def main():
parser = argparse.ArgumentParser()
parser.add_argument("-csv",
"--csv-files",
help="Str. Filenames as a comma separated list",
required=True)
parser.add_argument("--sample-rate",
type=int,
default=16000,
required=False,
help="Audio sample rate")
parser.add_argument("--channels",
type=int,
default=1,
required=False,
help="Audio channels")
parser.add_argument("--bits-per-sample",
type=int,
default=16,
required=False,
help="Audio bits per sample")
args = parser.parse_args()
in_files = [os.path.abspath(i) for i in args.csv_files.split(",")]
csv_dataframe = read_csvs(in_files)
total_bytes = csv_dataframe['wav_filesize'].sum()
total_files = len(csv_dataframe.index)
bytes_without_headers = total_bytes - 44 * total_files
total_time = bytes_without_headers / (args.sample_rate * args.channels *
args.bits_per_sample / 8)
print('total_bytes', total_bytes)
print('total_files', total_files)
print('bytes_without_headers', bytes_without_headers)
print('total_time', secs_to_hours(total_time))
def _maybe_convert_set(input_tsv,
audio_dir,
label_filter,
space_after_every_character=None):
output_csv = path.join(audio_dir,
os.path.split(input_tsv)[-1].replace('tsv', 'csv'))
print("Saving new DeepSpeech-formatted CSV file to: ", output_csv)
# Get audiofile path and transcript for each sentence in tsv
samples = []
with open(input_tsv, encoding='utf-8') as input_tsv_file:
reader = csv.DictReader(input_tsv_file, delimiter='\t')
for row in reader:
samples.append((row['path'], row['sentence']))
# Keep track of how many samples are good vs. problematic
counter = {
'all': 0,
'failed': 0,
'invalid_label': 0,
'too_short': 0,
'too_long': 0,
'total_time': 0
}
lock = RLock()
num_samples = len(samples)
rows = []
def one_sample(sample):
""" Take a audio file, and optionally convert it to 16kHz WAV """
mp3_filename = path.join(audio_dir, sample[0])
if not path.splitext(mp3_filename.lower())[1] == '.mp3':
mp3_filename += ".mp3"
# Storing wav files next to the mp3 ones - just with a different suffix
wav_filename = path.splitext(mp3_filename)[0] + ".wav"
_maybe_convert_wav(mp3_filename, wav_filename)
file_size = -1
frames = 0
if path.exists(wav_filename):
file_size = path.getsize(wav_filename)
frames = int(
subprocess.check_output(['soxi', '-s', wav_filename],
stderr=subprocess.STDOUT))
label = label_filter(sample[1])
with lock:
if file_size == -1:
# Excluding samples that failed upon conversion
counter['failed'] += 1
elif label is None:
# Excluding samples that failed on label validation
counter['invalid_label'] += 1
elif int(frames / SAMPLE_RATE * 1000 / 10 / 2) < len(str(label)):
# Excluding samples that are too short to fit the transcript
counter['too_short'] += 1
elif frames / SAMPLE_RATE > MAX_SECS:
# Excluding very long samples to keep a reasonable batch-size
counter['too_long'] += 1
else:
# This one is good - keep it for the target CSV
rows.append(
(os.path.split(wav_filename)[-1], file_size, label))
counter['all'] += 1
counter['total_time'] += frames
print("Importing mp3 files...")
