How to process audio stream in realtime - python-3.x

Question

I have a setup with a raspberry pi 3 running latest jessie with all updates installed in which i provide a A2DP bluetooth sink where i connect with a phone to play some music.
Via pulseaudio, the source (phone) is routed to the alsa output (sink). This works reasonably well.
I now want to analyze the audio stream using python3.4 with librosa and i found a promising example using pyaudio which got adjusted to use the pulseaudio input (which magically works because its the default) instead of a wavfile:
"""PyAudio Example: Play a wave file (callback version)."""
import pyaudio
import wave
import time
import sys
import numpy
# instantiate PyAudio (1)
p = pyaudio.PyAudio()
# define callback (2)
def callback(in_data, frame_count, time_info, status):
# convert data to array
data = numpy.fromstring(data, dtype=numpy.float32)
# process data array using librosa
# ...
return (None, pyaudio.paContinue)
# open stream using callback (3)
stream = p.open(format=p.paFloat32,
channels=1,
rate=44100,
input=True,
output=False,
frames_per_buffer=int(44100*10),
stream_callback=callback)
# start the stream (4)
stream.start_stream()
# wait for stream to finish (5)
while stream.is_active():
time.sleep(0.1)
# stop stream (6)
stream.stop_stream()
stream.close()
wf.close()
# close PyAudio (7)
p.terminate()
Now while the data flow works in principle, there is a delay (length of buffer) with which the stream_callback gets called. Since the docs state
Note that PyAudio calls the callback function in a separate thread.
i would have assumed that while the callback is worked on, the buffer keeps filling in the mainthread. Of course, there would be an initial delay to fill the buffer, afterwards i expected to get synchronous flow.
I need a longer portion in the buffer (see frames_in_buffer) for librosa to be able to perfom analysis correctly.
How is something like this possible? Is it a limitation of the software-ports for the raspberry ARM?
I found other answers, but they use the blocking I/O. How would i wrap this into a thread so that librosa analysis (which might take some time) does not block the buffer filling?
This blog seems to fight performance issues with cython, but i dont think the delay is a performance issue. Or might it? Others seem to need some ALSA tweaks but would this help while using pulseaudio?
Thanks, any input appreciated!