def load_mfa_align():
converter = MfaTextGridConverter(
use_phones=SETTINGS.training.token_type == 'phone')
id_align_map = {}
for tg_path in args.align_folder.glob('**/*.TextGrid'):
tg = TextGrid(str(tg_path.absolute()))
audio_id = tg_path.name.split('.', 1)[0]
id_align_map[audio_id] = converter.convert(tg)
return id_align_map
def prepare_same_close_mic(part3_path):
check_dependencies()
from textgrids import TextGrid
recordings = []
supervisions = []
for audio_path in tqdm(
(part3_path / "AudioSameCloseMic").glob("*.wav"),
desc="Creating manifests for SameCloseMic",
):
try:
recording_id = audio_path.stem
recording = Recording.from_file(audio_path)
tg = TextGrid(
part3_path / f"ScriptsSame/{recording_id}.TextGrid", coding="utf-16"
)
segments = [
s
for s in (
SupervisionSegment(
id=f"{recording_id}-{idx}",
recording_id=recording_id,
start=segment.xmin,
# We're trimming the last segment's duration as it exceeds the actual duration of the recording.
# This is safe because if we end up with a zero/negative duration, the validation will catch it.
duration=min(
round(segment.xmax - segment.xmin, ndigits=8),
recording.duration - segment.xmin,
),
text=segment.text,
language="Singaporean English",
speaker=recording_id,
)
for idx, segment in enumerate(tg[recording_id])
if segment.text not in ("<S>", "<Z>") # skip silences
)
if s.duration > 0 # NSC has some bad segments
]
recordings.append(recording)
supervisions.extend(segments)
except:
print(f"Error when processing {audio_path} - skipping...")
return {
"recordings": RecordingSet.from_recordings(recordings),
"supervisions": SupervisionSet.from_segments(supervisions),
}
def read_phonemes(textgrid_fname, mfcc_len, phone_inventory):
tg = TextGrid(textgrid_fname)
phone_ids = np.zeros(int(tg['phones'][-1].xmax * 100), dtype=np.int64)
phone_ids[:] = -1
for interval in tg['phones']:
phone = interval.text.lower()
if phone in ['', 'sp', 'spn']:
phone = 'sil'
if phone[-1] in string.digits:
phone = phone[:-1]
ph_id = phone_inventory.index(phone)
phone_ids[int(interval.xmin * 100):int(interval.xmax * 100)] = ph_id
assert (phone_ids >= 0).all(), 'missing aligned phones'
phone_ids = phone_ids[1:mfcc_len +
1] # mfccs is 2-3 shorter due to edge effects
return phone_ids
def prepare_separate_phone_mic(part3_path):
check_dependencies()
from textgrids import TextGrid
recordings = []
supervisions = []
for audio_path in tqdm(
(part3_path / 'AudioSeparateIVR').rglob('**/*.wav'),
desc='Creating manifests for SeparateIVR'
):
try:
recording_id = f'{audio_path.parent.name}_{audio_path.stem}'
recording = Recording.from_file(audio_path)
tg = TextGrid(part3_path / f'ScriptsSeparate/{recording_id}.TextGrid', coding='utf-16')
segments = [
s for s in (
SupervisionSegment(
id=f'{recording_id}-{idx}',
recording_id=recording_id,
start=segment.xmin,
# We're trimming the last segment's duration as it exceeds the actual duration of the recording.
# This is safe because if we end up with a zero/negative duration, the validation will catch it.
duration=min(round(segment.xmax - segment.xmin, ndigits=8), recording.duration - segment.xmin),
text=segment.text,
language='Singaporean English',
speaker=recording_id,
)
for idx, segment in enumerate(tg[recording_id])
if segment.text not in ('<S>', '<Z>') # skip silences
)
if s.duration > 0 # NSC has some bad segments
]
supervisions.extend(segments)
recordings.append(recording)
except:
print(f'Error when processing {audio_path} - skipping...')
return {
'recordings': RecordingSet.from_recordings(recordings),
'supervisions': SupervisionSet.from_segments(supervisions)
}
def prepare_same_close_mic(part3_path):
check_dependencies()
from textgrids import TextGrid
recordings = []
supervisions = []
for audio_path in tqdm(
(part3_path / 'AudioSameCloseMic').glob('*.wav'),
desc='Creating manifests for SameCloseMic'
):
try:
recording_id = audio_path.stem
recording = Recording.from_wav(audio_path)
tg = TextGrid(part3_path / f'ScriptsSame/{recording_id}.TextGrid', coding='utf-16')
segments = [
s for s in (
SupervisionSegment(
id=f'{recording_id}-{idx}',
recording_id=recording_id,
start=segment.xmin,
duration=round(segment.xmax - segment.xmin, ndigits=8),
text=segment.text,
language='Singaporean English',
speaker=recording_id,
)
for idx, segment in enumerate(tg[recording_id])
if segment.text not in ('<S>', '<Z>') # skip silences
)
if s.duration > 0 # NSC has some bad segments
]
recordings.append(recording)
supervisions.extend(segments)
except:
print(f'Error when processing {audio_path} - skipping...')
return {
'recordings': RecordingSet.from_recordings(recordings),
'supervisions': SupervisionSet.from_segments(supervisions)
}
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--input-folder',
'-i',
dest='align_folder',
type=Path,
required=True)
args = parser.parse_args()
converter = MfaTextGridConverter()
ds_kwargs = dict(sr=SETTINGS.audio.sample_rate,
mono=SETTINGS.audio.use_mono,
words=[])
ds_path = SETTINGS.dataset.dataset_path
train_ds, dev_ds, test_ds = AudioClipDatasetLoader().load_splits(
ds_path, **ds_kwargs)
id_align_map = {}
for tg_path in args.align_folder.glob('*.TextGrid'):
tg = TextGrid(str(tg_path.absolute()))
audio_id = tg_path.name.split('.', 1)[0]
id_align_map[audio_id] = converter.convert(tg)
for ds in (train_ds, dev_ds, test_ds):
with AudioClipDatasetMetadataWriter(ds_path,
ds.set_type,
'aligned-',
mode='w') as writer:
for ex in tqdm(ds, total=len(ds)):
try:
transcription = id_align_map[ex.metadata.path.name.split(
'.', 1)[0]]
writer.write(
AlignedAudioClipMetadata(path=ex.metadata.path,
transcription=transcription))
except KeyError:
pass
line.rstrip()
for line in open("src/PhonationModeling/data/filelists/textgrid.lst")
]
cnt = 1
for wf, tf in zip(wav_lst, txtgrd_lst):
print(f"Processing {wf}")
# Read wav
sample_rate, wav_raw = wavfile.read(os.path.join(
data_root, "wavs", wf))
# Convert from 16-bit int to 32-bit float
wav_data = (wav_raw / pow(2, 15)).astype("float32")
# Read textgrid
txtgrd = TextGrid(os.path.join(data_root, tf))
tier_phone = txtgrd[f"s{cnt} - phone"] # tier containing phones
tier_c = txtgrd["ipp"] # tier containing creaky voices
cnt = cnt + 1
# Get creaky phone segments
ph_intvls = dict_phone_interval(tier_phone)["AA1"] # NOTE: phone
ph_segs = get_phone_segments(tier_c, ph_intvls, from_creaky=False)
wav_segs = get_wav_segments(wav_data, sample_rate, ph_segs)
# Save to wav
for i, w_seg in enumerate(wav_segs):
wavfile.write(
os.path.join(save_dir,
wf.rstrip(".wav") + f"_phone_AA1_{i:d}.wav"),
sample_rate,