shnread.py

# read shorten (shn) files into numpy arrays

from __future__ import division
import os.path as path
import os
from subprocess import call
import struct
import tempfile

def shnread(shnfile):
    """Read shorten audio file, such as used by NIST, TIMIT and WSJ databases
    into numpy array. Assumes that `shorten` is in the PATH.
    Shorten can be found here: http://etree.org/shnutils/shorten/

    Arguments:
    :param shnfile: filename

    :return sig, fs:
      signal and sample rate
    """
    tmp = path.join(tempfile.mkdtemp(),
		    path.splitext(path.basename(sph_file))[0] + '.wav')
    with open(sph_file, 'rb') as f:
	f.read(8) # NIST_1A
	headerlen = int(f.read(8).strip())
	header = f.read(headerlen - 16).strip().split('\n')
	
    header = {e.split(' ')[0]: e.split(' ')[2] for e in header[:-1]}
    fs = int(header['sample_rate'])
    endianness = '<' if int(header['sample_byte_format']) == 1 else '>'
    bytes_per_sample = int(header['sample_n_bytes'])
    bit_depth = int(header['sample_sig_bits'])

    call(['shorten', '-x', '-d', str(headerlen), sph_file, tmp])
    with open(tmp, 'rb') as f:
	f.read(headerlen) # skip header
	bytes = f.read()

    format_map = {1: 'b', 2: 'h', 4: 'i', 8: 'q'}
    sig = np.fromiter((struct.unpack('{0}{1}'.format(endianness,
						     format_map[bytes_per_sample]),
				     bytes[i:i+bytes_per_sample])[0]
		       for i in xrange(0, len(bytes), bytes_per_sample)),
		      dtype='int{}'.format(bit_depth))
    sig = sig / max(abs(sig))
    os.unlink(tmp)
    return sig, fs