pool = Pool(cpu_count())
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, _ in enumerate(pool.imap_unordered(one_sample, samples), start=1):
bar.update(i)
bar.update(num_samples)
pool.close()
pool.join()
with open(output_csv, 'w', encoding='utf-8') as output_csv_file:
print('Writing CSV file for DeepSpeech.py as: ', output_csv)
writer = csv.DictWriter(output_csv_file, fieldnames=FIELDNAMES)
writer.writeheader()
bar = progressbar.ProgressBar(max_value=len(rows), widgets=SIMPLE_BAR)
for filename, file_size, transcript in bar(rows):
if space_after_every_character:
writer.writerow({
'wav_filename': filename,
'wav_filesize': file_size,
'transcript': ' '.join(transcript)
})
else:
writer.writerow({
'wav_filename': filename,
'wav_filesize': file_size,
'transcript': transcript
})
print('Imported %d samples.' %
(counter['all'] - counter['failed'] - counter['too_short'] -
counter['too_long']))
if counter['failed'] > 0:
print('Skipped %d samples that failed upon conversion.' %
counter['failed'])
if counter['invalid_label'] > 0:
print('Skipped %d samples that failed on transcript validation.' %
counter['invalid_label'])
if counter['too_short'] > 0:
print(
'Skipped %d samples that were too short to match the transcript.' %
counter['too_short'])
if counter['too_long'] > 0:
print('Skipped %d samples that were longer than %d seconds.' %
(counter['too_long'], MAX_SECS))
print('Final amount of imported audio: %s.' %
secs_to_hours(counter['total_time'] / SAMPLE_RATE))
def _maybe_convert_sets(target_dir, extracted_data):
extracted_dir = path.join(target_dir, extracted_data)
# override existing CSV with normalized one
target_csv_template = os.path.join(target_dir, ARCHIVE_DIR_NAME, ARCHIVE_NAME.replace('.tgz', '_{}.csv'))
if os.path.isfile(target_csv_template):
return
wav_root_dir = os.path.join(extracted_dir)
# Get audiofile path and transcript for each sentence in tsv
samples = []
glob_dir = os.path.join(wav_root_dir, '**/metadata.csv')
for record in glob(glob_dir, recursive=True):
for sk in SKIP_LIST:
if not (sk in record):
with open(record, 'r') as rec:
for re in rec.readlines():
re = re.strip().split('|')
audio = os.path.join(os.path.dirname(record), 'wavs', re[0] + '.wav')
transcript = re[2]
samples.append((audio, transcript))
# Keep track of how many samples are good vs. problematic
counter = {'all': 0, 'failed': 0, 'invalid_label': 0, 'too_short': 0, 'too_long': 0, 'total_time': 0}
lock = RLock()
num_samples = len(samples)
rows = []
def one_sample(sample):
""" Take a audio file, and optionally convert it to 16kHz WAV """
wav_filename = sample[0]
file_size = -1
frames = 0
if path.exists(wav_filename):
file_size = path.getsize(wav_filename)
frames = int(subprocess.check_output(['soxi', '-s', wav_filename], stderr=subprocess.STDOUT))
label = label_filter(sample[1])
with lock:
if file_size == -1:
# Excluding samples that failed upon conversion
counter['failed'] += 1
elif label is None:
# Excluding samples that failed on label validation
counter['invalid_label'] += 1
elif int(frames/SAMPLE_RATE*1000/15/2) < len(str(label)):
# Excluding samples that are too short to fit the transcript
counter['too_short'] += 1
elif frames/SAMPLE_RATE > MAX_SECS:
# Excluding very long samples to keep a reasonable batch-size
counter['too_long'] += 1
else:
# This one is good - keep it for the target CSV
rows.append((wav_filename, file_size, label))
counter['all'] += 1
counter['total_time'] += frames
print("Importing WAV files...")
