def get_segments(task: CTCSegmentationTask):
"""Obtain segments for given utterance texts and CTC log posteriors.
Args:
task: CTCSegmentationTask object that contains ground truth and
CTC posterior probabilities.
Returns:
result: Dictionary with alignments. Combine this with the task
object to obtain a human-readable segments representation.
"""
assert check_argument_types()
assert task.config is not None
config = task.config
lpz = task.lpz
ground_truth_mat = task.ground_truth_mat
utt_begin_indices = task.utt_begin_indices
text = task.text
# Align using CTC segmentation
timings, char_probs, state_list = ctc_segmentation(
config, lpz, ground_truth_mat)
# Obtain list of utterances with time intervals and confidence score
segments = determine_utterance_segments(config, utt_begin_indices,
char_probs, timings, text)
# Store results
result = {
"name": task.name,
"timings": timings,
"char_probs": char_probs,
"state_list": state_list,
"segments": segments,
"done": True,
}
return result
def test_determine_utterance_segments():
"""Test the generation of segments from aligned utterances.
This is a function that is used after a completed CTC segmentation.
Results are checked and compared with test vectors.
"""
config = CtcSegmentationParameters()
frame_duration_ms = 1000
config.index_duration = frame_duration_ms / 1000.0
config.score_min_mean_over_L = 2
utt_begin_indices = [1, 4, 9]
text = ["catzz#\n", "dogs!!\n"]
char_probs = np.array([-0.5] * 10)
timings = np.array(list(range(10))) + 0.5
segments = determine_utterance_segments(config, utt_begin_indices,
char_probs, timings, text)
correct_segments = [(2.0, 4.0, -0.5), (5.0, 9.0, -0.5)]
for i in [0, 1]:
for j in [0, 1, 2]:
assert segments[i][j] == correct_segments[i][j]
def ctc_align(args, device):
"""ESPnet-specific interface for CTC segmentation.
Parses configuration, infers the CTC posterior probabilities,
and then aligns start and end of utterances using CTC segmentation.
Results are written to the output file given in the args.
:param args: given configuration
:param device: for inference; one of ['cuda', 'cpu']
:return: 0 on success
"""
model, train_args = load_trained_model(args.model)
assert isinstance(model, ASRInterface)
load_inputs_and_targets = LoadInputsAndTargets(
mode="asr",
load_output=True,
sort_in_input_length=False,
preprocess_conf=train_args.preprocess_conf
if args.preprocess_conf is None
else args.preprocess_conf,
preprocess_args={"train": False},
)
logging.info(f"Decoding device={device}")
# Warn for nets with high memory consumption on long audio files
if hasattr(model, "enc"):
encoder_module = model.enc.__class__.__module__
elif hasattr(model, "encoder"):
encoder_module = model.encoder.__class__.__module__
else:
encoder_module = "Unknown"
logging.info(f"Encoder module: {encoder_module}")
logging.info(f"CTC module: {model.ctc.__class__.__module__}")
if "rnn" not in encoder_module:
logging.warning("No BLSTM model detected; memory consumption may be high.")
model.to(device=device).eval()
# read audio and text json data
with open(args.data_json, "rb") as f:
js = json.load(f)["utts"]
with open(args.utt_text, "r", encoding="utf-8") as f:
lines = f.readlines()
i = 0
text = {}
segment_names = {}
for name in js.keys():
text_per_audio = []
segment_names_per_audio = []
while i < len(lines) and lines[i].startswith(name):
text_per_audio.append(lines[i][lines[i].find(" ") + 1 :])
segment_names_per_audio.append(lines[i][: lines[i].find(" ")])
i += 1
text[name] = text_per_audio
segment_names[name] = segment_names_per_audio
# apply configuration
config = CtcSegmentationParameters()
subsampling_factor = 1
frame_duration_ms = 10
if args.subsampling_factor is not None:
subsampling_factor = args.subsampling_factor
if args.frame_duration is not None:
frame_duration_ms = args.frame_duration
# Backwards compatibility to ctc_segmentation <= 1.5.3
if hasattr(config, "index_duration"):
config.index_duration = frame_duration_ms * subsampling_factor / 1000
else:
config.subsampling_factor = subsampling_factor
config.frame_duration_ms = frame_duration_ms
if args.min_window_size is not None:
config.min_window_size = args.min_window_size
if args.max_window_size is not None:
config.max_window_size = args.max_window_size
config.char_list = train_args.char_list
if args.use_dict_blank is not None:
logging.warning(
"The option --use-dict-blank is deprecated. If needed,"
" use --set-blank instead."
