def _convert_audio_and_split_sentences(extracted_dir, data_set, dest_dir):
source_dir = os.path.join(extracted_dir, data_set)
target_dir = os.path.join(extracted_dir, dest_dir)
if not os.path.exists(target_dir):
os.makedirs(target_dir)
# Loop over transcription files and split each one
#
# The format for each file 1-2.trans.txt is:
# 1-2-0 transcription of 1-2-0.flac
# 1-2-1 transcription of 1-2-1.flac
# ...
#
# Each file is then split into several files:
# 1-2-0.txt (contains transcription of 1-2-0.flac)
# 1-2-1.txt (contains transcription of 1-2-1.flac)
# ...
#
# We also convert the corresponding FLACs to WAV in the same pass
files = []
for root, dirnames, filenames in os.walk(source_dir):
for filename in fnmatch.filter(filenames, "*.trans.txt"):
trans_filename = os.path.join(root, filename)
with codecs.open(trans_filename, "r", "utf-8") as fin:
for line in fin:
# Parse each segment line
first_space = line.find(" ")
seqid, transcript = line[:first_space], line[first_space + 1 :]
# We need to do the encode-decode dance here because encode
# returns a bytes() object on Python 3, and text_to_char_array
# expects a string.
transcript = (
unicodedata.normalize("NFKD", transcript)
.encode("ascii", "ignore")
.decode("ascii", "ignore")
)
transcript = transcript.lower().strip()
# Convert corresponding FLAC to a WAV
flac_file = os.path.join(root, seqid + ".flac")
wav_file = os.path.join(target_dir, seqid + ".wav")
if not os.path.exists(wav_file):
tfm = Transformer()
tfm.set_output_format(rate=SAMPLE_RATE)
tfm.build(flac_file, wav_file)
wav_filesize = os.path.getsize(wav_file)
files.append((os.path.abspath(wav_file), wav_filesize, transcript))
return pandas.DataFrame(
data=files, columns=["wav_filename", "wav_filesize", "transcript"]
)
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 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 sph_to_wav(source_dir, target_dir):
"""Convert .sph files to .wav files."""
assert path.exists(source_dir) is True
if not path.exists(target_dir):
makedirs(target_dir)
for sph_file in glob(path.join(source_dir, "*.sph")):
transformer = Transformer()
if hp.tedlium_rate != 16000:
transformer.set_output_format(encoding='signed-integer',
channels=1,
rate=hp.tedlium_rate)
wav_filename = path.splitext(path.basename(sph_file))[0] + ".wav"
wav_file = path.join(target_dir, wav_filename)
transformer.build(sph_file, wav_file)
def compressed_wav_to_full(source_dir, target_dir):
"""Convert compressed wav files to full wav files."""
assert path.exists(source_dir) is True
if not path.exists(target_dir):
makedirs(target_dir)
for compressed_file in glob(path.join(source_dir, "*.wav")):
transformer = Transformer()
if hp.callhome_rate == 8000:
transformer.set_output_format(encoding='signed-integer', channels=1) # Also set single channel.
else: # Do resampling if specified.
transformer.set_output_format(encoding='signed-integer', channels=1, rate=hp.callhome_rate)
wav_filename = path.basename(compressed_file)
wav_file = path.join(target_dir, wav_filename)
transformer.build(compressed_file, wav_file)
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
# !/usr/bin/python3
import os
import tqdm
from sox import Transformer
SAMPLE_RATE = 16000
remove_flac = False
path = '/home/dsmolen/agh/LibriSpeech/'
i = 0
tq = tqdm.tqdm(os.walk(path, topdown=False))
for root, dirs, files in tq:
for name in files:
if name.endswith('.flac'):
tq.set_postfix(converted=i)
i += 1
name = name[:-5]
flac_file = os.path.join(root, name + ".flac")
wav_file = os.path.join(root, name + ".wav")
if not os.path.exists(wav_file):
tfm = Transformer()
tfm.set_output_format(rate=SAMPLE_RATE)
tfm.build(flac_file, wav_file)
if remove_flac:
os.remove(flac_file)
import tensorflow as tf
import pandas as pd
import os
import unicodedata
import tqdm
import logging
import librosa
import numpy as np
import soundfile
from sklearn.model_selection import train_test_split
logging.basicConfig(level=logging.NOTSET)
logging.getLogger('sox').setLevel(logging.ERROR)
FLAGS = tf.compat.v1.app.flags.FLAGS
tfm = Transformer()
tfm.set_output_format(rate=16000)
def main(_):
source_dir = FLAGS.source_dir
data = []
df_details = pd.read_csv(os.path.join(source_dir, "validated.tsv"),
sep="\t",
header=0)
with tqdm.tqdm(total=len(df_details.index)) as bar:
for i in df_details.index:
file_name = df_details["path"][i]
source_file = os.path.join(source_dir, "clips/" + file_name)
wav_file = os.path.join(
os.path.dirname(__file__),
"../data/common-voice-mozilla/Common-Voice-Mozilla/wav-files/"