pool = Pool(cpu_count())
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, _ in enumerate(pool.imap_unordered(one_sample, samples), start=1):
bar.update(i)
bar.update(num_samples)
pool.close()
pool.join()
with open(target_csv_template.format('train'), 'w') as train_csv_file: # 80%
with open(target_csv_template.format('dev'), 'w') as dev_csv_file: # 10%
with open(target_csv_template.format('test'), 'w') as test_csv_file: # 10%
train_writer = csv.DictWriter(train_csv_file, fieldnames=FIELDNAMES)
train_writer.writeheader()
dev_writer = csv.DictWriter(dev_csv_file, fieldnames=FIELDNAMES)
dev_writer.writeheader()
test_writer = csv.DictWriter(test_csv_file, fieldnames=FIELDNAMES)
test_writer.writeheader()
for i, item in enumerate(rows):
transcript = validate_label(item[2])
if not transcript:
continue
wav_filename = item[0]
i_mod = i % 10
if i_mod == 0:
writer = test_writer
elif i_mod == 1:
writer = dev_writer
else:
writer = train_writer
writer.writerow(dict(
wav_filename=wav_filename,
wav_filesize=os.path.getsize(wav_filename),
transcript=transcript,
))
print('Imported %d samples.' % (counter['all'] - counter['failed'] - counter['too_short'] - counter['too_long']))
if counter['failed'] > 0:
print('Skipped %d samples that failed upon conversion.' % counter['failed'])
if counter['invalid_label'] > 0:
print('Skipped %d samples that failed on transcript validation.' % counter['invalid_label'])
if counter['too_short'] > 0:
print('Skipped %d samples that were too short to match the transcript.' % counter['too_short'])
if counter['too_long'] > 0:
print('Skipped %d samples that were longer than %d seconds.' % (counter['too_long'], MAX_SECS))
print('Final amount of imported audio: %s.' % secs_to_hours(counter['total_time'] / SAMPLE_RATE))
def _maybe_convert_set(_dir):
input_tsv = _clean_text(join(_dir, 'utt_spk_text.tsv'))
samples = []
with open(input_tsv) as input_tsv_file:
reader = csv.DictReader(input_tsv_file,
delimiter='\t',
fieldnames=['path', 'id', 'sentence'])
for row in reader:
samples.append((row['path'], row['sentence']))
counter = {
'all': 0,
'failed': 0,
'invalid_label': 0,
'too_short': 0,
'too_long': 0,
'total_time': 0
}
lock = RLock()
num_samples = len(samples)
rows = []
def one_sample(sample):
_path = sample[0]
flac_filename = os.path.abspath(
path.join(_dir, 'data', _path[:2], _path + '.flac'))
wav_file = join(_path[:2], _path + ".wav")
wav_filename = join(STORE, wav_file)
_maybe_convert_wav(flac_filename, wav_filename)
_file_size = -1
frames = 0
if path.exists(wav_filename):
_file_size = path.getsize(wav_filename)
frames = int(
subprocess.check_output(['soxi', '-s', wav_filename],
stderr=subprocess.STDOUT))
label = label_filter_fun(sample[1])
with lock:
if _file_size == -1:
# Excluding samples that failed upon conversion
counter['failed'] += 1
elif label is None:
# Excluding samples that failed on label validation
counter['invalid_label'] += 1
elif int(frames / SAMPLE_RATE * 1000 / 10 / 2) < len(str(label)):
# Excluding samples that are too short to fit the transcript
counter['too_short'] += 1
elif frames / SAMPLE_RATE > MAX_SECS:
# Excluding very long samples to keep a reasonable batch-size
counter['too_long'] += 1
else:
# This one is good - keep it for the target CSV
rows.append((wav_file, _file_size, label))
counter['all'] += 1
counter['total_time'] += frames
print("Importing mp3 files...")
pool = Pool(cpu_count())
bar = progressbar.ProgressBar(maxval=num_samples, widgets=SIMPLE_BAR)
for i, _ in enumerate(pool.imap_unordered(one_sample, samples), start=0):
bar.update(i)
bar.update(num_samples)
pool.close()
pool.join()
with open(OUTPUT_CSV, 'w') as output_csv_file:
print('Writing CSV file for DeepSpeech.py as: ', OUTPUT_CSV)
writer = csv.DictWriter(output_csv_file, fieldnames=FIELDNAMES)
writer.writeheader()
bar = progressbar.ProgressBar(maxval=len(rows), widgets=SIMPLE_BAR)
for filename, file_size, transcript in bar(rows):
writer.writerow({
'wav_filename': filename,
'wav_filesize': file_size,
'transcript': transcript
})
print('Imported %d samples.' %
(counter['all'] - counter['failed'] - counter['too_short'] -
counter['too_long']))
if counter['failed'] > 0:
print('Skipped %d samples that failed upon conversion.' %
counter['failed'])
if counter['invalid_label'] > 0:
