def run_vad(self):
"""
Determine which frames contain speech and nonspeech,
and store the resulting boolean mask internally.
"""
def _compute_runs(array):
"""
Compute runs as a list of arrays,
each containing the indices of a contiguous run.
:param array: the data array
:type array: :class:`numpy.ndarray` (1D)
:rtype: list of :class:`numpy.ndarray` (1D)
"""
if len(array) < 1:
return []
return numpy.split(array, numpy.where(numpy.diff(array) != 1)[0] + 1)
self.log(u"Creating VAD object")
vad = VAD(rconf=self.rconf, logger=self.logger)
self.log(u"Running VAD...")
self.__mfcc_mask = vad.run_vad(self.__mfcc[0])
self.__mfcc_mask_map = (numpy.where(self.__mfcc_mask))[0]
self.log(u"Running VAD... done")
self.log(u"Storing speech and nonspeech intervals...")
# where( == True) already computed, reusing
#runs = _compute_runs((numpy.where(self.__mfcc_mask))[0])
runs = _compute_runs(self.__mfcc_mask_map)
self.__speech_intervals = [(r[0], r[-1]) for r in runs]
# where( == False) not already computed, computing now
runs = _compute_runs((numpy.where(~self.__mfcc_mask))[0])
self.__nonspeech_intervals = [(r[0], r[-1]) for r in runs]
self.log(u"Storing speech and nonspeech intervals... done")
def run_vad(self):
"""
Determine which frames contain speech and nonspeech,
and store the resulting boolean mask internally.
"""
def _compute_runs(array):
"""
Compute runs as a list of arrays,
each containing the indices of a contiguous run.
:param array: the data array
:type array: :class:`numpy.ndarray` (1D)
:rtype: list of :class:`numpy.ndarray` (1D)
"""
if len(array) < 1:
return []
return numpy.split(array,
numpy.where(numpy.diff(array) != 1)[0] + 1)
self.log(u"Creating VAD object")
vad = VAD(rconf=self.rconf, logger=self.logger)
self.log(u"Running VAD...")
self.__mfcc_mask = vad.run_vad(self.__mfcc[0])
self.__mfcc_mask_map = (numpy.where(self.__mfcc_mask))[0]
self.log(u"Running VAD... done")
self.log(u"Storing speech and nonspeech intervals...")
# where( == True) already computed, reusing
# COMMENTED runs = _compute_runs((numpy.where(self.__mfcc_mask))[0])
runs = _compute_runs(self.__mfcc_mask_map)
self.__speech_intervals = [(r[0], r[-1]) for r in runs]
# where( == False) not already computed, computing now
runs = _compute_runs((numpy.where(~self.__mfcc_mask))[0])
self.__nonspeech_intervals = [(r[0], r[-1]) for r in runs]
self.log(u"Storing speech and nonspeech intervals... done")
def _extract_speech(self):
""" Extract speech intervals """
self._log("Running VAD...")
vad = VAD(frame_rate=self.frame_rate, logger=self.logger)
vad.wave_len = self.audio_file.audio_length
vad.wave_mfcc = self.audio_file.audio_mfcc
vad.compute_vad()
self.audio_speech = vad.speech
self._log("Running VAD... done")
def run_vad(
self,
log_energy_threshold=None,
min_nonspeech_length=None,
extend_before=None,
extend_after=None
):
"""
Determine which frames contain speech and nonspeech,
and store the resulting boolean mask internally.
The four parameters might be ``None``:
in this case, the corresponding RuntimeConfiguration values
are applied.
:param float log_energy_threshold: the minimum log energy threshold to consider a frame as speech
:param int min_nonspeech_length: the minimum length, in frames, of a nonspeech interval
:param int extend_before: extend each speech interval by this number of frames to the left (before)
:param int extend_after: extend each speech interval by this number of frames to the right (after)
"""
def _compute_runs(array):
"""
Compute runs as a list of arrays,
each containing the indices of a contiguous run.
