def train(filelist,out='./out_dir/',model='UBM_0'):
fin = open(filelist)
ubm_list = out+model
fout = open(ubm_list,'w')
for i in fin:
wavname = i.strip()
basename = tools.gen_uid(wavname)
featname = '%s/%s_feat.mfc'%(out,basename)
sadname = '%s/%s_sad.txt'%(out,basename)
attr = {'audio':wavname,'mfcc':featname,'sad':sadname}
# SAD
sad.run_sad(attr)
# MFCC
feat.run_mfcc(attr)
fout.write(attr['mfcc']+' '+attr['sad']+'\n')
fin.close()
fout.close()
path = tools.set_path()
ubmname = '%s/%s'%(out,model)
command = '%s/sgminit -q --label=speech --file-list=%s %s'
exe_cmd = command%(path['audioseg'],ubm_list,ubmname)
os.system(exe_cmd)
return ubmname
def diarization(wavname,ubmname,out_dir):
out = out_dir+'/'
basename = tools.gen_uid(wavname)
sadname = '%s/%s_sad.txt'%(out,basename)
featname = '%s/%s_feat.mfc'%(out,basename)
bicname = '%s/%s_bic.txt'%(out,basename)
clustname = '%s/%s_cluster.txt'%(out,basename)
viterbiname = '%s/%s_viterbi.txt'%(out,basename)
attr = {'audio':wavname,
'mfcc':featname,
'sad':sadname,
'bic':bicname,
'cluster':clustname,
'viterbi':viterbiname}
# SAD
sad.run_sad(attr)
# MFCC
feat.run_mfcc(attr)
# BIC
bic.run_bic(attr,'audioseg')
# CLUSTERING
cluster.run_clustering(attr)
# Pick top clusters
labels, segment_starts,segment_ends = tools.read_segs(attr['cluster'])
top_n = tools.top_n_clusters(labels, segment_starts,segment_ends,n=2)
# Adapt UBM for each cluster.
cluster_gmms = {}
for i in top_n:
cluster = 'C%s'%(str(i))
gmmname = gmm.adapt(attr,cluster,ubmname)
cluster_gmms[cluster] = gmmname
# Resegmentation
hmmname = '%s/%s_hmm.txt'%(out,basename)
resegment.viterbi(attr,cluster_gmms,hmmname)
labs,starts,ends = tools.merge_segs(attr['viterbi'],attr['sad'])
return labs,starts,ends
path = tools.set_path()
sadname = attr['sad']
featname = attr['mfcc']
bicname = attr['bic']
if mode == 'audioseg':
command = '%s/sbic --segmentation=%s --label=speech %s %s'
exe_cmd = command % (path['audioseg'], sadname, featname, bicname)
os.system(exe_cmd)
return
elif mode == 'uniform':
# For cases where we'd rather break the signal
# into equal length segments.
segment_lengths = 2. * 16000. # in samples
unifrom_segmentation(bicname, sadname, segment_lengths)
return
else:
raise ('Segmentation mode not available!')
if __name__ == '__main__':
wavname = '/Users/navidshokouhi/Downloads/unimaquarie/projects/dolby-annotations/data/1_222_2_7_001-ch6-speaker16.wav'
basename = tools.gen_uid(wavname)
sadname = './%s_sad.txt' % (basename)
featname = './%s.mfc' % (basename)
bicname = './%s_bic.txt' % (basename)
attr = {'audio': wavname, 'sad': sadname, 'mfcc': featname, 'bic': bicname}
sad.run_sad(attr)
feat.run_mfcc(attr)
run_bic(attr, 'uniform')