print('Skipped %d samples that failed on transcript validation.' %
counter['invalid_label'])
if counter['too_short'] > 0:
print(
'Skipped %d samples that were too short to match the transcript.' %
counter['too_short'])
if counter['too_long'] > 0:
print('Skipped %d samples that were longer than %d seconds.' %
(counter['too_long'], MAX_SECS))
print('Final amount of imported audio: %s.' %
secs_to_hours(counter['total_time'] / SAMPLE_RATE))
def _maybe_convert_sets(target_dir, extracted_data, english_compatible=False):
extracted_dir = path.join(target_dir, extracted_data)
# override existing CSV with normalized one
target_csv_template = os.path.join(target_dir, 'ts_' + ARCHIVE_NAME + '_{}.csv')
if os.path.isfile(target_csv_template):
return
path_to_original_csv = os.path.join(extracted_dir, 'data.csv')
with open(path_to_original_csv) as csv_f:
data = [
d for d in csv.DictReader(csv_f, delimiter=',')
if float(d['duration']) <= MAX_SECS
]
# Keep track of how many samples are good vs. problematic
counter = {'all': 0, 'failed': 0, 'invalid_label': 0, 'too_short': 0, 'too_long': 0, 'total_time': 0}
lock = RLock()
num_samples = len(data)
rows = []
wav_root_dir = extracted_dir
def one_sample(sample):
""" Take a audio file, and optionally convert it to 16kHz WAV """
orig_filename = path.join(wav_root_dir, sample['path'])
# Storing wav files next to the wav ones - just with a different suffix
wav_filename = path.splitext(orig_filename)[0] + ".converted.wav"
_maybe_convert_wav(orig_filename, wav_filename)
file_size = -1
frames = 0
if path.exists(wav_filename):
file_size = path.getsize(wav_filename)
frames = int(subprocess.check_output(['soxi', '-s', wav_filename], stderr=subprocess.STDOUT))
label = sample['text']
with lock:
if file_size == -1:
# Excluding samples that failed upon conversion
counter['failed'] += 1
elif label is None:
# Excluding samples that failed on label validation
counter['invalid_label'] += 1
elif int(frames/SAMPLE_RATE*1000/10/2) < len(str(label)):
# Excluding samples that are too short to fit the transcript
counter['too_short'] += 1
elif frames/SAMPLE_RATE > MAX_SECS:
# Excluding very long samples to keep a reasonable batch-size
counter['too_long'] += 1
else:
# This one is good - keep it for the target CSV
rows.append((wav_filename, file_size, label))
counter['all'] += 1
counter['total_time'] += frames
print("Importing wav files...")
pool = Pool(cpu_count())
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, _ in enumerate(pool.imap_unordered(one_sample, data), start=1):
bar.update(i)
bar.update(num_samples)
pool.close()
pool.join()
with open(target_csv_template.format('train'), 'w') as train_csv_file: # 80%
with open(target_csv_template.format('dev'), 'w') as dev_csv_file: # 10%
with open(target_csv_template.format('test'), 'w') as test_csv_file: # 10%
train_writer = csv.DictWriter(train_csv_file, fieldnames=FIELDNAMES)
train_writer.writeheader()
dev_writer = csv.DictWriter(dev_csv_file, fieldnames=FIELDNAMES)
dev_writer.writeheader()
test_writer = csv.DictWriter(test_csv_file, fieldnames=FIELDNAMES)
test_writer.writeheader()
for i, item in enumerate(rows):
print('item', item)
transcript = validate_label(cleanup_transcript(item[2], english_compatible=english_compatible))
if not transcript:
continue
wav_filename = os.path.join(target_dir, extracted_data, item[0])
i_mod = i % 10
if i_mod == 0:
writer = test_writer
elif i_mod == 1:
writer = dev_writer
else:
writer = train_writer
writer.writerow(dict(
wav_filename=wav_filename,
wav_filesize=os.path.getsize(wav_filename),
transcript=transcript,
))
print('Imported %d samples.' % (counter['all'] - counter['failed'] - counter['too_short'] - counter['too_long']))
if counter['failed'] > 0:
print('Skipped %d samples that failed upon conversion.' % counter['failed'])
if counter['invalid_label'] > 0:
print('Skipped %d samples that failed on transcript validation.' % counter['invalid_label'])
if counter['too_short'] > 0:
print('Skipped %d samples that were too short to match the transcript.' % counter['too_short'])
if counter['too_long'] > 0:
print('Skipped %d samples that were longer than %d seconds.' % (counter['too_long'], MAX_SECS))
print('Final amount of imported audio: %s.' % secs_to_hours(counter['total_time'] / SAMPLE_RATE))