def fstprintall(
in_fst: fst.Fst,
out_file: Optional[TextIO] = None,
exclude_meta: bool = True,
eps: str = "<eps>",
) -> List[List[str]]:
sentences = []
output_symbols = in_fst.output_symbols()
out_eps = output_symbols.find(eps)
zero_weight = fst.Weight.Zero(in_fst.weight_type())
state_queue: Deque[Tuple[int, List[str]]] = deque()
state_queue.append((in_fst.start(), []))
while len(state_queue) > 0:
state, sentence = state_queue.popleft()
if in_fst.final(state) != zero_weight:
if out_file:
print(" ".join(sentence), file=out_file)
else:
sentences.append(sentence)
for arc in in_fst.arcs(state):
arc_sentence = list(sentence)
if arc.olabel != out_eps:
out_symbol = output_symbols.find(arc.olabel).decode()
if exclude_meta and out_symbol.startswith("__"):
pass # skip __label__, etc.
else:
arc_sentence.append(out_symbol)
state_queue.append((arc.nextstate, arc_sentence))
return sentences
def fstprintall(
in_fst: fst.Fst,
out_file: Optional[TextIO] = None,
exclude_meta: bool = True,
state: Optional[int] = None,
path: Optional[List[fst.Arc]] = None,
zero_weight: Optional[fst.Weight] = None,
eps: int = 0,
) -> List[List[str]]:
sentences = []
path = path or []
state = state or in_fst.start()
zero_weight = zero_weight or fst.Weight.Zero(in_fst.weight_type())
for arc in in_fst.arcs(state):
path.append(arc)
if in_fst.final(arc.nextstate) != zero_weight:
# Final state
out_syms = in_fst.output_symbols()
sentence = []
for p_arc in path:
if p_arc.olabel != eps:
osym = out_syms.find(p_arc.olabel).decode()
if exclude_meta and osym.startswith("__"):
continue # skip __label__, etc.
if out_file:
print(osym, "", end="", file=out_file)
else:
sentence.append(osym)
if out_file:
print("", file=out_file)
else:
sentences.append(sentence)
else:
# Non-final state
sentences.extend(
fstprintall(
in_fst,
out_file=out_file,
state=arc.nextstate,
path=path,
zero_weight=zero_weight,
eps=eps,
exclude_meta=exclude_meta,
))
path.pop()
return sentences
def longest_path(the_fst: fst.Fst, eps: str = "<eps>") -> fst.Fst:
output_symbols = the_fst.output_symbols()
out_eps = output_symbols.find(eps)
visited_states: Set[int] = set()
best_path: List[int] = []
state_queue: Deque[Tuple[int, List[int]]] = deque()
state_queue.append((the_fst.start(), []))
# Determine longest path
while len(state_queue) > 0:
state, path = state_queue.popleft()
if state in visited_states:
continue
visited_states.add(state)
if len(path) > len(best_path):
best_path = path
for arc in the_fst.arcs(state):
next_path = list(path)
next_path.append(arc.olabel)
state_queue.append((arc.nextstate, next_path))
# Create FST with longest path
path_fst = fst.Fst()
input_symbols = fst.SymbolTable()
input_symbols.add_symbol(eps)
path_fst.set_output_symbols(output_symbols)
weight_one = fst.Weight.One(path_fst.weight_type())
state = path_fst.add_state()
path_fst.set_start(state)
for olabel in best_path:
osym = output_symbols.find(olabel).decode()
next_state = path_fst.add_state()
path_fst.add_arc(
state,
fst.Arc(input_symbols.add_symbol(osym), olabel, weight_one,
next_state),
)
state = next_state
path_fst.set_final(state)
path_fst.set_input_symbols(input_symbols)
return path_fst
def fst_to_graph(the_fst: fst.Fst) -> nx.MultiDiGraph:
"""Converts a finite state transducer to a directed graph."""
