def __init__(self, model_name, trans_df):
from espnet2.bin.asr_inference import Speech2Text
from espnet_model_zoo.downloader import ModelDownloader
import jiwer
self.model_name = model_name
d = ModelDownloader()
self.asr_model = Speech2Text(**d.download_and_unpack(model_name))
self.input_txt_list = []
self.clean_txt_list = []
self.output_txt_list = []
self.transcriptions = []
self.true_txt_list = []
self.sample_rate = int(
d.data_frame[d.data_frame["name"] == model_name]["fs"])
self.trans_df = trans_df
self.trans_dic = self._df_to_dict(trans_df)
self.mix_counter = Counter()
self.clean_counter = Counter()
self.est_counter = Counter()
self.transformation = jiwer.Compose([
jiwer.ToLowerCase(),
jiwer.RemovePunctuation(),
jiwer.RemoveMultipleSpaces(),
jiwer.Strip(),
jiwer.SentencesToListOfWords(),
jiwer.RemoveEmptyStrings(),
])
def on_record_button_toggled(self, button):
if button.get_active():
self.reset_playback()
f = open(
os.path.join(self.training_dir, "%d.txt" % self.sample_id),
'w')
f.write(self.sentence_text)
f.close()
self.wavrecfile.set_property(
"location",
os.path.join(self.training_dir, "%d.wav" % self.sample_id))
training_file = os.path.join("Deepspeech", "training-data",
"%d.wav" % self.sample_id)
stripped_text = jiwer.RemovePunctuation()(
self.sentence_text).lower().encode("ascii",
"ignore").decode().replace(
"'", "")
wavsize = os.stat(location).st_size
if self.sample_id % 6 < 3:
# Training sample
self.training_csv.write(
"%s,%d,%s" % (training_file, wavsize, stripped_text))
else:
# Testing sample
self.testing_csv.write("%s,%d,%s" %
(training_file, wavsize, stripped_text))
self.record_pipeline.set_state(Gst.State.PLAYING)
self.play_button.set_sensitive(True)
self.have_sample = True
else:
self.reset_recording()
def on_test_message(self, bus, message):
structure = message.get_structure()
if structure and structure.get_name(
) == "deepspeech" and structure.get_value("intermediate") == False:
self.recognised_text += structure.get_value("text") + "\n"
self.training_progress.set_fraction(
(self.testing_sample + 1) / self.sample_id)
if self.testing_sample < self.sample_id - 1:
self.test_pipeline.set_state(Gst.State.NULL)
self.test_sample(self.testing_sample + 1)
else:
self.test_text = jiwer.RemovePunctuation()(
self.test_text).lower().encode("ascii", "ignore").decode()
print("Expected:", self.test_text)
print("Got:", self.recognised_text)
accuracy = 100 - jiwer.wer(
self.test_text.replace("\n", " "),
self.recognised_text.replace("\n", " ").replace("'",
"")) * 100
if self.pretraining:
pretraining_accuracy_label = self.builder.get_object(
"pretraining_accuracy_label")
pretraining_accuracy_label.set_text("%.2f%%" % accuracy)
self.pretraining = False
self.training = True
status_label = self.builder.get_object("status_label")
status_label.set_text("Training...")
