phoneme extraction

main.py

# coding: utf-8

# In[ ]:

import re
import commands
import numpy as np
import scipy.io.wavfile
import itertools
import cPickle
# In[ ]:

#cd /u/sotelo/ez-phones
#115

# In[ ]:
numfile = 0

# In[ ]:

def sec2step(seconds):
    return np.round(16000*(float(seconds))).astype('int')

def phoneme_list(x):
    # Given the output of pocketsphinx compute the list with the phonemes.
    data = [ re.split(' ',line) for line in x[1:]]
    ph, t1, t2 = zip(*[ line[0:3] for line in data if re.match("^\d+?\.\d+?$", line[1]) is not None])

    t1 = map(float,t1)
    t2 = map(float,t2)
    ph = list(ph)

    if t1[0] > 0.0:
        t2 = [t1[0]] + t2
        t1 = [0.0] + t1
        ph = ['*SIL'] + ph

    if t2[-1] < 2.0:
        t1 = t1 + [t2[-1]]
        t2 = t2 + [2.0]
        ph = ph + ['*SIL']

    result = [sec2step(t1[i+1]-t2[i])*['*SIL'] + sec2step(t2[i+1]-t1[i+1])*[ph[i+1]] for i in range(len(t2)-1)]
    result = list(itertools.chain(*result))

    return sec2step(t2[0]-t1[0])*[ph[0]] + result

# In[ ]:

NUMFILES = range(96,100)
for numfile in NUMFILES:
    print "New File"
    x = np.load('/data/lisatmp3/Lessac_Blizzard2013_segmented/train/sf_train_segmented_%i.npy' % numfile)

    results = []
    for i, z in enumerate(x):
        # Create temporary wav file and process with pocketsphinx
        scipy.io.wavfile.write('/dev/shm/temp.wav',16000, z.astype(np.int16))
        txt =commands.getstatusoutput('bash ps_shortcut.sh /dev/shm/temp.wav')[1]
        
        # Hack to get the results of pocketsphinx
        txt = re.split('\n',txt)
        last = next(j for j,line in enumerate(txt[::-1]) if re.match('INFO', line) is not None)
        
        # Parse results and append to results
        try:
            results.append(phoneme_list(txt[-last:]))
        except:
            results.append(['*SIL']*32000)
            print "Error"
        print i

    phonemes = set(itertools.chain(*results))
    phonemes = sorted(list(phonemes))

    ph_list = cPickle.load( open( "ph_list.p", "rb" ) )
    assert ph_list == phonemes

    phonemes = {key:i for i,key in enumerate(phonemes)}

    results_array = np.array([np.array([phonemes[step] for step in phon_list]) for phon_list in results])

    np.save('/data/lisatmp3/Blizzard2013_phonemes_pocketsphinx/train/sf_train_segmented_ph_%i.npy' % numfile,results_array)