def one_sample(sample):
mp3_filename = sample[0]
# 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)
frames = int(
subprocess.check_output(['soxi', '-s', wav_filename],
stderr=subprocess.STDOUT))
file_size = -1
if path.exists(wav_filename):
file_size = path.getsize(wav_filename)
frames = int(
subprocess.check_output(['soxi', '-s', wav_filename],
stderr=subprocess.STDOUT))
label = validate_label(sample[1])
rows = []
counter = get_counter()
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
return (counter, 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])
counter = get_counter()
rows = []
if file_size == -1:
# Excluding samples that failed upon conversion
print("conversion failure", wav_filename)
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
return (counter, rows)
def _maybe_convert_set(input_tsv, audio_dir, 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((path.join(audio_dir,
row['path']), row['sentence']))
counter = get_counter()
num_samples = len(samples)
rows = []
print("Importing mp3 files...")
pool = Pool()
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, processed in enumerate(pool.imap_unordered(one_sample, samples),
start=1):
counter += processed[0]
rows += processed[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
})
imported_samples = get_imported_samples(counter)
assert counter['all'] == num_samples
assert len(rows) == imported_samples
print_import_report(counter, SAMPLE_RATE, MAX_SECS)
def _maybe_convert_set(extracted_dir, source_csv, target_csv):
print()
if path.exists(target_csv):
print('Found CSV file "%s" - not importing "%s".' %
(target_csv, source_csv))
return
print('No CSV file "%s" - importing "%s"...' % (target_csv, source_csv))
samples = []
with open(source_csv) as source_csv_file:
reader = csv.DictReader(source_csv_file)
for row in reader:
samples.append((os.path.join(extracted_dir,
row['filename']), row['text']))
# Mutable counters for the concurrent embedded routine
counter = get_counter()
num_samples = len(samples)
rows = []
print('Importing mp3 files...')
pool = Pool()
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, processed in enumerate(pool.imap_unordered(one_sample, samples),
start=1):
counter += processed[0]
rows += processed[1]
bar.update(i)
bar.update(num_samples)
pool.close()
pool.join()
print('Writing "%s"...' % target_csv)
with open(target_csv, 'w') as target_csv_file:
writer = csv.DictWriter(target_csv_file, fieldnames=FIELDNAMES)
writer.writeheader()
bar = progressbar.ProgressBar(max_value=len(rows), widgets=SIMPLE_BAR)
for filename, file_size, transcript in bar(rows):
writer.writerow({
'wav_filename': filename,
'wav_filesize': file_size,
'transcript': transcript
})
imported_samples = get_imported_samples(counter)
assert counter['all'] == num_samples
assert len(rows) == imported_samples
print_import_report(counter, SAMPLE_RATE, MAX_SECS)
def one_sample(sample):
""" Take a audio file, and optionally convert it to 16kHz WAV """
orig_filename = 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']
rows = []
# Keep track of how many samples are good vs. problematic
counter = get_counter()
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
return (counter, rows)
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
]
for line in data:
line['path'] = os.path.join(extracted_dir, line['path'])
num_samples = len(data)
rows = []
counter = get_counter()
print("Importing {} wav files...".format(num_samples))
pool = Pool()
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, processed in enumerate(pool.imap_unordered(one_sample, data),
start=1):
counter += processed[0]
rows += processed[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(
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,
))
imported_samples = get_imported_samples(counter)
assert counter['all'] == num_samples
assert len(rows) == imported_samples
print_import_report(counter, SAMPLE_RATE, MAX_SECS)
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('.zip', '_{}.csv'))
if os.path.isfile(target_csv_template):
return
ogg_root_dir = os.path.join(extracted_dir,
ARCHIVE_NAME.replace('.zip', ''))
# Get audiofile path and transcript for each sentence in tsv
samples = []
glob_dir = os.path.join(ogg_root_dir, '**/*.ogg')
for record in glob(glob_dir, recursive=True):
record_file = record.replace(ogg_root_dir + os.path.sep, '')
if record_filter(record_file):
samples.append(
(os.path.join(ogg_root_dir, record_file),
os.path.splitext(os.path.basename(record_file))[0]))
counter = get_counter()
num_samples = len(samples)
rows = []
print("Importing ogg files...")
pool = Pool()
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, processed in enumerate(pool.imap_unordered(one_sample, samples),
start=1):
counter += processed[0]
rows += processed[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 = os.path.join(
ogg_root_dir, item[0].replace('.ogg', '.wav'))
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,
))
imported_samples = get_imported_samples(counter)
assert counter['all'] == num_samples
assert len(rows) == imported_samples
print_import_report(counter, SAMPLE_RATE, MAX_SECS)
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):
if any(map(lambda sk: sk in record, SKIP_LIST)): # pylint: disable=cell-var-from-loop
continue
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))
counter = get_counter()
num_samples = len(samples)
rows = []
print("Importing WAV files...")
pool = Pool()
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, processed in enumerate(pool.imap_unordered(one_sample, samples), start=1):
counter += processed[0]
rows += processed[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=os.path.relpath(wav_filename, extracted_dir),
wav_filesize=os.path.getsize(wav_filename),
transcript=transcript,
))
imported_samples = get_imported_samples(counter)
assert counter['all'] == num_samples
assert len(rows) == imported_samples
print_import_report(counter, SAMPLE_RATE, MAX_SECS)
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('.tar.gz', '_{}.csv'))
if os.path.isfile(target_csv_template):
return
wav_root_dir = os.path.join(extracted_dir)
all_files = [
'transcripts/train/yaounde/fn_text.txt',
'transcripts/train/ca16_conv/transcripts.txt',
'transcripts/train/ca16_read/conditioned.txt',
'transcripts/dev/niger_west_african_fr/transcripts.txt',
'speech/dev/niger_west_african_fr/niger_wav_file_name_transcript.tsv',
'transcripts/devtest/ca16_read/conditioned.txt',
'transcripts/test/ca16/prompts.txt',
]
transcripts = {}
for tr in all_files:
with open(os.path.join(target_dir, ARCHIVE_DIR_NAME, tr),
'r') as tr_source:
for line in tr_source.readlines():
line = line.strip()
if '.tsv' in tr:
sep = ' '
else:
sep = ' '
audio = os.path.basename(line.split(sep)[0])
if not ('.wav' in audio):
if '.tdf' in audio:
audio = audio.replace('.tdf', '.wav')
else:
audio += '.wav'
transcript = ' '.join(line.split(sep)[1:])
transcripts[audio] = transcript
# Get audiofile path and transcript for each sentence in tsv
samples = []
glob_dir = os.path.join(wav_root_dir, '**/*.wav')
for record in glob(glob_dir, recursive=True):
record_file = os.path.basename(record)
if record_file in transcripts:
samples.append((record, transcripts[record_file]))
# Keep track of how many samples are good vs. problematic
counter = get_counter()
num_samples = len(samples)
rows = []
print("Importing WAV files...")
pool = Pool()
bar = progressbar.ProgressBar(max_value=num_samples, widgets=SIMPLE_BAR)
for i, processed in enumerate(pool.imap_unordered(one_sample, samples),
start=1):
counter += processed[0]
rows += processed[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,
))
imported_samples = get_imported_samples(counter)
assert counter['all'] == num_samples
assert len(rows) == imported_samples
print_import_report(counter, SAMPLE_RATE, MAX_SECS)