zero_weight = fst.Weight.Zero(the_fst.weight_type())
in_symbols = the_fst.input_symbols()
out_symbols = the_fst.output_symbols()
g = nx.MultiDiGraph()
# Add nodes
for state in the_fst.states():
# Mark final states
is_final = the_fst.final(state) != zero_weight
g.add_node(state, final=is_final, start=False)
# Add edges
for arc in the_fst.arcs(state):
in_label = in_symbols.find(arc.ilabel).decode()
out_label = out_symbols.find(arc.olabel).decode()
g.add_edge(state,
arc.nextstate,
in_label=in_label,
out_label=out_label)
# Mark start state
g.add_node(the_fst.start(), start=True)
return g
def make_slot_fsts(intent_fst: fst.Fst) -> Dict[str, Dict[str, fst.Fst]]:
out_symbols = intent_fst.output_symbols()
intent_to_slots: Dict[str, Dict[str, fst.Fst]] = {}
start_state = intent_fst.start()
for intent_arc in intent_fst.arcs(start_state):
# Extract intent name from output label
intent_label = out_symbols.find(intent_arc.olabel).decode()
assert intent_label.startswith("__label__"), intent_label
intent_name = intent_label[9:]
# Create mapping from slot (tag) name to acceptor FST
slot_to_fst: Dict[str, fst.Fst] = {}
intent_to_slots[intent_name] = slot_to_fst
_make_slot_fst(intent_arc.nextstate, intent_fst, slot_to_fst)
return intent_to_slots
def _make_slot_fst(state: int, intent_fst: fst.Fst,
slot_to_fst: Dict[str, fst.Fst]):
out_symbols = intent_fst.output_symbols()
one_weight = fst.Weight.One(intent_fst.weight_type())
for arc in intent_fst.arcs(state):
label = out_symbols.find(arc.olabel).decode()
if label.startswith("__begin__"):
slot_name = label[9:]
# Big assumption here that each instance of a slot (e.g., location)
# will produce the same FST, and therefore doesn't need to be
# processed again.
if slot_name in slot_to_fst:
continue # skip duplicate slots
end_label = f"__end__{slot_name}"
# Create new FST
slot_fst = fst.Fst()
slot_fst.set_input_symbols(intent_fst.input_symbols())
slot_fst.set_output_symbols(intent_fst.output_symbols())
start_state = slot_fst.add_state()
slot_fst.set_start(start_state)
q = [arc.nextstate]
state_map = {arc.nextstate: start_state}
# Copy states/arcs from intent FST until __end__ is found
while len(q) > 0:
q_state = q.pop()
for q_arc in intent_fst.arcs(q_state):
slot_arc_label = out_symbols.find(q_arc.olabel).decode()
if slot_arc_label != end_label:
if not q_arc.nextstate in state_map:
state_map[q_arc.nextstate] = slot_fst.add_state()
# Create arc
slot_fst.add_arc(
state_map[q_state],
fst.Arc(
q_arc.ilabel,
q_arc.olabel,
one_weight,
state_map[q_arc.nextstate],
),
)
# Continue copy
q.append(q_arc.nextstate)
else:
# Mark previous state as final
slot_fst.set_final(state_map[q_state])
slot_to_fst[slot_name] = minimize_fst(slot_fst)
# Recurse
_make_slot_fst(arc.nextstate, intent_fst, slot_to_fst)
def _all_valid_strings(td: fst.Fst) -> List[Tuple[List[int], float]]:
"""Return an enumeration of the emission language. Essentially the equivalent of
fstprint, but not handling the de-interning of strings.
The weight returned is not the weight of the whole sequence, but the weight
of the final state in the sequence as a final state.
Does not check for duplicate emissions or cycles in the transducer.
:param td: transducer, the emission language of which to enumerate
:returns: a list of (interned emission symbols, weight) tuples
"""
if td.start() == -1:
return []
stack = [(td.start(), [])]
complete_emissions = []
while stack:
state, output = stack.pop()
final_weight = float(td.final(state))
if np.isfinite(final_weight):
complete_emissions.append((output, final_weight))
stack += [(a.nextstate, output + [a.olabel]) for a in td.arcs(state)]
return complete_emissions
def linear_fst(
elements: List[str],
automata_op: fst.Fst,
keep_isymbols: bool = True,
**kwargs: Mapping[Any, Any],
) -> fst.Fst:
"""Produce a linear automata."""
assert len(elements) > 0, "No elements"
compiler = fst.Compiler(
isymbols=automata_op.input_symbols().copy(),
acceptor=keep_isymbols,
keep_isymbols=keep_isymbols,
**kwargs,
)
num_elements = 0
for i, el in enumerate(elements):
print("{} {} {}".format(i, i + 1, el), file=compiler)
num_elements += 1
print(str(num_elements), file=compiler)
return compiler.compile()
def replace_and_patch(
outer_fst: fst.Fst,
outer_start_state: int,
outer_final_state: int,
inner_fst: fst.Fst,
label_sym: int,
eps: int = 0,
) -> None:
"""Copies an inner FST into an outer FST, creating states and mapping symbols.
