I have the following code for reading HTK feature files. The code below is working completely correct (verified it with unit tests and the output of the original HTK toolkit).

from HTK_model import FLOAT_TYPE
from numpy import array
from struct import unpack

def feature_reader(file_name):
    with open(file_name, 'rb') as in_f:
        #There are four standard headers. Sample period is not used
        num_samples = unpack('>i', in_f.read(4))[0]
        sample_period = unpack('>i', in_f.read(4))[0]
        sample_size = unpack('>h', in_f.read(2))[0]
        param_kind = unpack('>h', in_f.read(2))[0]

        compressed = bool(param_kind & 02000)

        #If compression is used, two matrices are defined. In that case the values are shorts, and the real values are:
        # (x+B)/A
        A = B = 0
        if compressed:
            A = array([unpack('>f',in_f.read(4))[0] for _ in xrange(sample_size/2)], dtype=FLOAT_TYPE)
            B = array([unpack('>f',in_f.read(4))[0] for _ in xrange(sample_size/2)], dtype=FLOAT_TYPE)
            #The first 4 samples were the matrices
            num_samples -= 4

        for _ in xrange(0,num_samples):
            if compressed:
                yield ((array( unpack('>' + ('h' * (sample_size//2)),in_f.read(sample_size)) ,dtype=FLOAT_TYPE) + B) / A)
            else:
                yield (array( unpack('>' + ('f' * (sample_size//4)),in_f.read(sample_size)), dtype=FLOAT_TYPE))

How can I speed up this code? And are there any things I should improve in this code?

I have the following code for reading HTK feature files. The code below is working completely correct (verified it with unit tests and the output of the original HTK toolkit).

from HTK_model import FLOAT_TYPE
from numpy import array
from struct import unpack

def feature_reader(file_name):
    with open(file_name, 'rb') as in_f:
        #There are four standard headers. Sample period is not used
        num_samples = unpack('>i', in_f.read(4))[0]
        sample_period = unpack('>i', in_f.read(4))[0]
        sample_size = unpack('>h', in_f.read(2))[0]
        param_kind = unpack('>h', in_f.read(2))[0]

        compressed = bool(param_kind & 02000)

        #If compression is used, two matrices are defined. In that case the values are shorts, and the real values are:
        # (x+B)/A
        A = B = 0
        if compressed:
            A = array([unpack('>f',in_f.read(4))[0] for _ in xrange(sample_size/2)], dtype=FLOAT_TYPE)
            B = array([unpack('>f',in_f.read(4))[0] for _ in xrange(sample_size/2)], dtype=FLOAT_TYPE)
            #The first 4 samples were the matrices
            num_samples -= 4

        for _ in xrange(0,num_samples):
            if compressed:
                yield ((array( unpack('>' + ('h' * (sample_size//2)),in_f.read(sample_size)) ,dtype=FLOAT_TYPE) + B) / A)
            else:
                yield (array( unpack('>' + ('f' * (sample_size//4)),in_f.read(sample_size)), dtype=FLOAT_TYPE))

How can I speed up this code, are there things I should improve in the code?