)
if args.set_blank is not None:
config.blank = args.set_blank
if args.replace_spaces_with_blanks is not None:
if args.replace_spaces_with_blanks:
config.replace_spaces_with_blanks = True
else:
config.replace_spaces_with_blanks = False
if args.gratis_blank:
config.blank_transition_cost_zero = True
if config.blank_transition_cost_zero and args.replace_spaces_with_blanks:
logging.error(
"Blanks are inserted between words, and also the transition cost of blank"
" is zero. This configuration may lead to misalignments!"
)
if args.scoring_length is not None:
config.score_min_mean_over_L = args.scoring_length
logging.info(f"Frame timings: {frame_duration_ms}ms * {subsampling_factor}")
# Iterate over audio files to decode and align
for idx, name in enumerate(js.keys(), 1):
logging.info("(%d/%d) Aligning " + name, idx, len(js.keys()))
batch = [(name, js[name])]
feat, label = load_inputs_and_targets(batch)
feat = feat[0]
with torch.no_grad():
# Encode input frames
enc_output = model.encode(torch.as_tensor(feat).to(device)).unsqueeze(0)
# Apply ctc layer to obtain log character probabilities
lpz = model.ctc.log_softmax(enc_output)[0].cpu().numpy()
# Prepare the text for aligning
ground_truth_mat, utt_begin_indices = prepare_text(config, text[name])
# Align using CTC segmentation
timings, char_probs, state_list = ctc_segmentation(
config, lpz, ground_truth_mat
)
logging.debug(f"state_list = {state_list}")
# Obtain list of utterances with time intervals and confidence score
segments = determine_utterance_segments(
config, utt_begin_indices, char_probs, timings, text[name]
)
# Write to "segments" file
for i, boundary in enumerate(segments):
utt_segment = (
f"{segment_names[name][i]} {name} {boundary[0]:.2f}"
f" {boundary[1]:.2f} {boundary[2]:.9f}\n"
)
args.output.write(utt_segment)
return 0
def get_segments(
log_probs: np.ndarray,
path_wav: Union[PosixPath, str],
transcript_file: Union[PosixPath, str],
output_file: str,
vocabulary: List[str],
window_size: int = 8000,
frame_duration_ms: int = 20,
) -> None:
"""
Segments the audio into segments and saves segments timings to a file
Args:
log_probs: Log probabilities for the original audio from an ASR model, shape T * |vocabulary|.
values for blank should be at position 0
path_wav: path to the audio .wav file
transcript_file: path to
output_file: path to the file to save timings for segments
vocabulary: vocabulary used to train the ASR model, note blank is at position 0
window_size: the length of each utterance (in terms of frames of the CTC outputs) fits into that window.
frame_duration_ms: frame duration in ms
"""
config = cs.CtcSegmentationParameters()
config.char_list = vocabulary
config.min_window_size = window_size
config.frame_duration_ms = frame_duration_ms
config.blank = config.space
config.subsampling_factor = 2
with open(transcript_file, "r") as f:
text = f.readlines()
text = [t.strip() for t in text if t.strip()]
# add corresponding original text without pre-processing
transcript_file_no_preprocessing = transcript_file.replace(
'.txt', '_with_punct.txt')
if not os.path.exists(transcript_file_no_preprocessing):
raise ValueError(f'{transcript_file_no_preprocessing} not found.')
with open(transcript_file_no_preprocessing, "r") as f:
text_no_preprocessing = f.readlines()
text_no_preprocessing = [
t.strip() for t in text_no_preprocessing if t.strip()
]
if len(text_no_preprocessing) != len(text):
raise ValueError(
f'{transcript_file} and {transcript_file_no_preprocessing} do not match'
)
ground_truth_mat, utt_begin_indices = cs.prepare_text(config, text)
logging.debug(f"Syncing {transcript_file}")
logging.debug(
f"Audio length {os.path.basename(path_wav)}: {log_probs.shape[0]}. "
f"Text length {os.path.basename(transcript_file)}: {len(ground_truth_mat)}"
)
timings, char_probs, char_list = cs.ctc_segmentation(
config, log_probs, ground_truth_mat)
segments = cs.determine_utterance_segments(config, utt_begin_indices,
char_probs, timings, text)
write_output(output_file, path_wav, segments, text, text_no_preprocessing)
pred_phon = tf.nn.log_softmax(pred_phon)
iphon_tar = model.text_pipeline.tokenizer.decode(
phonemes[j][:phon_len[j]])
iphon_tar = iphon_tar.split()
char_list = [''] + list(
model.text_pipeline.tokenizer.idx_to_token.values())
config = CtcSegmentationParameters(char_list=char_list)
config.index_duration = 0.0115545
text = [phonemes[j][:phon_len[j]]]
ground_truth_mat, utt_begin_indices = prepare_token_list(config, text)
timings, char_probs, state_list = ctc_segmentation(
config, pred_phon.numpy(), ground_truth_mat)
utt_begin_indices = list(range(2, len(timings)))
segments = determine_utterance_segments(config, utt_begin_indices,
char_probs, timings, text[0])
tg = tgt.core.TextGrid('haa')
tier = tgt.core.IntervalTier(name='phonemes')
if (segments[0][-1] < -0.001):
segments[0] = (0, segments[0][1], segments[0][2])
else:
itv = tgt.core.Interval(0, segments[0][0], text='sp')
tier.add_interval(itv)
if (segments[-1][-1] < -0.001):
segments[-1] = (segments[-1][0], segments[-1][1] + 0.15,
segments[-1][2])
if (segments[-1][1] > mel_len[j] * config.index_duration):
pass
def prediction_one_song(model,
audio_filename,
transcript,
lp_dir='tmp',
lp_ext='_logprobs.py',
word_dir='../lyrics',
word_ext='.words.txt',
prediction_dir='metadata',
prediction_ext='_align.csv'):
'''
model - nemo model object
lp_dir - path with logprobabilities
audio_filename - file name of audio song that is being proceesed
'''
basename = audio_filename[:-4] #crop extension (mp3 or wav)
alphabet = [t for t in model.cfg['labels']
] + ['%'] # converting to list and adding blank character.
