Hi all, long time lurker and first time poster. I'm writing a Python program which does analysis on mp3 files, so I'd like to read in the mp3 data as a numpy array. I'm using pymad, which seems stable and works well. When you read in a frame of data from pymad, it gets returned as a buffer of unicode hex values and characters, eg,


The format of these values is Left channel MSB, LSB, Right channel MSB, LSB, ... and so on. So, to convert this into a numpy array, here's my function:

def bufToNumpy( inBuf ):
  bufLength = len(inBuf)
  outBuf = np.zeros(bufLength/2)
  for i in range(bufLength/2):
    outBuf[i] = struct.unpack('h', inBuf[2*i] + inBuf[2*i+1])[0]
  return outBuf

As you can see, I am using struct's unpack function to grab each pair of MSB, LSB and converting it into a 16 bit int, which is stored in a numpy array. However, because I'm doing this per-number, it takes a LONG time (for a typical song I call unpack about 10 million times). Is there a way I can vectorize this? In other words, I'd like to just call struct.unpack once and get out an array (of some kind) of 16 bit integers. Thanks for any tips!


6 Years
Discussion Span
Last Post by craffel

Nevermind, figured it out! You can just use numpy's built-in frombuffer() function. Eg, my function above can be replaced by

np.frombuffer(buffy, 'h')

That function is about 2000 times faster than the one I posted above, for the buffer size I was dealing with (4000 values).


This question has already been answered. Start a new discussion instead.
Have something to contribute to this discussion? Please be thoughtful, detailed and courteous, and be sure to adhere to our posting rules.