfile.read(path) # read in a mono wav file
spec, freqs, time = specgram(x, NFFT=4096, Fs=sr, mode='psd') # compute power spectral density spectogram
spec2 = np.diff(spec, axis=1) # discrete difference in each frequency bin
spec2[spec2<0] = 0 # half-wave rectification
diff = np.sum(spec2, axis=0) # sum positive difference in each time bin
for peak in argrelmax(diff)[0]: # find peaks
print("onset between %f and %f." % (time[peak], time[peak+1]))