:param array: the data array
:type array: :class:`numpy.ndarray` (1D)
:rtype: list of :class:`numpy.ndarray` (1D)
"""
if len(array) < 1:
return []
return numpy.split(array, numpy.where(numpy.diff(array) != 1)[0] + 1)
self.log(u"Creating VAD object")
vad = VAD(rconf=self.rconf, logger=self.logger)
self.log(u"Running VAD...")
self.__mfcc_mask = vad.run_vad(
wave_energy=self.__mfcc[0],
log_energy_threshold=log_energy_threshold,
min_nonspeech_length=min_nonspeech_length,
extend_before=extend_before,
extend_after=extend_after
)
self.__mfcc_mask_map = (numpy.where(self.__mfcc_mask))[0]
self.log(u"Running VAD... done")
self.log(u"Storing speech and nonspeech intervals...")
# where( == True) already computed, reusing
# COMMENTED runs = _compute_runs((numpy.where(self.__mfcc_mask))[0])
runs = _compute_runs(self.__mfcc_mask_map)
self.__speech_intervals = [(r[0], r[-1]) for r in runs]
# where( == False) not already computed, computing now
runs = _compute_runs((numpy.where(~self.__mfcc_mask))[0])
self.__nonspeech_intervals = [(r[0], r[-1]) for r in runs]
self.log(u"Storing speech and nonspeech intervals... done")
def run_vad(self,
log_energy_threshold=None,
min_nonspeech_length=None,
extend_before=None,
extend_after=None):
"""
Determine which frames contain speech and nonspeech,
and store the resulting boolean mask internally.
The four parameters might be ``None``:
in this case, the corresponding RuntimeConfiguration values
are applied.
:param float log_energy_threshold: the minimum log energy threshold to consider a frame as speech
:param int min_nonspeech_length: the minimum length, in frames, of a nonspeech interval
:param int extend_before: extend each speech interval by this number of frames to the left (before)
:param int extend_after: extend each speech interval by this number of frames to the right (after)
"""
def _compute_runs(array):
"""
Compute runs as a list of arrays,
each containing the indices of a contiguous run.
:param array: the data array
:type array: :class:`numpy.ndarray` (1D)
:rtype: list of :class:`numpy.ndarray` (1D)
"""
if len(array) < 1:
return []
return numpy.split(array,
numpy.where(numpy.diff(array) != 1)[0] + 1)
self.log(u"Creating VAD object")
vad = VAD(rconf=self.rconf, logger=self.logger)
self.log(u"Running VAD...")
self.__mfcc_mask = vad.run_vad(
wave_energy=self.__mfcc[0],
log_energy_threshold=log_energy_threshold,
min_nonspeech_length=min_nonspeech_length,
extend_before=extend_before,
extend_after=extend_after)
self.__mfcc_mask_map = (numpy.where(self.__mfcc_mask))[0]
self.log(u"Running VAD... done")
self.log(u"Storing speech and nonspeech intervals...")
# where( == True) already computed, reusing
# COMMENTED runs = _compute_runs((numpy.where(self.__mfcc_mask))[0])
runs = _compute_runs(self.__mfcc_mask_map)
self.__speech_intervals = [(r[0], r[-1]) for r in runs]
# where( == False) not already computed, computing now
runs = _compute_runs((numpy.where(~self.__mfcc_mask))[0])
self.__nonspeech_intervals = [(r[0], r[-1]) for r in runs]
self.log(u"Storing speech and nonspeech intervals... done")
def _adjust_boundaries(self, text_map, real_wave_full_mfcc,
real_wave_length):
"""
Adjust the boundaries between consecutive fragments.
