def __init__(self, vad_cfg, speech_thresh=0.7, non_speech_thresh=0.1):
self.vad_cfg = vad_cfg
self.speech_thresh = speech_thresh
self.non_speech_thresh = non_speech_thresh
logging.info('Loading VAD model.')
self.vad = FFNNVAD(**vad_cfg)
def __init__(self, vad_cfg, speech_thresh=0.7, non_speech_thresh=0.1):
self.vad_cfg = vad_cfg
self.speech_thresh = speech_thresh
self.non_speech_thresh = non_speech_thresh
logging.info('Loading VAD model.')
self.vad = FFNNVAD(**vad_cfg)
class RecordingSplitter(object):
CHANGE_TO_NON_SPEECH = 2
CHANGE_TO_SPEECH = 1
speech_thresh = 0.7
non_speech_thresh = 0.1
read_buffer_size = 128
def __init__(self, vad_cfg, speech_thresh=0.7, non_speech_thresh=0.1):
self.vad_cfg = vad_cfg
self.speech_thresh = speech_thresh
self.non_speech_thresh = non_speech_thresh
logging.info('Loading VAD model.')
self.vad = FFNNVAD(**vad_cfg)
def split_single_channel_wav(self, file_name, out_dir, out_prefix):
logging.info('Splitting %s' % file_name)
wave_in = wave.open(file_name)
sample_rate = wave_in.getframerate()
sample_width = wave_in.getsampwidth()
bytes_per_second = sample_rate * sample_width
frames_per_second = bytes_per_second / self.read_buffer_size
(detection_window_sil, detection_window_speech,
pre_detection_buffer) = self._initialize_buffers(frames_per_second)
res_files = []
res_file_cntr = 0
frames = []
is_speech = False
n_read = 0
n_read_beg = None
while 1:
audio_data = wave_in.readframes(self.read_buffer_size)
n_read += self.read_buffer_size
if len(audio_data) == 0:
break
raw_vad_decision = self.vad.decide(audio_data)
is_speech, change = self._smoothe_decison(raw_vad_decision,
is_speech,
detection_window_speech,
detection_window_sil)
if not is_speech:
pre_detection_buffer.append(audio_data)
if change == self.CHANGE_TO_SPEECH:
n_read_beg = n_read - self.read_buffer_size
frames = []
elif change == self.CHANGE_TO_NON_SPEECH:
#if not is_speech and len(frames) > 1:
self._save_part(res_file_cntr,
list(pre_detection_buffer) + frames, out_dir,
res_files, wave_in, out_prefix, n_read_beg,
n_read, bytes_per_second)
res_file_cntr += 1
pre_detection_buffer.extend(
frames[-pre_detection_buffer.maxlen:])
if is_speech:
frames.append(audio_data)
if n_read_beg:
self._save_part(res_file_cntr, frames, out_dir, res_files, wave_in,
out_prefix, n_read_beg, n_read, bytes_per_second)
return res_files
def _initialize_buffers(self, frames_per_second):
pre_detection_buffer_frames = int(frames_per_second * 0.5)
smoothe_decision_window_sil = int(frames_per_second * 0.2)
smoothe_decision_window_speech = int(frames_per_second * 0.2)
detection_window_speech = deque(maxlen=smoothe_decision_window_speech)
detection_window_sil = deque(maxlen=smoothe_decision_window_sil)
pre_detection_buffer = deque(maxlen=pre_detection_buffer_frames)
return detection_window_sil, detection_window_speech, pre_detection_buffer
def _smoothe_decison(self, decision, last_vad, detection_window_speech,
detection_window_sil):
detection_window_speech.append(decision)
detection_window_sil.append(decision)
speech = float(sum(detection_window_speech)) / (
len(detection_window_speech) + 1.0)
sil = float(
sum(detection_window_sil)) / (len(detection_window_sil) + 1.0)
vad = last_vad
change = None
if last_vad:
# last decision was speech
if sil < self.non_speech_thresh:
vad = False
change = self.CHANGE_TO_NON_SPEECH
else:
if speech > self.speech_thresh:
vad = True
change = self.CHANGE_TO_SPEECH
return vad, change
def _save_part(self, cntr, frames, out_dir, res_files, wave_in, out_prefix,
n_read_beg, n_read_end, bytes_per_second):
content = b''.join(frames)
logging.info('Saving part %d (%.1f s).' %
(cntr, len(content) * 1.0 / bytes_per_second))
res_file = os.path.join(out_dir, 'part.%s.%.3d.wav' % (
out_prefix,
cntr,
))
wf = wave.open(res_file, 'wb')
wf.setnchannels(wave_in.getnchannels())
wf.setsampwidth(wave_in.getsampwidth())
wf.setframerate(wave_in.getframerate())
wf.writeframes(content)
wf.close()
res_files.append(((n_read_beg * 1.0 / bytes_per_second,
n_read_end * 1.0 / bytes_per_second), res_file))
class RecordingSplitter(object):
CHANGE_TO_NON_SPEECH = 2
CHANGE_TO_SPEECH = 1
speech_thresh = 0.7
non_speech_thresh = 0.1
read_buffer_size = 128
def __init__(self, vad_cfg, speech_thresh=0.7, non_speech_thresh=0.1):
self.vad_cfg = vad_cfg
self.speech_thresh = speech_thresh
self.non_speech_thresh = non_speech_thresh
logging.info('Loading VAD model.')
