def compute_wav_path(wav, feat_scp, feat_ark, utt2dur, utt2num_frames):
feat, duration = Make_Spect(wav_path=wav[1], windowsize=0.02, stride=0.01, duration=True)
# np_fbank = Make_Fbank(filename=uid2path[uid], use_energy=True, nfilt=c.TDNN_FBANK_FILTER)
len_vec = len(feat.tobytes())
key = wav[0]
kaldi_io.write_mat(feat_ark, feat, key=key)
feat_scp.write(str(key) + ' ' + str(feat_ark.name) + ':' + str(feat_ark.tell() - len_vec - 15) + '\n')
utt2dur.write('%s %.6f\n' % (str(key), duration))
utt2num_frames.write('%s %d\n' % (str(key), len(feat)))
def MakeFeatsProcess(out_dir, proid, t_queue, e_queue):
# wav_scp = os.path.join(data_path, 'wav.scp')
feat_scp = os.path.join(out_dir, 'feat.%d.scp' % proid)
feat_ark = os.path.join(out_dir, 'feat.%d.ark' % proid)
utt2dur = os.path.join(out_dir, 'utt2dur.%d' % proid)
utt2num_frames = os.path.join(out_dir, 'utt2num_frames.%d' % proid)
feat_scp = open(feat_scp, 'w')
feat_ark = open(feat_ark, 'wb')
utt2dur = open(utt2dur, 'w')
utt2num_frames = open(utt2num_frames, 'w')
while not t_queue.empty():
wav = t_queue.get()
pair = wav.split()
try:
feat, duration = Make_Spect(wav_path=pair[1],
windowsize=0.02,
stride=0.01,
duration=True)
# np_fbank = Make_Fbank(filename=uid2path[uid], use_energy=True, nfilt=c.TDNN_FBANK_FILTER)
len_vec = len(feat.tobytes())
key = pair[0]
kaldi_io.write_mat(feat_ark, feat, key='')
feat_scp.write(
str(key) + ' ' + str(feat_ark.name) + ':' +
str(feat_ark.tell() - len_vec - 15) + '\n')
utt2dur.write('%s %.6f' % (str(key), duration))
utt2num_frames.write('%s %d' % (str(key), len(feat)))
except:
print("Error: %s" % pair[0])
e_queue.put(pair[0])
# if self.queue.qsize() % 1000 == 0:
print('==> Process %s: %s left' % (str(proid), str(t_queue.qsize())))
feat_scp.close()
feat_ark.close()
utt2dur.close()
utt2num_frames.close()
print('>> Process {} finished!'.format(proid))
def compute_wav_path(wav, feat_scp, feat_path, utt2dur, utt2num_frames):
feat, duration = Make_Spect(wav_path=wav[1],
windowsize=0.02,
stride=0.01,
duration=True)
# np_fbank = Make_Fbank(filename=uid2path[uid], use_energy=True, nfilt=c.TDNN_FBANK_FILTER)
key = wav[0]
# pdb.set_trace()
save_path = os.path.join(feat_path, wav[0] + '.npy')
# print('save path:' + save_path)
np.save(save_path, feat)
feat_scp.write(str(key) + ' ' + save_path + '\n')
utt2dur.write('%s %.6f\n' % (str(key), duration))
utt2num_frames.write('%s %d\n' % (str(key), len(feat)))
def MakeFeatsProcess(out_dir, proid, t_queue, e_queue):
# wav_scp = os.path.join(data_path, 'wav.scp')
feat_scp = os.path.join(out_dir, 'feat.%d.scp' % proid)
feat_path = os.path.join(out_dir, 'feats.%d' % proid)
if not os.path.exists(feat_path):
os.makedirs(feat_path)
utt2dur = os.path.join(out_dir, 'utt2dur.%d' % proid)
utt2num_frames = os.path.join(out_dir, 'utt2num_frames.%d' % proid)
feat_scp = open(feat_scp, 'w')
utt2dur = open(utt2dur, 'w')
utt2num_frames = open(utt2num_frames, 'w')
while not t_queue.empty():
comm = task_queue.get()
# print('111')
pair = comm.split()
key = pair[0]
try:
feat, duration = Make_Spect(wav_path=pair[1],
windowsize=0.02,
stride=0.01,
duration=True)
# np_fbank = Make_Fbank(filename=uid2path[uid], use_energy=True, nfilt=c.TDNN_FBANK_FILTER)
save_path = os.path.join(feat_path, key + '.npy')
np.save(save_path, feat)
feat_scp.write(str(key) + ' ' + save_path + '\n')
utt2dur.write('%s %.6f\n' % (str(key), duration))
utt2num_frames.write('%s %d\n' % (str(key), len(feat)))
except:
e_queue.put(key)
if t_queue.qsize() % 100 == 0:
print(
'\rProcess [%3s] There are [%6s] utterances left, with [%6s] errors.'