self.training_progress.set_fraction(0)
if self.posttraining:
posttraining_accuracy_label = self.builder.get_object(
"posttraining_accuracy_label")
posttraining_accuracy_label.set_text("%.2f%%" % accuracy)
spinner = self.builder.get_object("spinner")
spinner.set_active = False
self.posttraining = False
def rmPunctuation(values):
"""preprocess the list of words to RemovePunctuation"""
newValues = []
for v in values:
newValue = jiwer.RemovePunctuation()(v)
newValue = jiwer.Strip()(newValue)
newValue = jiwer.RemoveMultipleSpaces()(newValue)
newValues.append(newValue)
return newValues
def calc_wer(ground_truth, hypothesis):
transformation = jiwer.Compose([
jiwer.ToLowerCase(),
jiwer.RemoveMultipleSpaces(),
jiwer.Strip(),
jiwer.ExpandCommonEnglishContractions(),
jiwer.RemovePunctuation()
])
wer = jiwer.wer(ground_truth,
hypothesis,
truth_transform=transformation,
hypothesis_transform=transformation)
return wer
def compute_perc_script_missing(original_script, transcript, language):
'''
Check how much of original_script is missing in transcript. Clean and remove stopwords
'''
# print(original_script)
# print(transcript)
cleaning = jiwer.Compose([
jiwer.SubstituteRegexes({"¡": "", "¿":"", "á": "a", "é": "e", "í": "i", "ó": "o","ú": "u"}),
jiwer.SubstituteWords({ "tardes": "dias",
"noches": "dias",
" uno ": " 1 ",
" dos ": " 2 ",
" tres ": " 3 ",
" cuatro ": " 4 ",
" cinco ": " 5 ",
" seis ": " 6 ",
" siete ": " 7 ",
" ocho ": " 8 ",
" nueve ": " 9 "}),
jiwer.RemovePunctuation(),
jiwer.ToLowerCase(),
jiwer.SentencesToListOfWords(word_delimiter=" "),
jiwer.RemoveEmptyStrings()
])
#Remove anything between ${variable} from original_script
original_script_transformed = re.sub(r'\${.*?\}','',original_script)
# print(original_script_transformed)
#Clean both
original_script_transformed = cleaning(original_script_transformed)
transcript_transformed = cleaning(transcript)
# print(original_script_transformed)
#Remove stopwords from original_script
original_script_transformed_no_stopwords = remove_stopwords(original_script_transformed, language)
if len(original_script_transformed_no_stopwords) != 0: #Sometimes removing stopwords removes all words from script
original_script_transformed = original_script_transformed_no_stopwords
#Lemmatize transcript
stemmer = get_stemmer(language)
transcript_transformed_stem = [stemmer.stem(word) for word in transcript_transformed]
#Get words form original_script_transformed whose stem is not in transcript_transformed_stem
words_missing = [word for word in original_script_transformed if stemmer.stem(word) not in transcript_transformed_stem]
return len(words_missing)/len(original_script_transformed), words_missing
def sentence_wer(reference: str, prediction: str):
transformation = jiwer.Compose([
jiwer.RemoveMultipleSpaces(),
jiwer.RemovePunctuation(),
jiwer.Strip(),
jiwer.ToLowerCase(),
jiwer.ExpandCommonEnglishContractions(),
jiwer.RemoveWhiteSpace(replace_by_space=True),
jiwer.SentencesToListOfWords(),
jiwer.RemoveEmptyStrings(),
])
return jiwer.wer(reference.strip(),
prediction.strip(),
truth_transform=transformation,
hypothesis_transform=transformation)
def normalize_sentence(sentence):
"""Normalize sentence"""
# Convert all characters to upper.
sentence = sentence.upper()
# Delete punctuations.
sentence = jiwer.RemovePunctuation()(sentence)
# Remove \n, \t, \r, \x0c.
sentence = jiwer.RemoveWhiteSpace(replace_by_space=True)(sentence)
# Remove multiple spaces.
sentence = jiwer.RemoveMultipleSpaces()(sentence)
# Remove white space in two end of string.
sentence = jiwer.Strip()(sentence)
