def create_datapoints(transformer: sox.Transformer, writers: Writers,
grid: pathlib.Path, audio: pathlib.Path):
"""Creates datapoints from a TextGrid."""
audio_file = audio / grid.parts[-2] / f"{grid.stem}.wav"
if audio_file.is_file():
resampled_audio = transformer.build_array(
input_filepath=str(audio_file))
tg = textgrid.TextGrid.fromFile(grid)
for interval in tg[0]:
start_time = interval.minTime
end_time = interval.maxTime
text = interval.mark
if text in writers.word_counts:
start_sample = int(
max((start_time - 0.1) * transformer.output_format["rate"],
0))
end_sample = int(
min((end_time + 0.1) * transformer.output_format["rate"],
resampled_audio.size))
utterance = resampled_audio[start_sample:end_sample]
writers.write(word=text,
sample_rate=transformer.output_format["rate"],
audio=utterance)
else:
print(f"File not found: {audio_file}")
def preprocess_wav(cls, fpath: Union[str, Path]) -> np.ndarray:
"""Load, resample, normalize and trim a waveform."""
transformer = Transformer()
transformer.norm()
transformer.silence(silence_threshold=1, min_silence_duration=0.1)
transformer.set_output_format(rate=cls.sample_rate,
bits=16,
channels=1)
wav = transformer.build_array(input_filepath=str(fpath))
wav = wav / (2**15)
return wav.astype(np.float32)
def sample_triplet(
transformer: sox.Transformer, utterances: Utterances
) -> ((np.ndarray, str), (np.ndarray, str), (np.ndarray, str)):
anchor = random.choice(utterances.words)
negative = random.choice(utterances.words)
while negative == anchor:
negative = random.choice(utterances.words)
anchor_file = random.choice(utterances.word_files[anchor])
positive_file = random.choice(utterances.word_files[anchor])
while positive_file == anchor_file:
positive_file = random.choice(utterances.word_files[anchor])
negative_file = random.choice(utterances.word_files[negative])
# print("anchor file", anchor_file)
# print("positive file", positive_file)
# print("negative file", negative_file)
anchor_word = anchor_file.parent.stem
anchor_audio = transformer.build_array(input_filepath=str(anchor_file))
positive_word = positive_file.parent.stem
positive_audio = transformer.build_array(input_filepath=str(positive_file))
negative_word = negative_file.parent.stem
negative_audio = transformer.build_array(input_filepath=str(negative_file))
#print("anchor-word", anchor_word)
#print("positive-word", positive_word)
#print("negative-word", negative_word)
return (anchor_audio, anchor_word), (positive_audio,
positive_word), (negative_audio,
negative_word)
def loadFile(data, max_timestep):
transformer = Transformer()
transformer.norm()
# transformer.silence(silence_threshold=1, min_silence_duration=0.1)
transformer.set_output_format(rate=16000, bits=16, channels=1)
wav = transformer.build_array(input_filepath=str(data))
wav = torch.tensor(wav / (2**15)).float()
length = len(wav)
if length > max_timestep:
start = 0
end = max_timestep
length = max_timestep
wav = wav[start:end]
length = torch.tensor(length).long()
return wav, length
def loadFile_thread_exec(data):
wavs = []
lengths = []
for i in range(len(data)):
fullPath = data[i]
transformer = Transformer()
transformer.norm()
transformer.silence(silence_threshold=1, min_silence_duration=0.1)
transformer.set_output_format(rate=16000, bits=16, channels=1)
wav = transformer.build_array(input_filepath=str(fullPath))
wav = torch.tensor(wav / (2**15)).float()
length = len(wav)
if length > max_timestep:
start = random.randint(0, int(length - max_timestep))
end = start + max_timestep
length = max_timestep
wav = wav[start:end]
wavs.append(wav)
lengths.append(torch.tensor(length).long())
return wavs, lengths