Creates arcs from outer start/final states to inner start/final states."""
in_symbols = outer_fst.input_symbols()
out_symbols = outer_fst.output_symbols()
inner_zero = fst.Weight.Zero(inner_fst.weight_type())
outer_one = fst.Weight.One(outer_fst.weight_type())
state_map = {}
in_symbol_map = {}
out_symbol_map = {}
for i in range(inner_fst.output_symbols().num_symbols()):
sym_str = inner_fst.output_symbols().find(i).decode()
out_symbol_map[i] = out_symbols.find(sym_str)
for i in range(inner_fst.input_symbols().num_symbols()):
sym_str = inner_fst.input_symbols().find(i).decode()
in_symbol_map[i] = in_symbols.find(sym_str)
# Create states in outer FST
for inner_state in inner_fst.states():
state_map[inner_state] = outer_fst.add_state()
# Create arcs in outer FST
for inner_state in inner_fst.states():
if inner_state == inner_fst.start():
outer_fst.add_arc(
outer_start_state,
fst.Arc(eps, label_sym, outer_one, state_map[inner_state]),
)
for inner_arc in inner_fst.arcs(inner_state):
outer_fst.add_arc(
state_map[inner_state],
fst.Arc(
in_symbol_map[inner_arc.ilabel],
out_symbol_map[inner_arc.olabel],
outer_one,
state_map[inner_arc.nextstate],
),
)
if inner_fst.final(inner_arc.nextstate) != inner_zero:
outer_fst.add_arc(
state_map[inner_arc.nextstate],
fst.Arc(eps, eps, outer_one, outer_final_state),
)
def write_dictionary(self, intent_fst: fst.Fst) -> Set[str]:
"""Writes all required words to a CMU dictionary.
Unknown words have their pronunciations guessed and written to a separate dictionary.
Fails if any unknown words are found."""
start_time = time.time()
words_needed: Set[str] = set()
# Gather all words needed
in_symbols = intent_fst.input_symbols()
for i in range(in_symbols.num_symbols()):
word = in_symbols.find(i).decode()
if word.startswith("__") or word.startswith("<"):
continue # skip metadata
# Dictionary uses upper-case letters
if self.dictionary_upper:
word = word.upper()
else:
word = word.lower()
words_needed.add(word)
# Load base and custom dictionaries
base_dictionary_path = self.profile.read_path(
self.profile.get(
f"speech_to_text.{self.system}.base_dictionary", "base_dictionary.txt"
)
)
custom_path = self.profile.read_path(
self.profile.get(
f"speech_to_text.{self.system}.custom_words", "custom_words.txt"
)
)
word_dict: Dict[str, List[str]] = {}
for word_dict_path in [base_dictionary_path, custom_path]:
if os.path.exists(word_dict_path):
self._logger.debug(f"Loading dictionary from {word_dict_path}")
with open(word_dict_path, "r") as dictionary_file:
read_dict(dictionary_file, word_dict)
# Add words from wake word if using pocketsphinx
if self.profile.get("wake.system") == "pocketsphinx":
wake_keyphrase = self.profile.get("wake.pocketsphinx.keyphrase", "")
if len(wake_keyphrase) > 0:
self._logger.debug(f"Adding words from keyphrase: {wake_keyphrase}")
_, wake_tokens = sanitize_sentence(
wake_keyphrase,
self.dictionary_casing,
self.replace_patterns,
self.split_pattern,
)
for word in wake_tokens:
# Dictionary uses upper-case letters
if self.dictionary_upper:
word = word.upper()
else:
word = word.lower()
words_needed.add(word)
# Determine if we need to include the entire base dictionary
mix_weight = float(
self.profile.get(f"speech_to_text.{self.system}.mix_weight", 0)
)
if mix_weight > 0:
self._logger.debug(
"Including base dictionary because base language model will be mixed"
)
# Add in all the words
words_needed.update(word_dict.keys())
# Write out dictionary with only the necessary words (speeds up loading)
dictionary_path = self.profile.write_path(
self.profile.get(
