def init_directory(self):
# init directory for save the audio
make_dir(os.path.join(self.audio_dir, self.tts.getName()))
# init directory for save the transcription
for asr in self.asrs:
make_dir(
os.path.join(self.transcription_dir, self.tts.getName(),
asr.getName()))
def get_outputfile_for_failed_test_case(self):
asrs_dir = "_".join([asr.getName() for asr in self.asrs])
result_dir = os.path.join(
self.output_dir, "result", self.tts.getName(), asrs_dir,
f"num_iteration_{self.num_iteration}",
f"text_batch_size_{self.text_batch_size if self.text_batch_size else 'global' }"
)
make_dir(result_dir)
experiment_name = f"with-estimator-{self.estimator.getName().replace('/','-')}" if self.estimator else "without-estimator"
return os.path.join(result_dir, experiment_name + ".json")
def saveFailedTestCases(self, processed_texts, cases):
failed_test_case_dir = os.path.join(self.output_dir,
"failed_test_cases",
self.tts.getName(),
self.target_asr)
make_dir(failed_test_case_dir)
ids = self.get_id_only(processed_texts)
input_texts = self.get_text_only(processed_texts)
source_audio_dir = os.path.join(self.audio_dir, self.tts.getName())
for input_text, filename, case in zip(input_texts, ids, cases):
if case[self.target_asr] == FAILED_TEST_CASE:
src_audio_fpath = source_audio_dir + f"/{filename}.wav"
trgt_audio_fpath = failed_test_case_dir + f"/{filename}.wav"
os.system(f"cp {src_audio_fpath} {trgt_audio_fpath}")
ground_truth_file = failed_test_case_dir + f"/{filename}.txt"
f = open(ground_truth_file, 'w+')
f.write(input_text)
f.close()
def generate(tts_name: str, corpus_path: str, data_dir: str, execution_time_dir:str):
tts = create_tts_by_name(tts_name)
audio_dir = os.path.join(data_dir, AUDIO_DIR)
execution_time_dir = os.path.join(execution_time_dir, AUDIO_DIR, tts_name)
make_dir(execution_time_dir)
corpus = read_corpus(corpus_path)
for i in range(0, 3) :
c = corpus[i]
text = c.getText()
filename = c.getId()
start = time.time()
tts.generateAudio(text=text, audio_dir=audio_dir, filename=filename)
end = time.time()
execution_time = end - start
fpath = os.path.join(execution_time_dir, filename + ".txt")
save_execution_time(fpath=fpath, execution_time=execution_time)
print(f"Generate {i}")
i += 1
if tts_name in ["google"]:
random_number = float(random.randint(15, 40))/10.
time.sleep(random_number)
df = df.reset_index(drop=True)
print("get sample: " + str(datetime.now()))
# get sample data
df = df.sample(frac=1, random_state=seed).reset_index(drop=True)
N = 20000
sample_df = get_sample_data(df, int(2 * N))
print("preprocess data: " + str(datetime.now()))
# text preprocessing
data = preprocess_data(sample_df, N)
print("write data: " + str(datetime.now()))
# prepare folder to save the data
directory = config["output_dir"]
if len(config["corpus_fpath"].split("/")) > 1:
directory = os.path.join(
directory, "/".join(config["corpus_fpath"].split("/")[:-1]))
make_dir(directory)
outfile = os.path.join(config["output_dir"], config["corpus_fpath"])
file = open(outfile, "w+")
for s in data:
file.write("%s\n" % s)
file.close()
def saveTranscription(self, transcription_dir: str, filename: str):
transcription_dir = os.path.join(transcription_dir, self.getName())
make_dir(transcription_dir)
transcription_path = os.path.join(transcription_dir, filename + ".txt")
with open(transcription_path, "w+") as f:
f.write(self.getTranscription())
def processText(self, text: str, filename: str):
"""
Run CrossASR on a single text
Description: Given a sentence as input, the program will generate a test case. The program needs some parameters, i.e. a TTS and ASRs used
:params text:
:params filename:
:returns case:
:returns execution time:
"""
execution_time = 0.
directory = os.path.join(self.execution_time_dir, AUDIO_DIR,
self.getTTS().getName())
make_dir(directory)
time_for_generating_audio_fpath = os.path.join(directory,
filename + ".txt")
audio_fpath = self.getTTS().getAudioPath(text=text,
audio_dir=self.audio_dir,
filename=filename)
if self.recompute or not os.path.exists(audio_fpath):
# print(audio_fpath)
start_time = time.time()
self.getTTS().generateAudio(text=text, audio_fpath=audio_fpath)
save_execution_time(fpath=time_for_generating_audio_fpath,
execution_time=time.time() - start_time)
## add execution time for generating audio
execution_time += get_execution_time(
fpath=time_for_generating_audio_fpath)
transcription_dir = os.path.join(self.transcription_dir,
self.getTTS().getName())
transcriptions = {}
for asr in self.asrs:
directory = os.path.join(self.execution_time_dir,
TRANSCRIPTION_DIR,
self.getTTS().getName(), asr.getName())
make_dir(directory)
time_for_recognizing_audio_fpath = os.path.join(
directory, filename + ".txt")
if self.recompute:
start_time = time.time()
# TODO:
# change recognize audio -> input audio instead of fpath
# audio = asr.loadAudio(audio_fpath=audio_fpath)
# transcription = asr.recognizeAudio(audio=audio)
# asr.saveTranscription(transcription_fpath, transcription)
transcription = asr.recognizeAudio(audio_fpath=audio_fpath)
asr.setTranscription(transcription)
asr.saveTranscription(transcription_dir=transcription_dir,
filename=filename)
save_execution_time(fpath=time_for_recognizing_audio_fpath,
execution_time=time.time() - start_time)
transcription = asr.loadTranscription(
transcription_dir=transcription_dir, filename=filename)
num_retry = 0
while transcription == "" and num_retry < self.max_num_retry:
start_time = time.time()
asr.recognizeAudio(audio_fpath=audio_fpath)
asr.saveTranscription(transcription_dir=transcription_dir,
filename=filename)
save_execution_time(fpath=time_for_recognizing_audio_fpath,
execution_time=time.time() - start_time)
transcription = asr.loadTranscription(
transcription_dir=transcription_dir, filename=filename)
if asr.getName() == "wit":
random_number = float(random.randint(9, 47)) / 10.