# adapted example from here:
# https://github.com/lumaku/ctc-segmentation
config = ctc.CtcSegmentationParameters()
config.frame_duration_ms = 20 #frame duration is the window of the predictions (i.e. logprobs prediction window)
config.blank = len(
alphabet
) - 1 #index for character that is intended for 'blank' - in our case, we specify the last character in alphabet.
logprobs_filenames = glob.glob(
os.path.join(lp_dir, basename + '*' + lp_ext))
logprobs_filenames.sort()
logprobs_list = []
for f in logprobs_filenames:
logprobs_list.append(np.load(f))
logprobs = logprobs_list[0]
for i in range(1, len(logprobs_list)):
logprobs = np.concatenate((logprobs, logprobs_list[i]))
print('Prepare Text.', flush=True)
ground_truth_mat, utt_begin_indices = ctc.prepare_text(
config, transcript, alphabet)
print('Segmentation.', flush=True)
timings, char_probs, state_list = ctc.ctc_segmentation(
config, logprobs, ground_truth_mat)
print('Get time intervals.', flush=True)
# Obtain list of utterances with time intervals and confidence score
segments = ctc.determine_utterance_segments(config, utt_begin_indices,
char_probs, timings,
transcript)
tend = time.time()
pred_fname = prediction_dir + '/' + basename + '_align.csv' #jamendolyrics convention
fname = open(pred_fname, 'w')
offset = 0 #offset is used to compensate for the re.search command which only finds the first
# match in the string. so the transcript is iteratively cropped to ensure that the
# previous words in the transcript are not found again.
for i in transcript.split():
#
# taking each word, and writing out the word timings from segments variable
#
# re.search performs regular expression operations.
# .format inserts characters into {}.
# r'<string>' is considered a raw string.
# char.start() gives you the start index of the starting character of the word (i) in transcript string
# char.end() gives you the last index of the ending character** of the word (i) in transcript string
# **the ending character is offset by one for the regex command, so a -1 is required to get the right
# index
char = re.search(r'{}'.format(i), transcript[offset:])
# segments[index of character][start time of char=0]
onset = segments[char.start() + offset][0]
# segments[index of character][end time of char=1]
term = segments[char.end() - 1 + offset][1]
offset += char.end()
fname.write(str(onset) + ',' + str(term) + '\n')
fname.close()
def validate_asr_with_alignment(asr_model,val_ds,num_to_validate):
val_set = []
with open(val_ds) as F:
for line in F:
val = json.loads(line)
val_set.append(val)
val_files = [t["audio_filepath"] for t in val_set[0:num_to_validate]]
val_text = [t["text"] for t in val_set[0:num_to_validate]]
test_cfg = asr_model.cfg['validation_ds']
test_cfg['manifest_filepath'] = val_ds
asr_model.setup_test_data(test_cfg) #TODO: what is this doing?
calc_wer(asr_model)
asr_model.preprocessor._sample_rate = test_cfg['sample_rate']
print("batch size: ", test_cfg['batch_size'],
"preprocessor sample_rate: ", asr_model.preprocessor._sample_rate)
logprobs_list = asr_model.transcribe(val_files, batch_size=test_cfg['batch_size'], logprobs=True)
nlogprobs = len(logprobs_list)
alphabet = [t for t in asr_model.cfg['labels']] + ['%'] # converting to list and adding blank character.