f"speech_to_text.{self.system}.dictionary", "dictionary.txt"
)
)
words_written = 0
number_duplicates = self.profile.get(
"training.dictionary_number_duplicates", True
)
with open(dictionary_path, "w") as dictionary_file:
for word in sorted(words_needed):
if not word in word_dict:
continue
for i, pronounce in enumerate(word_dict[word]):
if (i < 1) or (not number_duplicates):
print(word, pronounce, file=dictionary_file)
else:
print("%s(%s)" % (word, i + 1), pronounce, file=dictionary_file)
words_written += 1
dictionary_time = time.time() - start_time
self._logger.debug(
f"Wrote {words_written} word(s) to {dictionary_path} in {dictionary_time} second(s)"
)
# Check for unknown words
return words_needed - word_dict.keys()
def make_slot_acceptor(intent_fst: fst.Fst, eps: str = "<eps>") -> fst.Fst:
in_eps = intent_fst.input_symbols().find(eps)
out_eps = intent_fst.output_symbols().find(eps)
slot_fst = fst.Fst()
# Copy symbol tables
all_symbols = fst.SymbolTable()
meta_keys = set()
for table in [intent_fst.input_symbols(), intent_fst.output_symbols()]:
for i in range(table.num_symbols()):
key = table.get_nth_key(i)
sym = table.find(key).decode()
all_key = all_symbols.add_symbol(sym)
if sym.startswith("__"):
meta_keys.add(all_key)
weight_one = fst.Weight.One(slot_fst.weight_type())
weight_zero = fst.Weight.Zero(slot_fst.weight_type())
# States that will be set to final
final_states: Set[int] = set()
# States that already have all-word loops
loop_states: Set[int] = set()
all_eps = all_symbols.find(eps)
# Add self transitions to a state for all input words (besides <eps>)
def add_loop_state(state):
for sym_idx in range(all_symbols.num_symbols()):
all_key = all_symbols.get_nth_key(sym_idx)
if (all_key != all_eps) and (all_key not in meta_keys):
slot_fst.add_arc(state,
fst.Arc(all_key, all_key, weight_one, state))
slot_fst.set_start(slot_fst.add_state())
# Queue of (intent state, acceptor state, copy count)
state_queue: Deque[Tuple[int, int, int]] = deque()
state_queue.append((intent_fst.start(), slot_fst.start(), 0))
# BFS
while len(state_queue) > 0:
intent_state, slot_state, do_copy = state_queue.popleft()
final_states.add(slot_state)
for intent_arc in intent_fst.arcs(intent_state):
out_symbol = intent_fst.output_symbols().find(
intent_arc.olabel).decode()
all_key = all_symbols.find(out_symbol)
if out_symbol.startswith("__label__"):
# Create corresponding __label__ arc
next_state = slot_fst.add_state()
slot_fst.add_arc(
slot_state,
fst.Arc(all_key, all_key, weight_one, next_state))
# Must create a loop here for intents with no slots
add_loop_state(next_state)
loop_states.add(slot_state)
else:
# Non-label arc
if out_symbol.startswith("__begin__"):
# States/arcs will be copied until __end__ is reached
do_copy += 1
# Add loop transitions to soak up non-tag words
if not slot_state in loop_states:
add_loop_state(slot_state)
loop_states.add(slot_state)
if (do_copy > 0) and ((intent_arc.ilabel != in_eps) or
(intent_arc.olabel != out_eps)):
# Copy state/arc
in_symbol = (intent_fst.input_symbols().find(
intent_arc.ilabel).decode())
next_state = slot_fst.add_state()
slot_fst.add_arc(
slot_state,
fst.Arc(all_symbols.find(in_symbol), all_key,
weight_one, next_state),
)
final_states.discard(slot_state)
else:
next_state = slot_state
if out_symbol.startswith("__end__"):
# Stop copying after this state until next __begin__
do_copy -= 1
next_info = (intent_arc.nextstate, next_state, do_copy)
state_queue.append(next_info)
# Mark all dangling states as final (excluding start)
for state in final_states:
if state != slot_fst.start():