time.sleep(random_number)
num_retry += 1
transcriptions[asr.getName()] = preprocess_text(transcription)
## add execution time for generating audio
execution_time += get_execution_time(
fpath=time_for_recognizing_audio_fpath)
cases = self.caseDeterminer(text, transcriptions)
# if sum(cases.values()) == 0 :
# print(text)
# print(transcriptions["wav2vec2"])
# print(cases)
# print()
for asr_name, case in cases.items():
self.saveCase(self.case_dir,
self.getTTS().getName(), asr_name, filename,
str(case))
# print(f"Execution time: {execution_time}")
return cases, execution_time
class CrossASR:
def __init__(self,
tts: TTS,
asrs: [ASR],
output_dir: "",
target_asr=None,
recompute=False,
num_iteration=5,
time_budget=3600,
max_num_retry=0,
text_batch_size=None,
seed=None,
estimator=None):
self.tts = tts
self.asrs = asrs
self.target_asr = target_asr
self.output_dir = output_dir
self.audio_dir = os.path.join(output_dir, DATA_DIR, AUDIO_DIR)
self.transcription_dir = os.path.join(output_dir, DATA_DIR,
TRANSCRIPTION_DIR)
self.init_directory()
## TODO: make init directory for execution time and case
self.execution_time_dir = os.path.join(output_dir, EXECUTION_TIME_DIR)
self.case_dir = os.path.join(output_dir, CASE_DIR)
self.recompute = recompute
self.num_iteration = num_iteration
self.time_budget = time_budget
self.max_num_retry = max_num_retry
self.text_batch_size = text_batch_size
self.estimator = estimator
self.outputfile_failed_test_case = self.get_outputfile_for_failed_test_case(
)
if seed:
crossasr.utils.set_seed(seed)
## TODO: convert print into global logging
def init_directory(self):
# init directory for save the audio
make_dir(os.path.join(self.audio_dir, self.tts.getName()))
# init directory for save the transcription
for asr in self.asrs:
make_dir(
os.path.join(self.transcription_dir, self.tts.getName(),
asr.getName()))
def get_outputfile_for_failed_test_case(self):
asrs_dir = "_".join([asr.getName() for asr in self.asrs])
result_dir = os.path.join(
self.output_dir, "result", self.tts.getName(), asrs_dir,
f"num_iteration_{self.num_iteration}",
f"text_batch_size_{self.text_batch_size if self.text_batch_size else 'global' }"
)
make_dir(result_dir)
experiment_name = f"with-estimator-{self.estimator.getName().replace('/','-')}" if self.estimator else "without-estimator"
return os.path.join(result_dir, experiment_name + ".json")
def getTTS(self):
return self.tts
def setTTS(self, tts: TTS):
self.tts = tts
def getASRS(self):
return self.asrs
def addASR(self, asr: ASR):
for curr_asr in self.asrs:
if asr.getName() == curr_asr.getName():
# asr is already on the list of asrs
return
self.asrs.append(asr)
def removeASR(self, asr_name: str):
for i, asr in enumerate(self.asrs):
if asr_name == asr.getName():
break
del self.asrs[i]
def getOutputDir(self):
return self.audio_dir
def setOutputDir(self, output_dir: str):
self.output_dir = output_dir
self.audio_dir = os.path.join(output_dir, DATA_DIR, AUDIO_DIR)
self.transcription_dir = os.path.join(output_dir, DATA_DIR,
TRANSCRIPTION_DIR)
self.execution_time_dir = os.path.join(output_dir, EXECUTION_TIME_DIR)
self.case_dir = os.path.join(output_dir, CASE_DIR)
def caseDeterminer(self, text: str, transcriptions: str):
# word error rate
wers = {}
is_determinable = False
for k, transcription in transcriptions.items():
word_error_rate = wer(text, transcription)
wers[k] = word_error_rate
if word_error_rate == 0:
is_determinable = True
case = {}
if is_determinable:
for k in transcriptions.keys():
if wers[k] == 0:
case[k] = SUCCESSFUL_TEST_CASE
else:
case[k] = FAILED_TEST_CASE
else:
for k in transcriptions.keys():
case[k] = INDETERMINABLE_TEST_CASE
return case
def saveCase(self, case_dir: str, tts_name: str, asr_name: str,
filename: str, case: str):
case_dir = os.path.join(case_dir, tts_name, asr_name)
make_dir(case_dir)
fpath = os.path.join(case_dir, filename + ".txt")
file = open(fpath, "w+")
file.write(case)
file.close()