# adapted example from here:
# https://github.com/lumaku/ctc-segmentation
config = CtcSegmentationParameters()
config.frame_duration_ms = 20 #frame duration is the window of the predictions (i.e. logprobs prediction window)
config.blank = len(alphabet)-1 #index for character that is intended for 'blank' - in our case, we specify the last character in alphabet.
for ii in range(nlogprobs):
transcript = val_text[ii]
ground_truth_mat, utt_begin_indices = prepare_text(config,transcript,alphabet)
timings, char_probs, state_list = ctc_segmentation(config,logprobs_list[ii].cpu().numpy(),ground_truth_mat)
# Obtain list of utterances with time intervals and confidence score
segments = determine_utterance_segments(config, utt_begin_indices, char_probs, timings, transcript)
quartznet_transcript = asr_model.transcribe([val_files[ii]])
print('Ground Truth Transcript:',transcript)
print('Quartznet Transcript:',quartznet_transcript[0])
print('CTC Segmentation Dense Sequnce:\n',''.join(state_list))
#save onset per word.
print('Saving timing prediction.')
fname = open(val_files[ii][:-4]+'_align.csv','w') #jamendolyrics convention
for i in transcript.split():
# re.search performs regular expression operations.
# .format inserts characters into {}.
# r'<string>' is considered a raw string.
# char.start() gives you the start index of the starting character of the word (i) in transcript string
# char.end() gives you the last index of the ending character** of the word (i) in transcript string
# **the ending character is offset by one for the regex command, so a -1 is required to get the right
# index
char = re.search(r'\b({})\b'.format(i),transcript)
# segments[index of character][start time of char=0]
onset = segments[char.start()][0]
# segments[index of character][end time of char=1]
term = segments[char.end()-1][1]
fname.write(str(onset)+','+str(term)+'\n')
fname.close()
def ctc_align(args, device):
"""ESPnet-specific interface for CTC segmentation.
Parses configuration, infers the CTC posterior probabilities,
and then aligns start and end of utterances using CTC segmentation.
Results are written to the output file given in the args.
:param args: given configuration
:param device: for inference; one of ['cuda', 'cpu']
:return: 0 on success
"""
model, train_args = load_trained_model(args.model)
assert isinstance(model, ASRInterface)
load_inputs_and_targets = LoadInputsAndTargets(
mode="asr",
load_output=True,
sort_in_input_length=False,
preprocess_conf=train_args.preprocess_conf
if args.preprocess_conf is None else args.preprocess_conf,
preprocess_args={"train": False},
)
logging.info(f"Decoding device={device}")
model.to(device=device).eval()
# read audio and text json data
with open(args.data_json, "rb") as f:
js = json.load(f)["utts"]
with open(args.utt_text, "r") as f:
lines = f.readlines()
i = 0
text = {}
segment_names = {}
for name in js.keys():
text_per_audio = []
segment_names_per_audio = []
while i < len(lines) and lines[i].startswith(name):
text_per_audio.append(lines[i][lines[i].find(" ") + 1:])
segment_names_per_audio.append(lines[i][:lines[i].find(" ")])
i += 1
text[name] = text_per_audio
segment_names[name] = segment_names_per_audio
# apply configuration
config = CtcSegmentationParameters()
if args.subsampling_factor is not None:
config.subsampling_factor = args.subsampling_factor
if args.frame_duration is not None:
config.frame_duration_ms = args.frame_duration
if args.min_window_size is not None:
config.min_window_size = args.min_window_size
if args.max_window_size is not None:
config.max_window_size = args.max_window_size
char_list = train_args.char_list
if args.use_dict_blank:
config.blank = char_list[0]
logging.debug(
f"Frame timings: {config.frame_duration_ms}ms * {config.subsampling_factor}"
)
# Iterate over audio files to decode and align
for idx, name in enumerate(js.keys(), 1):
logging.info("(%d/%d) Aligning " + name, idx, len(js.keys()))
batch = [(name, js[name])]
feat, label = load_inputs_and_targets(batch)
feat = feat[0]
with torch.no_grad():
# Encode input frames
enc_output = model.encode(
torch.as_tensor(feat).to(device)).unsqueeze(0)
# Apply ctc layer to obtain log character probabilities
lpz = model.ctc.log_softmax(enc_output)[0].cpu().numpy()
# Prepare the text for aligning
ground_truth_mat, utt_begin_indices = prepare_text(
config, text[name], char_list)
# Align using CTC segmentation
timings, char_probs, state_list = ctc_segmentation(
config, lpz, ground_truth_mat)
# Obtain list of utterances with time intervals and confidence score
segments = determine_utterance_segments(config, utt_begin_indices,
char_probs, timings,
text[name])
# Write to "segments" file
for i, boundary in enumerate(segments):
utt_segment = (f"{segment_names[name][i]} {name} {boundary[0]:.2f}"
f" {boundary[1]:.2f} {boundary[2]:.9f}\n")
args.output.write(utt_segment)
return 0
