def dumpFeaturePho(wav_path,
textgrid_path,
recordings,
syllableTierName,
phonemeTierName):
"""
dump the log-mel feature for each phoneme
:param wav_path: path of the wavs
:param textgrid_path: path of the textgrid
:param recordings: recording file names
:param syllableTierName: syllable textgrid tier name - dianSilence
:param phonemeTierName: phoneme textgrid tier name - details
:return:
"""
# feature dictionary
dic_pho_embedding = {}
for _, pho in enumerate(set(dic_pho_map.values())):
dic_pho_embedding[pho] = []
for artist_path, recording in recordings:
nestedPhonemeLists, numSyllables, numPhonemes \
= syllableTextgridExtraction(textgrid_path,
join(artist_path, recording),
syllableTierName,
phonemeTierName)
wav_full_filename = join(wav_path, artist_path, recording+'.wav')
log_mel = getMFCCBandsMadmom(audio_fn=wav_full_filename, fs=fs, hopsize_t=hopsize_t)
# go through all the phonemes in the list
for ii, pho in enumerate(nestedPhonemeLists):
print('calculating ', recording, ' and phoneme ', str(ii), ' of ', str(len(nestedPhonemeLists)))
for p in pho[1]:
# map from annotated xsampa to a readable notation with a dictionary
try:
key = dic_pho_map[p[2]]
except KeyError:
print(artist_path, recording)
print(ii, p[2])
raise
sf = int(round(p[0] * fs / float(hopsize))) # starting frame
ef = int(round(p[1] * fs / float(hopsize))) # ending frame
log_mel_p = log_mel[sf:ef, :] # log-mel phoneme
if len(log_mel_p):
dic_pho_embedding[key].append(log_mel_p)
return dic_pho_embedding
def dumpAudioPhn(wav_path,
textgrid_path,
recordings,
lineTierName,
phonemeTierName):
"""
dump audio of each phone
:param wav_path:
:param textgrid_path:
:param recordings:
:param lineTierName:
:param phonemeTierName:
:return:
"""
##-- dictionary feature
dic_pho_wav = {}
for _, pho in enumerate(set(dic_pho_map.values())):
dic_pho_wav[pho] = []
for artist_path, recording in recordings:
nestedPhonemeLists, numSyllables, numPhonemes \
= syllableTextgridExtraction(textgrid_path,
join(artist_path, recording),
lineTierName,
phonemeTierName)
# audio
wav_full_filename = join(wav_path, artist_path, recording + '.wav')
data_wav, fs_wav = sf.read(wav_full_filename)
for ii, pho in enumerate(nestedPhonemeLists):
print('calculating ', recording, ' and phoneme ', str(ii), ' of ', str(len(nestedPhonemeLists)))
for p in pho[1]:
# map from annotated xsampa to readable notation
try:
key = dic_pho_map[p[2]]
except KeyError:
print(artist_path, recording)
print(ii, p[2])
raise
st = int(round(p[0] * fs_wav)) # starting time
et = int(round(p[1] * fs_wav)) # ending time
pho_wav = data_wav[st: et]
if len(pho_wav):
dic_pho_wav[key].append(pho_wav)
return dic_pho_wav
def dumpFeaturePho(wav_path, textgrid_path, recordings, syllableTierName,
phonemeTierName):
'''
dump the MFCC for each phoneme
:param recordings:
:return:
'''
##-- dictionary feature
dic_pho_feature = {}
for _, pho in enumerate(set(dic_pho_map.values())):
dic_pho_feature[pho] = np.array([])
for artist_path, recording in recordings:
nestedPhonemeLists, numSyllables, numPhonemes \
= syllableTextgridExtraction(textgrid_path,
join(artist_path,recording),
syllableTierName,
phonemeTierName)
# audio
wav_full_filename = join(wav_path, artist_path, recording + '.wav')
mfcc = getMFCCBands2DMadmom(wav_full_filename,
fs,
hopsize_t,
channel=1)
for ii, pho in enumerate(nestedPhonemeLists):
print('calculating ', recording, ' and phoneme ', str(ii), ' of ',
str(len(nestedPhonemeLists)))
for p in pho[1]:
# map from annotated xsampa to readable notation
try:
key = dic_pho_map[p[2]]
except KeyError:
print(artist_path, recording)
print(ii, p[2])
raise
sf = int(round(p[0] * fs / float(hopsize))) # starting frame
ef = int(round(p[1] * fs / float(hopsize))) # ending frame
mfcc_p = mfcc[sf:ef, :] # phoneme syllable
if not len(dic_pho_feature[key]):
dic_pho_feature[key] = mfcc_p
else:
dic_pho_feature[key] = np.vstack(
(dic_pho_feature[key], mfcc_p))
return dic_pho_feature
def extract_log_mel_of_line(wav_path, textgrid_path, filename, num_line):
nestedPhonemeLists, numlines, numPhonemes \
= syllableTextgridExtraction(textgrid_path,
filename,
"line",
"details")
# audio filename
wav_full_filename = os.path.join(wav_path, filename + '.wav')
log_mel = getMFCCBandsMadmom(audio_fn=wav_full_filename, fs=fs, hopsize_t=hopsize_t)
line_list = nestedPhonemeLists[num_line]
sf = int(round(line_list[0][0] * fs / float(hopsize))) # starting frame
ef = int(round(line_list[0][1] * fs / float(hopsize))) # ending frame
return (log_mel[sf:ef], line_list[1])
def s_check(textgrid_path,
recordings,
parentTierName,
childTierName):
for artist_path, recording in recordings:
nestedLists, _, _ \
= syllableTextgridExtraction(textgrid_path=textgrid_path,
recording=os.path.join(artist_path,recording),
tier0=parentTierName,
tier1=childTierName)
for ii, line_list in enumerate(nestedLists):
print(artist_path, recording ,ii, len(line_list[1]))
if childTierName=='details':
for phn in line_list[1]:
try:
key = dic_pho_map[phn[2]]
except:
print(artist_path, ii, recording, phn[2])
raise KeyError