def __init__(self,
lexicon: Lexicon,
P: k2.Fsa,
device: torch.device,
oov: str = '<UNK>'):
'''
Args:
L_inv:
Its labels are words, while its aux_labels are phones.
P:
A phone bigram LM if the pronunciations in the lexicon are in phones;
a word piece bigram if the pronunciations in the lexicon are word pieces.
phones:
The phone symbol table.
words:
The word symbol table.
oov:
Out of vocabulary word.
'''
self.lexicon = lexicon
L_inv = self.lexicon.L_inv.to(device)
P = P.to(device)
if L_inv.properties & k2.fsa_properties.ARC_SORTED != 0:
L_inv = k2.arc_sort(L_inv)
assert L_inv.requires_grad is False
assert oov in self.lexicon.words
self.L_inv = L_inv
self.oov_id = self.lexicon.words[oov]
self.oov = oov
self.device = device
phone_symbols = get_phone_symbols(self.lexicon.phones)
phone_symbols_with_blank = [0] + phone_symbols
ctc_topo = build_ctc_topo(phone_symbols_with_blank).to(device)
assert ctc_topo.requires_grad is False
ctc_topo_inv = k2.arc_sort(ctc_topo.invert_())
P_with_self_loops = k2.add_epsilon_self_loops(P)
ctc_topo_P = k2.intersect(ctc_topo_inv,
P_with_self_loops,
treat_epsilons_specially=False).invert()
self.ctc_topo_P = k2.arc_sort(ctc_topo_P)
def __init__(self,
L_inv: k2.Fsa,
L_disambig: k2.Fsa,
G: k2.Fsa,
phones: k2.SymbolTable,
words: k2.SymbolTable,
device: torch.device,
oov: str = '<UNK>'):
'''
Args:
L_inv:
Its labels are words, while its aux_labels are phones.
L_disambig:
L with disambig symbols. Its labels are phones and aux_labels
are words.
G:
The language model.
phones:
The phone symbol table.
words:
The word symbol table.
device:
The target device that all FSAs should be moved to.
oov:
Out of vocabulary word.
'''
L_inv = L_inv.to(device)
G = G.to(device)
if L_inv.properties & k2.fsa_properties.ARC_SORTED != 0:
L_inv = k2.arc_sort(L_inv)
if G.properties & k2.fsa_properties.ARC_SORTED != 0:
G = k2.arc_sort(G)
assert L_inv.requires_grad is False
assert G.requires_grad is False
assert oov in words
L = L_inv.invert()
L = k2.arc_sort(L)
self.L_inv = L_inv
self.L = L
self.phones = phones
self.words = words
self.device = device
self.oov_id = self.words[oov]
phone_symbols = get_phone_symbols(phones)
phone_symbols_with_blank = [0] + phone_symbols
ctc_topo = k2.arc_sort(
build_ctc_topo(phone_symbols_with_blank).to(device))
assert ctc_topo.requires_grad is False
self.ctc_topo = ctc_topo
self.ctc_topo_inv = k2.arc_sort(ctc_topo.invert())
lang_dir = Path('data/lang_nosp')
if not (lang_dir / 'HLG_uni.pt').exists():
logging.info("Composing (ctc_topo, L_disambig, G)")
first_phone_disambig_id = find_first_disambig_symbol(phones)
first_word_disambig_id = find_first_disambig_symbol(words)
# decoding_graph is the result of composing (ctc_topo, L_disambig, G)
decoding_graph = compile_HLG(
L=L_disambig.to('cpu'),
G=G.to('cpu'),
H=ctc_topo.to('cpu'),
labels_disambig_id_start=first_phone_disambig_id,
aux_labels_disambig_id_start=first_word_disambig_id)
torch.save(decoding_graph.as_dict(), lang_dir / 'HLG_uni.pt')
else:
logging.info("Loading pre-compiled HLG")
decoding_graph = k2.Fsa.from_dict(
torch.load(lang_dir / 'HLG_uni.pt'))
assert hasattr(decoding_graph, 'phones')
self.decoding_graph = decoding_graph.to(device)