Return a pair:
1. a success bool flag
2. the computed interval map, that is,
a list of triples ``[start_time, end_time, fragment_id]``
"""
self._log("Adjusting boundaries")
algo = self.task.configuration.adjust_boundary_algorithm
value = None
if algo is None:
self._log("No adjust boundary algorithm specified: returning")
return (True, text_map)
elif algo == AdjustBoundaryAlgorithm.AUTO:
self._log("Requested adjust boundary algorithm AUTO: returning")
return (True, text_map)
elif algo == AdjustBoundaryAlgorithm.AFTERCURRENT:
value = self.task.configuration.adjust_boundary_aftercurrent_value
elif algo == AdjustBoundaryAlgorithm.BEFORENEXT:
value = self.task.configuration.adjust_boundary_beforenext_value
elif algo == AdjustBoundaryAlgorithm.OFFSET:
value = self.task.configuration.adjust_boundary_offset_value
elif algo == AdjustBoundaryAlgorithm.PERCENT:
value = self.task.configuration.adjust_boundary_percent_value
elif algo == AdjustBoundaryAlgorithm.RATE:
value = self.task.configuration.adjust_boundary_rate_value
elif algo == AdjustBoundaryAlgorithm.RATEAGGRESSIVE:
value = self.task.configuration.adjust_boundary_rate_value
self._log(["Requested algo %s and value %s", algo, value])
try:
self._log("Running VAD...")
vad = VAD(logger=self.logger)
vad.wave_mfcc = real_wave_full_mfcc
vad.wave_len = real_wave_length
vad.compute_vad()
self._log("Running VAD... done")
except Exception as e:
self._log("Adjusting boundaries: failed")
self._log(["Message: %s", str(e)])
return (False, None)
self._log("Creating AdjustBoundaryAlgorithm object")
adjust_boundary = AdjustBoundaryAlgorithm(algorithm=algo,
text_map=text_map,
speech=vad.speech,
nonspeech=vad.nonspeech,
value=value,
logger=self.logger)
self._log("Adjusting boundaries...")
adjusted_map = adjust_boundary.adjust()
self._log("Adjusting boundaries... done")
self._log("Adjusting boundaries: succeeded")
return (True, adjusted_map)
def _extract_speech(self):
""" Extract speech intervals """
self._log("Running VAD...")
vad = VAD(frame_rate=self.frame_rate, logger=self.logger)
vad.wave_len = self.audio_file.audio_length
vad.wave_mfcc = self.audio_file.audio_mfcc
vad.compute_vad()
self.audio_speech = vad.speech
self._log("Running VAD... done")
def perform(self, input_file_path, speech_length, nonspeech_length):
vad = VAD(get_abs_path(input_file_path))
vad.compute_mfcc()
vad.compute_vad()
self.assertEqual(len(vad.speech), speech_length)
self.assertEqual(len(vad.nonspeech), nonspeech_length)
def _adjust_boundaries(
self,
text_map,
real_wave_full_mfcc,
real_wave_length
):
"""
Adjust the boundaries between consecutive fragments.
Return a pair:
1. a success bool flag
2. the computed interval map, that is,
a list of triples ``[start_time, end_time, fragment_id]``
"""
self._log("Adjusting boundaries")
algo = self.task.configuration.adjust_boundary_algorithm
value = None
if algo is None:
self._log("No adjust boundary algorithm specified: returning")
return (True, text_map)
elif algo == AdjustBoundaryAlgorithm.AUTO:
self._log("Requested adjust boundary algorithm AUTO: returning")
return (True, text_map)
elif algo == AdjustBoundaryAlgorithm.AFTERCURRENT:
value = self.task.configuration.adjust_boundary_aftercurrent_value
elif algo == AdjustBoundaryAlgorithm.BEFORENEXT:
value = self.task.configuration.adjust_boundary_beforenext_value
elif algo == AdjustBoundaryAlgorithm.OFFSET:
value = self.task.configuration.adjust_boundary_offset_value
elif algo == AdjustBoundaryAlgorithm.PERCENT:
value = self.task.configuration.adjust_boundary_percent_value
elif algo == AdjustBoundaryAlgorithm.RATE:
value = self.task.configuration.adjust_boundary_rate_value
elif algo == AdjustBoundaryAlgorithm.RATEAGGRESSIVE:
value = self.task.configuration.adjust_boundary_rate_value
self._log(["Requested algo %s and value %s", algo, value])
try:
self._log("Running VAD...")