slot_fst.set_final(state)
# Fix symbol tables
slot_fst.set_input_symbols(all_symbols)
slot_fst.set_output_symbols(all_symbols)
return slot_fst
def filter_words(words: Iterable[str], the_fst: fst.Fst) -> List[str]:
input_symbols = the_fst.input_symbols()
return [w for w in words if input_symbols.find(w) >= 0]
def _replace_fsts(outer_fst: fst.Fst,
replacements: Dict[int, fst.Fst],
eps="<eps>") -> fst.Fst:
input_symbol_map: Dict[Union[int, Tuple[int, int]], int] = {}
output_symbol_map: Dict[Union[int, Tuple[int, int]], int] = {}
state_map: Dict[Union[int, Tuple[int, int]], int] = {}
# Create new FST
new_fst = fst.Fst()
new_input_symbols = fst.SymbolTable()
new_output_symbols = fst.SymbolTable()
weight_one = fst.Weight.One(new_fst.weight_type())
weight_zero = fst.Weight.Zero(new_fst.weight_type())
weight_final = fst.Weight.Zero(outer_fst.weight_type())
# Copy symbols
outer_input_symbols = outer_fst.input_symbols()
for i in range(outer_input_symbols.num_symbols()):
key = outer_input_symbols.get_nth_key(i)
input_symbol_map[key] = new_input_symbols.add_symbol(
outer_input_symbols.find(key))
outer_output_symbols = outer_fst.output_symbols()
for i in range(outer_output_symbols.num_symbols()):
key = outer_output_symbols.get_nth_key(i)
output_symbol_map[key] = new_output_symbols.add_symbol(
outer_output_symbols.find(key))
in_eps = new_input_symbols.add_symbol(eps)
out_eps = new_output_symbols.add_symbol(eps)
# Copy states
for outer_state in outer_fst.states():
new_state = new_fst.add_state()
state_map[outer_state] = new_state
if outer_fst.final(outer_state) != weight_final:
new_fst.set_final(new_state)
# Set start state
new_fst.set_start(state_map[outer_fst.start()])
# Copy arcs
for outer_state in outer_fst.states():
new_state = state_map[outer_state]
for outer_arc in outer_fst.arcs(outer_state):
next_state = state_map[outer_arc.nextstate]
replace_fst = replacements.get(outer_arc.olabel)
if replace_fst is not None:
# Replace in-line
r = outer_arc.olabel
replace_final = fst.Weight.Zero(replace_fst.weight_type())
replace_input_symbols = replace_fst.input_symbols()
replace_output_symbols = replace_fst.output_symbols()
# Copy states
for replace_state in replace_fst.states():
state_map[(r, replace_state)] = new_fst.add_state()
# Create final arc to next state
if replace_fst.final(replace_state) != replace_final:
new_fst.add_arc(
state_map[(r, replace_state)],
fst.Arc(in_eps, out_eps, weight_one, next_state),
)
# Copy arcs
for replace_state in replace_fst.states():
for replace_arc in replace_fst.arcs(replace_state):
new_fst.add_arc(
state_map[(r, replace_state)],
fst.Arc(
new_input_symbols.add_symbol(
replace_input_symbols.find(
replace_arc.ilabel)),
new_output_symbols.add_symbol(
replace_output_symbols.find(
replace_arc.olabel)),
weight_one,
state_map[(r, replace_arc.nextstate)],
),
)
# Create arc into start state
new_fst.add_arc(
new_state,
fst.Arc(in_eps, out_eps, weight_one,
state_map[(r, replace_fst.start())]),
)
else:
# Copy arc as-is
new_fst.add_arc(
new_state,
fst.Arc(
input_symbol_map[outer_arc.ilabel],
output_symbol_map[outer_arc.olabel],
weight_one,
next_state,
),
)
# Fix symbol tables
new_fst.set_input_symbols(new_input_symbols)
new_fst.set_output_symbols(new_output_symbols)
return new_fst
def minimize_fst(the_fst: fst.Fst) -> fst.Fst:
# BUG: Fst.minimize does not pass allow_nondet through, so we have to call out to the command-line
minimize_cmd = ["fstminimize", "--allow_nondet"]
return fst.Fst.read_from_string(
subprocess.check_output(minimize_cmd, input=the_fst.write_to_string()))
