I am manipulating audio using numpy with IPython Notebook and want to use IPython.display.Audio for listening to the numpy arrays. Unfortunately, auto-normalization tampers with the results. Example:

# Generate a sound
import IPython
import numpy as np
framerate = 44100
t = np.linspace(0,5,framerate*5)
tone = np.sin(2*np.pi*220*t)
antitone = np.sin(2*np.pi*220*t + np.pi)
IPython.display.Audio(tone+antitone, rate=framerate)

Adding a sin wav to itself shifted by 180 deg should give total silence. Instead, auto-normalization amplifies the floating point errors. The problem is in the IPython.lib.display.Audio _make_wav method, which always normalizes the numpy array(see 'scaled' variable). I think that we should have a normalize keyword argument, so that Audio can be used for audio analysis. Something like:

Audio(tone+antitone, rate=framerate, normalize=False)