self.vad = FFNNVAD(**vad_cfg)
def split_single_channel_wav(self, file_name, out_dir, out_prefix):
logging.info('Splitting %s' % file_name)
wave_in = wave.open(file_name)
sample_rate = wave_in.getframerate()
sample_width = wave_in.getsampwidth()
bytes_per_second = sample_rate * sample_width
frames_per_second = bytes_per_second / self.read_buffer_size
(detection_window_sil,
detection_window_speech,
pre_detection_buffer) = self._initialize_buffers(frames_per_second)
res_files = []
res_file_cntr = 0
frames = []
is_speech = False
n_read = 0
n_read_beg = None
while 1:
audio_data = wave_in.readframes(self.read_buffer_size)
n_read += self.read_buffer_size
if len(audio_data) == 0:
break
raw_vad_decision = self.vad.decide(audio_data)
is_speech, change = self._smoothe_decison(raw_vad_decision, is_speech, detection_window_speech, detection_window_sil)
if not is_speech:
pre_detection_buffer.append(audio_data)
if change == self.CHANGE_TO_SPEECH:
n_read_beg = n_read - self.read_buffer_size
frames = []
elif change == self.CHANGE_TO_NON_SPEECH:
#if not is_speech and len(frames) > 1:
self._save_part(res_file_cntr, list(pre_detection_buffer) + frames, out_dir, res_files, wave_in, out_prefix, n_read_beg, n_read, bytes_per_second)
res_file_cntr += 1
pre_detection_buffer.extend(frames[-pre_detection_buffer.maxlen:])
if is_speech:
frames.append(audio_data)
if n_read_beg:
self._save_part(res_file_cntr, frames, out_dir, res_files, wave_in, out_prefix, n_read_beg, n_read, bytes_per_second)
return res_files
def _initialize_buffers(self, frames_per_second):
pre_detection_buffer_frames = int(frames_per_second * 0.5)
smoothe_decision_window_sil = int(frames_per_second * 0.2)
smoothe_decision_window_speech = int(frames_per_second * 0.2)
detection_window_speech = deque(maxlen=smoothe_decision_window_speech)
detection_window_sil = deque(maxlen=smoothe_decision_window_sil)
pre_detection_buffer = deque(maxlen=pre_detection_buffer_frames)
return detection_window_sil, detection_window_speech, pre_detection_buffer
def _smoothe_decison(self, decision, last_vad, detection_window_speech, detection_window_sil):
detection_window_speech.append(decision)
detection_window_sil.append(decision)
speech = float(sum(detection_window_speech)) / (len(detection_window_speech) + 1.0)
sil = float(sum(detection_window_sil)) / (len(detection_window_sil) + 1.0)
vad = last_vad
change = None
if last_vad:
# last decision was speech
if sil < self.non_speech_thresh:
vad = False
change = self.CHANGE_TO_NON_SPEECH
else:
if speech > self.speech_thresh:
vad = True
change = self.CHANGE_TO_SPEECH
return vad, change
def _save_part(self, cntr, frames, out_dir, res_files, wave_in, out_prefix, n_read_beg, n_read_end, bytes_per_second):
content = b''.join(frames)
logging.info('Saving part %d (%.1f s).' % (cntr, len(content) * 1.0 / bytes_per_second))
res_file = os.path.join(out_dir, 'part.%s.%.3d.wav' % (out_prefix, cntr, ))
wf = wave.open(res_file, 'wb')
wf.setnchannels(wave_in.getnchannels())
wf.setsampwidth(wave_in.getsampwidth())
wf.setframerate(wave_in.getframerate())
wf.writeframes(content)
wf.close()
res_files.append(((n_read_beg * 1.0 / bytes_per_second, n_read_end * 1.0 / bytes_per_second), res_file))
