def compute_raw_scores(audio_paths: List[str]) -> Iterable[List[AsrOutput]]:
from deepspeech_training.train import create_model # pylint: disable=cyclic-import,import-outside-toplevel
from deepspeech_training.util.checkpoints import load_graph_for_evaluation
initialize_globals()
with tf.Session(config=Config.session_config) as session:
tf.train.get_or_create_global_step()
for idx, audio_path in enumerate(audio_paths):
with AudioFile(audio_path, as_path=True) as wav_path:
data_set = split_audio_file(wav_path,
batch_size=FLAGS.batch_size,
aggressiveness=1,
outlier_duration_ms=FLAGS.outlier_duration_ms,
outlier_batch_size=FLAGS.outlier_batch_size)
iterator = tf.data.Iterator.from_structure(data_set.output_types, data_set.output_shapes,
output_classes=data_set.output_classes)
batch_time_start, batch_time_end, batch_x, batch_x_len = iterator.get_next()
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x, seq_length=batch_x_len, dropout=no_dropout)
transposed = tf.nn.softmax(tf.transpose(logits, [1, 0, 2]))
session.run(iterator.make_initializer(data_set))
if idx == 0:
load_graph_for_evaluation(session)
output_list = []
while True:
try:
starts, ends, batch_logits, batch_lengths = session.run([batch_time_start, batch_time_end, transposed, batch_x_len])
except tf.errors.OutOfRangeError:
break
for start, end, logits, length in zip(starts, ends, batch_logits, batch_lengths):
output_list.append(AsrOutput(start, end, logits[:length]))
tf.get_variable_scope().reuse_variables()
yield output_list
def transcribe_file(audio_path, tlog_path):
from deepspeech_training.train import create_model # pylint: disable=cyclic-import,import-outside-toplevel
from deepspeech_training.util.checkpoints import load_graph_for_evaluation
initialize_globals()
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.scorer_path,
Config.alphabet)
try:
num_processes = cpu_count()
except NotImplementedError:
num_processes = 1
with AudioFile(audio_path, as_path=True) as wav_path:
data_set = split_audio_file(
wav_path,
batch_size=FLAGS.batch_size,
aggressiveness=FLAGS.vad_aggressiveness,
outlier_duration_ms=FLAGS.outlier_duration_ms,
outlier_batch_size=FLAGS.outlier_batch_size)
iterator = tf.data.Iterator.from_structure(
data_set.output_types,
data_set.output_shapes,
output_classes=data_set.output_classes)
batch_time_start, batch_time_end, batch_x, batch_x_len = iterator.get_next(
)
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
seq_length=batch_x_len,
dropout=no_dropout)
transposed = tf.nn.softmax(tf.transpose(logits, [1, 0, 2]))
tf.train.get_or_create_global_step()
with tf.Session(config=Config.session_config) as session:
load_graph_for_evaluation(session)
session.run(iterator.make_initializer(data_set))
transcripts = []
while True:
try:
starts, ends, batch_logits, batch_lengths = \
session.run([batch_time_start, batch_time_end, transposed, batch_x_len])
except tf.errors.OutOfRangeError:
break
decoded = ctc_beam_search_decoder_batch(
batch_logits,
batch_lengths,
Config.alphabet,
FLAGS.beam_width,
num_processes=num_processes,
scorer=scorer)
decoded = list(d[0][1] for d in decoded)
transcripts.extend(zip(starts, ends, decoded))
transcripts.sort(key=lambda t: t[0])
transcripts = [{
'start': int(start),
'end': int(end),
'transcript': transcript
} for start, end, transcript in transcripts]
with open(tlog_path, 'w') as tlog_file:
json.dump(transcripts, tlog_file, default=float)