class WlDecoder:
"""
Wav2Letter-based decoder. Follows the official examples for the python bindings,
see https://github.com/facebookresearch/wav2letter/blob/master/bindings/python/examples/decoder_example.py
"""
def __init__(self,
lm_weight=2.0,
lexicon_path="WER_data/lexicon.txt",
token_path="WER_data/letters.lst",
lm_path="WER_data/4-gram.bin"):
lexicon = load_words(lexicon_path)
word_dict = create_word_dict(lexicon)
self.token_dict = Dictionary(token_path)
self.lm = KenLM(lm_path, word_dict)
self.sil_idx = self.token_dict.get_index("|")
self.unk_idx = word_dict.get_index("<unk>")
self.token_dict.add_entry("#")
self.blank_idx = self.token_dict.get_index('#')
self.trie = Trie(self.token_dict.index_size(), self.sil_idx)
start_state = self.lm.start(start_with_nothing=False)
for word, spellings in lexicon.items():
usr_idx = word_dict.get_index(word)
_, score = self.lm.score(start_state, usr_idx)
for spelling in spellings:
# max_reps should be 1; using 0 here to match DecoderTest bug
spelling_idxs = tkn_to_idx(spelling,
self.token_dict,
max_reps=0)
self.trie.insert(spelling_idxs, usr_idx, score)
self.trie.smear(SmearingMode.MAX)
self.opts = DecoderOptions(beam_size=2500,
beam_threshold=100.0,
lm_weight=lm_weight,
word_score=2.0,
unk_score=-math.inf,
log_add=False,
sil_weight=-1,
criterion_type=CriterionType.CTC)
def collapse(self, prediction):
result = []
for p in prediction:
if result and p == result[-1]:
continue
result.append(p)
blank = '#'
space = '|'
result = [x for x in result if x != blank]
result = [(x if x != space else ' ') for x in result if x != blank]
return result
def predictions(self, emissions):
t, n = emissions.size()
emissions = emissions.cpu().numpy()
decoder = WordLMDecoder(self.opts, self.trie, self.lm, self.sil_idx,
self.blank_idx, self.unk_idx, [])
results = decoder.decode(emissions.ctypes.data, t, n)
prediction = [
self.token_dict.get_entry(x) for x in results[0].tokens if x >= 0
]
prediction = self.collapse(prediction)
return prediction
lm_state, lm_score = lm.finish(lm_state)
total_score += lm_score
assert_near(total_score, -19.5123, 1e-5)
# build trie
# Trie is necessary to do beam-search decoding with word-level lm
# We restrict our search only to the words from the lexicon
# Trie is constructed from the lexicon, each node is a token
# path from the root to a leaf corresponds to a word spelling in the lexicon
# get silence index
sil_idx = token_dict.get_index("|")
# get unknown word index
unk_idx = word_dict.get_index("<unk>")
# create the trie, specifying how many tokens we have and silence index
trie = Trie(token_dict.index_size(), sil_idx)
start_state = lm.start(False)
# use heuristic for the trie, called smearing:
# predict lm score for each word in the lexicon, set this score to a leaf
# (we predict lm score for each word as each word starts a sentence)
# word score of a leaf is propagated up to the root to have some proxy score
# for any intermediate path in the trie
# SmearingMode defines the function how to process scores
# in a node came from the children nodes:
# could be max operation or logadd or none
for word, spellings in lexicon.items():
usr_idx = word_dict.get_index(word)
_, score = lm.score(start_state, usr_idx)
for spelling in spellings:
# max_reps should be 1; using 0 here to match DecoderTest bug