def register(self, option_parser: OptionsParser):
self.samp_freq = option_parser.get("samp_freq",
16000,
type_function=np.int)
self.frame_shift = option_parser.get("frame_shift",
10.0,
type_function=np.float)
self.frame_length = option_parser.get("frame_length",
25.0,
type_function=np.float)
self.dither = option_parser.get("dither", 1.0, type_function=np.float)
self.preemph_coeff = option_parser.get("preemph_coeff",
0.97,
type_function=np.float)
self.remove_dc_offset = option_parser.get("remove_dc_offset",
"True",
type_function=np.bool)
self.window_type = option_parser.get("window_type",
"povey",
type_function=np.str)
self.blackman_coeff = option_parser.get("blackman_coeff",
0.42,
type_function=np.float)
self.snip_edges = option_parser.get("snip_edges",
True,
type_function=np.bool)
self.allow_downsample = option_parser.get("allow_downsample",
"False",
type_function=np.bool)
def register(self, option_parser: OptionsParser):
# self.order = option_parser.get("delta_order", 2, type_function=np.int)
# self.window = option_parser.get("delta_window", 2, type_function=np.int)
self.cmn_window = option_parser.get("cmn_window", 600, type_function=np.int)
self.min_window = option_parser.get("min_window", 100, type_function=np.int)
self.max_warnings = option_parser.get("max_warnings", 5, type_function=np.int)
self.normalize_variance = option_parser.get("normalize_variance", "False", type_function=np.bool)
self.center = option_parser.get("center", "False", type_function=np.bool)
def register(self, option_parser: OptionsParser):
self.frame_opts.register(option_parser)
self.mel_opts.register(option_parser)
self.use_energy = option_parser.get("use_energy", False, type_function=np.bool)
self.energy_floor = option_parser.get("energy_floor", 0.0, type_function=np.float)
self.raw_energy = option_parser.get("raw_energy", True, type_function=np.bool)
self.use_log_fbank = option_parser.get("use_log_fbank", True, type_function=np.bool)
self.use_power = option_parser.get("use_power", True, type_function=np.bool)
def register(self, option_parser: OptionsParser):
# self.use_energy = option_parser.get("use_energy", False, type_function=np.bool)
self.vad_energy_threshold = option_parser.get("vad_energy_threshold",
5.0,
type_function=np.float)
self.vad_energy_mean_scale = option_parser.get("vad_energy_mean_scale",
0.5,
type_function=np.float)
self.vad_frames_context = option_parser.get("vad_frames_context",
0,
type_function=np.int)
self.vad_proportion_threshold = option_parser.get(
"vad_proportion_threshold", 0.6, type_function=np.float)
def register(self, option_parser: OptionsParser):
self.frame_opts.register(option_parser)
self.mel_opts.register(option_parser)
self.num_ceps = option_parser.get("num_ceps", 13, type_function=np.int)
self.use_energy = option_parser.get("use_energy",
"True",
type_function=np.bool)
self.energy_floor = option_parser.get("energy_floor",
0.0,
type_function=np.float) # 0.0
self.raw_energy = option_parser.get("raw_energy",
"True",
type_function=np.bool)
self.cepstral_lifter = option_parser.get("cepstral_lifter",
22.0,
type_function=np.float)
def main(args):
data_path = args.data_path
save_path = args.save_path
config_file = args.config_file
config_section = args.config_section
option_parser = OptionsParser(conf_file=config_file,
conf_section=config_section)
vad_opts = VadEnergyOptions()
vad_opts.register(option_parser)
feats_scp = os.path.join(data_path, "feats.scp")
feature_reader = FeatureReader(feats_scp)
vad_writer = FeatureWriter(save_path, split_num=1)
num_done = 0
num_err = 0
num_unvoiced = 0
tot_length = 0.0
tot_decision = 0.0
for (utt_id, utt_feat) in feature_reader:
if utt_feat.shape[0] == 0:
logging.warning(f"Empty feature matrix for utterance {utt_id}")
num_err += 1
continue
vad_result = compute_vad_energy(vad_opts, utt_feat)
vad_sum = np.sum(vad_result)
if vad_sum == 0.0:
logging.warning(
f"No frames were judged voiced for utterance {utt_id}")
num_unvoiced += 1
else:
num_done += 1
tot_decision += np.sum(vad_result)
tot_length += vad_result.shape[0]
if vad_sum != 0.0:
vad_writer.write(utt_id, vad_result)
logging.info(f"Applied energy based voice activity detection "
f"{num_done} utterances successfully"
f"{num_err} had empty features, and "
f"{num_unvoiced} were completely unvoiced.")
logging.info(f"Proportion of voiced frames was"
f"{float(tot_decision) / float(tot_length)} over"
f"{tot_length} frames")
def main(args):
data_path = args.data_path
save_path = args.save_path
config_file = args.config_file
config_section = args.config_section
wav_reader = WavReader(data_path)
option_parser = OptionsParser(conf_file=config_file, conf_section=config_section)
feature_extractor = FeatureExtractor(feature_type="fbank", option_parser=option_parser)
feature_writer = FeatureWriter(save_path, split_num=1)
for utt_id, (wav, sample_rate) in wav_reader:
# print(f"waveform:\n{wav}\n")
result = feature_extractor.compute_features(wav, sample_rate)
print(f"utt_id = {utt_id}, result.shape = {result.shape}")
feature_writer.write(utt_id, result)
feature_writer.flush()
def register(self, option_parser: OptionsParser):
self.frame_opts.register(option_parser)
self.energy_floor = option_parser.get("energy_floor", 0.0, np.float)
self.raw_energy = option_parser.get("raw_energy", "True", np.bool)
def register(self, option_parser: OptionsParser):
self.num_bins = option_parser.get("num_mel_bins", 25, type_function=np.int)
self.low_freq = option_parser.get("low_freq", 20.0, type_function=np.float)
self.high_freq = option_parser.get("high_freq", 7600.0, type_function=np.float)
def register(self, option_parser: OptionsParser):
self.order = option_parser.get("delta_order", 2, type_function=np.int)
self.window = option_parser.get("delta_window", 2, type_function=np.int)