# Convert all characters to upper.
sentence = sentence.upper()
return sentence
def evaluate(testset, audio_directory):
model = deepspeech.Model('deepspeech-0.7.0-models.pbmm')
model.enableExternalScorer('deepspeech-0.7.0-models.scorer')
predictions = []
targets = []
for i, datapoint in enumerate(testset):
audio, rate = sf.read(
os.path.join(audio_directory, f'example_output_{i}.wav'))
assert rate == model.sampleRate(), 'wrong sample rate'
audio_int16 = (audio * (2**15)).astype(np.int16)
text = model.stt(audio_int16)
predictions.append(text)
target_text = unidecode(datapoint['text'])
targets.append(target_text)
transformation = jiwer.Compose(
[jiwer.RemovePunctuation(),
jiwer.ToLowerCase()])
targets = transformation(targets)
predictions = transformation(predictions)
logging.info(f'targets: {targets}')
logging.info(f'predictions: {predictions}')
logging.info(f'wer: {jiwer.wer(targets, predictions)}')
def analyze():
try:
req_data = request.get_json()
compose_rule_set = []
if req_data.get('to_lower_case', False) == True:
compose_rule_set.append(jiwer.ToLowerCase())
if req_data.get('strip_punctuation', False) == True:
compose_rule_set.append(jiwer.RemovePunctuation())
if req_data.get('strip_words', False) == True:
compose_rule_set.append(jiwer.Strip())
if req_data.get('strip_multi_space', False) == True:
compose_rule_set.append(jiwer.RemoveMultipleSpaces())
word_excepts = req_data.get('t_words', '')
if word_excepts != '':
words = [a.strip() for a in word_excepts.split(",")]
compose_rule_set.append(jiwer.RemoveSpecificWords(words))
compose_rule_set.append(
jiwer.RemoveWhiteSpace(
replace_by_space=req_data.get('replace_whitespace', False)))
transformation = jiwer.Compose(compose_rule_set)
measures = jiwer.compute_measures(req_data.get('s_truth', ""),
req_data.get('s_hypo', ""),
truth_transform=transformation,
hypothesis_transform=transformation)
return jsonify({
"wer": measures['wer'],
"mer": measures['mer'],
"wil": measures['wil']
})
except:
return jsonify("API endpoint Error")
def metric(ref_trans, asr_trans, lang):
if lang == "en":
transformation = jiwer.Compose([
jiwer.Strip(),
jiwer.ToLowerCase(),
jiwer.RemoveWhiteSpace(replace_by_space=True),
jiwer.RemoveMultipleSpaces(),
jiwer.SentencesToListOfWords(word_delimiter=" "),
jiwer.RemoveEmptyStrings(),
jiwer.RemovePunctuation(),
])
wer = jiwer.wer(
ref_trans,
asr_trans,
truth_transform=transformation,
hypothesis_transform=transformation,
)
elif lang == "cn":
del_symblos = re.compile(r"[^\u4e00-\u9fa5]+")
for idx in range(len(asr_trans)):
sentence = re.sub(del_symblos, "", asr_trans[idx])
sentence = list(sentence)
sentence = " ".join(sentence)
asr_trans[idx] = sentence
sentence = re.sub(del_symblos, "", ref_trans[idx])
sentence = list(sentence)
sentence = " ".join(sentence)
ref_trans[idx] = sentence
asr_valid = set(asr_trans)
assert len(asr_valid) == len(asr_trans)
wer = jiwer.wer(ref_trans, asr_trans)
else:
raise ("Args error!")
return wer
srt.push(next_sub)
break
end = next_sub.end.hours * 3600 + next_sub.end.minutes * 60 + next_sub.end.seconds + next_sub.end.milliseconds / 1000
ground_truth = ground_truth + " " + next_sub.text_without_tags
hypothesis = kd.query_text(start, end)
else:
break
kd.mark_words(start, end)
transformation = jiwer.Compose([
jiwer.ToLowerCase(),
jiwer.RemoveMultipleSpaces(),
jiwer.RemoveWhiteSpace(replace_by_space=True),
jiwer.SentencesToListOfWords(),
jiwer.RemovePunctuation(),
jiwer.RemoveEmptyStrings(),
jiwer.SubstituteRegexes({r"ё": r"е"})
])
gt = transformation([ground_truth])
hp = transformation([hypothesis])
gt, hp = replace_pairs(gt, hp)
hp, gt = replace_pairs(hp, gt)
wer(gt, hp)
r = jiwer.compute_measures(
gt,
hp
)
def RemovePunctuation(replace_by_space=False):
if not replace_by_space:
return jiwer.RemovePunctuation()
return replace_punctuations_by_space