% (str(proid), str(t_queue.qsize()), str(e_queue.qsize())),
end='')
feat_scp.close()
utt2dur.close()
utt2num_frames.close()
# command = command.rstrip(' - ') + " " + temp_wav
# print(command)
spid, stdout, error = RunCommand(command)
# os.waitpid(spid, 0)
# with open(temp_wav, 'wb') as wav_f:
# wav_f.write(stdout)
>>>>>>> Server/Server
if args.feat_type == 'fbank':
feat, duration = Make_Fbank(filename=temp_wav, filtertype=args.filter_type, use_energy=True,
lowfreq=args.lowfreq, log_scale=args.log_scale,
nfft=args.nfft, nfilt=args.filters, normalize=args.normalize,
duration=True, windowsize=args.windowsize,
multi_weight=args.multi_weight)
elif args.feat_type == 'spectrogram':
feat, duration = Make_Spect(wav_path=temp_wav, windowsize=args.windowsize,
<<<<<<< HEAD
lowfreq=args.lowfreq, stride=args.stride, duration=True,
nfft=args.nfft, normalize=args.normalize)
elif args.feat_type == 'mfcc':
feat, duration = Make_MFCC(filename=temp_wav, numcep=args.numcep, nfilt=args.filters,
lowfreq=args.lowfreq, normalize=args.normalize, duration=True,
use_energy=True)
=======
bandpass=args.bandpass, lowfreq=args.lowfreq,
highfreq=args.highfreq,
log_scale=args.log_scale,
stride=args.stride, duration=True, nfft=args.nfft,
normalize=args.normalize)
elif args.feat_type == 'mfcc':
feat, duration = Make_MFCC(filename=temp_wav, numcep=args.numcep, nfilt=args.filters,
def MakeFeatsProcess(lock, out_dir, ark_dir, ark_prefix, proid, t_queue,
e_queue):
# wav_scp = os.path.join(data_path, 'wav.scp')
feat_scp = os.path.join(out_dir, 'feat.%d.temp.scp' % proid)
utt2dur = os.path.join(out_dir, 'utt2dur.%d' % proid)
utt2num_frames = os.path.join(out_dir, 'utt2num_frames.%d' % proid)
feat_scp_f = open(feat_scp, 'w')
utt2dur_f = open(utt2dur, 'w')
if args.feat_format == 'kaldi':
feat_ark = os.path.join(out_dir,
'%s_feat.%d.ark' % (ark_prefix, proid))
feat_ark_f = open(feat_ark, 'wb')
utt2num_frames_f = open(utt2num_frames, 'w')
feat_dir = os.path.join(ark_dir, ark_prefix)
if not os.path.exists(feat_dir):
os.makedirs(feat_dir)
temp_dir = out_dir + '/temp'
if not os.path.exists(temp_dir):
os.makedirs(temp_dir)
while True:
lock.acquire() # 加上锁
if not t_queue.empty():
comm = task_queue.get()
lock.release() # 释放锁
pair = comm.split()
key = pair[0]
try:
if len(pair) > 2:
command = ' '.join(pair[1:])
if command.endswith('|'):
command = command.rstrip('|')
spid, stdout, error = RunCommand(command)
# os.waitpid(spid, 0)
temp_wav = temp_dir + '/%s.%s' % (key, args.data_format)
with open(temp_wav, 'wb') as wav_f:
wav_f.write(stdout)
if args.feat_type == 'fbank':
feat, duration = Make_Fbank(
filename=temp_wav,
filtertype=args.filter_type,
use_energy=True,
lowfreq=args.lowfreq,
nfft=args.nfft,
nfilt=args.filters,
normalize=args.normalize,
duration=True,
windowsize=args.windowsize,
multi_weight=args.multi_weight)
elif args.feat_type == 'spectrogram':
feat, duration = Make_Spect(wav_path=temp_wav,
windowsize=args.windowsize,
lowfreq=args.lowfreq,
stride=args.stride,
duration=True,
nfft=args.nfft,
normalize=args.normalize)
elif args.feat_type == 'mfcc':
feat, duration = Make_MFCC(filename=temp_wav,
numcep=args.numcep,
nfilt=args.filters,
