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)
return np.frombuffer(file.read(N * N * 4), dtype=np.float32)
def assert_near(x, y, tol):
assert abs(x - y) <= tol
# load test files
T, N = load_TN(os.path.join(testing_data_path, "TN.bin"))
emissions = load_emissions(os.path.join(testing_data_path, "emission.bin"))
transitions = load_transitions(
os.path.join(testing_data_path, "transition.bin"))
lexicon = loadWords(os.path.join(testing_data_path, "words.lst"))
wordDict = createWordDict(lexicon)
tokenDict = Dictionary(os.path.join(testing_data_path, "letters.lst"))
tokenDict.addEntry("1")
lm = KenLM(os.path.join(testing_data_path, "lm.arpa"), wordDict)
# test LM
#sentence = ["the", "cat", "sat", "on", "the", "mat"]
#lm_state = lm.start(False)
#total_score = 0
#lm_score_target = [-1.05971, -4.19448, -3.33383, -2.76726, -1.16237, -4.64589]
#for i in range(len(sentence)):
# lm_state, lm_score = lm.score(lm_state, wordDict.getIndex(sentence[i]))
# assert_near(lm_score, lm_score_target[i], 1e-5)
# total_score += lm_score
#lm_state, lm_score = lm.finish(lm_state)
#total_score += lm_score
# load test files
# load time and number of tokens for dumped acoustic scores
T, N = load_tn(os.path.join(data_path, "TN.bin"))
# load emissions [Batch=1, Time, Ntokens]
emissions = load_emissions(os.path.join(data_path, "emission.bin"))
# load transitions (from ASG loss optimization) [Ntokens, Ntokens]
transitions = load_transitions(os.path.join(data_path, "transition.bin"))
# load lexicon file, which defines spelling of words
# the format word and its tokens spelling separated by the spaces,
# for example for letters tokens with ASG loss:
# ann a n 1 |
lexicon = load_words(os.path.join(data_path, "words.lst"))
# read lexicon and store it in the w2l dictionary
word_dict = create_word_dict(lexicon)
# create w2l dict with tokens set (letters in this example)
token_dict = Dictionary(os.path.join(data_path, "letters.lst"))
# add repetition symbol as soon as we have ASG acoustic model
token_dict.add_entry("1")
# create Kenlm language model
lm = KenLM(os.path.join(data_path, "lm.arpa"), word_dict)
# test LM
sentence = ["the", "cat", "sat", "on", "the", "mat"]
# start LM with nothing, get its current state
lm_state = lm.start(False)
total_score = 0
lm_score_target = [
-1.05971, -4.19448, -3.33383, -2.76726, -1.16237, -4.64589
]
# iterate over words in the sentence
for i in range(len(sentence)):