vad = VAD(logger=self.logger)
vad.wave_mfcc = real_wave_full_mfcc
vad.wave_len = real_wave_length
vad.compute_vad()
self._log("Running VAD... done")
except Exception as e:
self._log("Adjusting boundaries: failed")
self._log(["Message: %s", str(e)])
return (False, None)
self._log("Creating AdjustBoundaryAlgorithm object")
adjust_boundary = AdjustBoundaryAlgorithm(
algorithm=algo,
text_map=text_map,
speech=vad.speech,
nonspeech=vad.nonspeech,
value=value,
logger=self.logger
)
self._log("Adjusting boundaries...")
adjusted_map = adjust_boundary.adjust()
self._log("Adjusting boundaries... done")
self._log("Adjusting boundaries: succeeded")
return (True, adjusted_map)
def perform(self, input_file_path, speech_length, nonspeech_length):
vad = VAD(get_abs_path(input_file_path))
vad.compute_mfcc()
vad.compute_vad()
self.assertEqual(len(vad.speech), speech_length)
self.assertEqual(len(vad.nonspeech), nonspeech_length)
def main():
""" Entry point """
if len(sys.argv) < 4:
usage()
return
audio_file_path = sys.argv[1]
tmp_handler, tmp_file_path = tempfile.mkstemp(
suffix=".wav",
dir=gf.custom_tmp_dir()
)
mode = sys.argv[2]
output_file_path = sys.argv[3]
verbose = (sys.argv[-1] == "-v")
if mode not in ["speech", "nonspeech", "both"]:
usage()
return
if not gf.can_run_c_extension():
print "[WARN] Unable to load Python C Extensions"
print "[WARN] Running the slower pure Python code"
print "[WARN] See the README file for directions to compile the Python C Extensions"
logger = Logger(tee=verbose)
print "[INFO] Converting audio file to mono..."
converter = FFMPEGWrapper(logger=logger)
converter.convert(audio_file_path, tmp_file_path)
print "[INFO] Converting audio file to mono... done"
vad = VAD(tmp_file_path, logger=logger)
print "[INFO] Extracting MFCCs..."
vad.compute_mfcc()
print "[INFO] Extracting MFCCs... done"
print "[INFO] Executing VAD..."
vad.compute_vad()
print "[INFO] Executing VAD... done"
print "[INFO] Cleaning up..."
cleanup(tmp_handler, tmp_file_path)
print "[INFO] Cleaning up... done"
if mode == "speech":
print "[INFO] Creating speech file..."
output_file = open(output_file_path, "w")
for interval in vad.speech:
output_file.write("%.3f\t%.3f\n" % (interval[0], interval[1]))
output_file.close()
print "[INFO] Creating speech file... done"
if mode == "nonspeech":
print "[INFO] Creating nonspeech file..."
output_file = open(output_file_path, "w")
for interval in vad.nonspeech:
output_file.write("%.3f\t%.3f\n" % (interval[0], interval[1]))
output_file.close()
print "[INFO] Creating nonspeech file... done"
if mode == "both":
print "[INFO] Creating speech and nonspeech file..."
output_file = open(output_file_path, "w")
speech = [[x[0], x[1], "speech"] for x in vad.speech]
nonspeech = [[x[0], x[1], "nonspeech"] for x in vad.nonspeech]
both = sorted(speech + nonspeech)
for interval in both:
output_file.write("%.3f\t%.3f\t%s\n" % (
interval[0],
interval[1],
interval[2]
))
output_file.close()
print "[INFO] Creating speech and nonspeech file... done"
print "[INFO] Created file %s" % output_file_path