lowfreq=args.lowfreq,
normalize=args.normalize,
duration=True,
use_energy=True)
os.remove(temp_wav)
else:
if args.feat_type == 'fbank':
feat, duration = Make_Fbank(
filename=pair[1],
filtertype=args.filter_type,
use_energy=True,
nfft=args.nfft,
windowsize=args.windowsize,
lowfreq=args.lowfreq,
nfilt=args.filters,
duration=True,
normalize=args.normalize,
multi_weight=args.multi_weight)
elif args.feat_type == 'spectrogram':
feat, duration = Make_Spect(wav_path=pair[1],
windowsize=args.windowsize,
lowfreq=args.lowfreq,
stride=args.stride,
duration=True,
nfft=args.nfft,
normalize=args.normalize)
elif args.feat_type == 'mfcc':
feat, duration = Make_MFCC(filename=pair[1],
numcep=args.numcep,
nfilt=args.filters,
lowfreq=args.lowfreq,
normalize=args.normalize,
duration=True,
use_energy=True)
# feat = np.load(pair[1]).astype(np.float32)
feat = feat.astype(np.float32)
if args.feat_format == 'kaldi':
kaldi_io.write_mat(feat_ark_f, feat, key='')
offsets = feat_ark + ':' + str(feat_ark_f.tell() -
len(feat.tobytes()) - 15)
# print(offsets)
feat_scp_f.write(key + ' ' + offsets + '\n')
elif args.feat_format == 'npy':
npy_path = os.path.join(feat_dir, '%s.npy' % key)
np.save(npy_path, feat)
feat_scp_f.write(key + ' ' + npy_path + '\n')
utt2dur_f.write('%s %.6f\n' % (key, duration))
utt2num_frames_f.write('%s %d\n' % (key, len(feat)))
except Exception as e:
print(e)
e_queue.put(key)
# if t_queue.qsize() % 100 == 0:
print('\rProcess [%6s] There are [%6s] utterances' \
' left, with [%6s] errors.' % (str(os.getpid()), str(t_queue.qsize()), str(e_queue.qsize())),
end='')
else:
lock.release() # 释放锁
# print('\n>> Process {}: queue empty!'.format(os.getpid()))
break
feat_scp_f.close()
utt2dur_f.close()
if args.feat_format == 'kaldi':
feat_ark_f.close()
utt2num_frames_f.close()
new_feat_scp = os.path.join(out_dir, 'feat.%d.scp' % proid)
if args.feat_format == 'kaldi' and args.compress:
new_feat_ark = os.path.join(feat_dir, 'feat.%d.ark' % proid)
compress_command = "copy-feats --compress=true scp:{} ark,scp:{},{}".format(
feat_scp, new_feat_ark, new_feat_scp)
pid, stdout, stderr = RunCommand(compress_command)
# print(stdout)
if os.path.exists(new_feat_scp) and os.path.exists(new_feat_ark):
os.remove(feat_ark)
else:
shutil.copy(feat_scp, new_feat_scp)
@Time: 2020/11/29 20:31
@Overview:
"""
import matplotlib.pyplot as plt
from Process_Data.Compute_Feat.compute_vad import ComputeVadEnergy
from Process_Data.audio_processing import Make_Spect
import numpy as np
# v1 = 'Data/dataset/voxceleb1/8k_radio_v3/id10001/1zcIwhmdeo4/00001.wav'
v1 = 'Data/dataset/aishell-2/data/C0001/IC0001W0001-8k.wav'
a1 = 'Data/dataset/wav_test/01-yangxiaokang/tmp_0001-U000013_15s_8k.wav'
r1 = 'Data/dataset/wav_test/00-yangwenhao/tmp_0009-U000000_15s_8k.wav'
v1_spect = Make_Spect(v1,
windowsize=0.02,
stride=0.01,
nfft=320,
normalize=False)
a1_spect = Make_Spect(a1,
windowsize=0.02,
stride=0.01,
nfft=320,
normalize=False)
r1_spect = Make_Spect(r1,
windowsize=0.02,
stride=0.01,
nfft=320,
normalize=False)
v1_energy = np.log(np.exp(v1_spect).sum(axis=1)).reshape(-1, 1)
a1_energy = np.log(np.exp(a1_spect).sum(axis=1)).reshape(-1, 1)