def extract_mfcc(input_wav_dir, output_feature_file, output_feature_dir):
# secure destination
mkdir_for_file(output_feature_file)
mkdir_for_dir(output_feature_dir)
# extract mfcc in single file format and dir format
import util
mkdir_for_dir(default_temp_dir)
util.make_feature(input_wav_dir, default_temp_dir)
def write_feat(feat, feat_id, outfile):
outfile.write(feat_id+'.wav\n')
for i in range(feat.shape[0]):
fline = '{:04d} {:04d} #' + ' {:f}'*feat.shape[1] +'\n'
outfile.write(fline.format(i, i+1, *feat[i]))
outfile.write('\n')
with open(output_feature_file, 'w') as outfile:
for f in sorted(os.listdir(default_temp_dir)):
feat = util.read_feature(os.path.join(default_temp_dir, f))
write_feat(feat, f[:-4], outfile)
for f in sorted(os.listdir(default_temp_dir)):
feat = util.read_feature(os.path.join(default_temp_dir, f))
util.write_feature(feat, os.path.join(output_feature_dir, f))
# cleanup, remove large files
mkdir_for_dir(default_temp_dir)
def read_feature_dir(feature_dir):
import util
feat, feat_id = [], []
for f in sorted(os.listdir(feature_dir)):
feature = util.read_feature(os.path.join(feature_dir, f))
feat.append(feature)
feat_id.append(f[:-4])
return feat, feat_id
def feature_dtw(self, query_index, inst_index):
Q = self.feature_query(query_index, inst_index)
similarity = []
for ___ in range(len(self.query_mlf.wav_list)):
similarity.append(float('Inf'))
for w in range(len(self.query_mlf.wav_list)):
f_file = self.feature_fold + self.query_mlf.wav_list[w] + '.mfc'
D = util.read_feature(f_file)
similarity[w] = util.warp(util.cos_dist(D, Q))
# similarity[w] = util.warp(- np.dot(D,Q.T))
print query_index, w, similarity[w]
return similarity
def feature_query(self, query_index, inst_index):
w_index, t_beg, t_end = self.query_answer[query_index][inst_index]
# print self.feature_fold + self.query_mlf.wav_list[w_index]
f_file = self.feature_fold + self.query_mlf.wav_list[w_index] + '.mfc'
return util.read_feature(f_file)[t_beg:t_end]