def train_model(in_filename, mlf_filename, scp_filename, out_filename, scp_prefix_subst=None):
# Read models and create managers
with open(in_filename) as f:
model_dict, gmm_mgr, hmm_mgr = read_acoustic_model(f, log_domain=True)
# Some tricky business to make sure we have an epsilon Hmm to use later
assert hmm_mgr.input_arity_set == hmm_mgr.output_arity_set
assert len(hmm_mgr.input_arity_set) == 1
in_out_arity = hmm_mgr.input_arity_set.pop()
epsilon_index = hmm_mgr.add_epsilon_model(gmm_mgr, in_out_arity, log_domain=True)
# Initialize managers for training
gmm_mgr.set_adaptation_state("INITIALIZING")
hmm_mgr.set_adaptation_state("INITIALIZING")
gmm_mgr.clear_all_accumulators()
utterance_num = [1]
def train_one_utterance(labeled_data_event):
print("Beginning training on utterance %d (%s) (%d frames)" % (utterance_num[0], labeled_data_event.name,
labeled_data_event.data.shape[0]))
sys.stdout.flush()
label_lattice = labels_to_lattice(labeled_data_event.labels)
# Build model lattice
model_lattice = build_model_lattice(label_lattice, model_dict, epsilon_index)
# Build training graph
training_graph = TrainingGraph(model_lattice, hmm_mgr, dict())
# Train on data
if gmm_mgr.get_adaptation_state() != "INITIALIZING":
gmm_mgr.set_adaptation_state("INITIALIZING")
hmm_mgr.set_adaptation_state("INITIALIZING")
training_graph.begin_training()
gmm_mgr.set_adaptation_state("ACCUMULATING")
hmm_mgr.set_adaptation_state("ACCUMULATING")
training_graph.train_one_sequence(labeled_data_event.data)
training_graph.end_training()
print("Finished training on utterance %d (%s)" % (utterance_num[0], labeled_data_event.name))
sys.stdout.flush()
utterance_num[0] += 1
# Do processing:
with contextlib.nested(open(mlf_filename), open(scp_filename)) as (mlf_file, scp_file):
process_htk_files(mlf_file, scp_file, train_one_utterance, scp_prefix_subst)
# Finalize training
gmm_mgr.set_adaptation_state("APPLYING")
hmm_mgr.set_adaptation_state("APPLYING")
gmm_mgr.apply_all_accumulators()
hmm_mgr.apply_all_accumulators()
gmm_mgr.set_adaptation_state("NOT_ADAPTING")
hmm_mgr.set_adaptation_state("NOT_ADAPTING")
# Write out models
with open(out_filename, "w") as f:
write_acoustic_model(model_dict, gmm_mgr, hmm_mgr, f)
def train_model(in_filename,
mlf_filename,
scp_filename,
out_filename,
scp_prefix_subst=None):
# Read models and create managers
with open(in_filename) as f:
model_dict, gmm_mgr, hmm_mgr = read_acoustic_model(f, log_domain=True)
# Some tricky business to make sure we have an epsilon Hmm to use later
assert hmm_mgr.input_arity_set == hmm_mgr.output_arity_set
assert len(hmm_mgr.input_arity_set) == 1
in_out_arity = hmm_mgr.input_arity_set.pop()
epsilon_index = hmm_mgr.add_epsilon_model(gmm_mgr,
in_out_arity,
log_domain=True)
# Initialize managers for training
gmm_mgr.set_adaptation_state("INITIALIZING")
hmm_mgr.set_adaptation_state("INITIALIZING")
gmm_mgr.clear_all_accumulators()
utterance_num = [1]
def train_one_utterance(labeled_data_event):
print("Beginning training on utterance %d (%s) (%d frames)" %
(utterance_num[0], labeled_data_event.name,
labeled_data_event.data.shape[0]))
sys.stdout.flush()
label_lattice = labels_to_lattice(labeled_data_event.labels)
# Build model lattice
model_lattice = build_model_lattice(label_lattice, model_dict,
epsilon_index)
# Build training graph
training_graph = TrainingGraph(model_lattice, hmm_mgr, dict())
# Train on data
if gmm_mgr.get_adaptation_state() != "INITIALIZING":
gmm_mgr.set_adaptation_state("INITIALIZING")
hmm_mgr.set_adaptation_state("INITIALIZING")
training_graph.begin_training()
gmm_mgr.set_adaptation_state("ACCUMULATING")
hmm_mgr.set_adaptation_state("ACCUMULATING")
training_graph.train_one_sequence(labeled_data_event.data)
training_graph.end_training()
print("Finished training on utterance %d (%s)" %
(utterance_num[0], labeled_data_event.name))
sys.stdout.flush()
utterance_num[0] += 1
# Do processing:
with contextlib.nested(open(mlf_filename),
open(scp_filename)) as (mlf_file, scp_file):
process_htk_files(mlf_file, scp_file, train_one_utterance,
scp_prefix_subst)
# Finalize training
gmm_mgr.set_adaptation_state("APPLYING")
hmm_mgr.set_adaptation_state("APPLYING")
gmm_mgr.apply_all_accumulators()
hmm_mgr.apply_all_accumulators()
gmm_mgr.set_adaptation_state("NOT_ADAPTING")
hmm_mgr.set_adaptation_state("NOT_ADAPTING")
# Write out models
with open(out_filename, "w") as f:
write_acoustic_model(model_dict, gmm_mgr, hmm_mgr, f)
def htkmmf2native(htk_file_name, native_file_name):
with open(htk_file_name) as f_in:
models, hmm_mgr, gmm_mgr = read_htk_mmf_file(f_in)
with open(native_file_name, 'wb') as f_out:
write_acoustic_model(models, gmm_mgr, hmm_mgr, f_out)
def htkmmf2native(htk_file_name, native_file_name):
with open(htk_file_name) as f_in:
models, hmm_mgr, gmm_mgr = read_htk_mmf_file(f_in)
with open(native_file_name, 'wb') as f_out:
write_acoustic_model(models, gmm_mgr, hmm_mgr, f_out)
