def reco_files(self):
fnames = QFileDialog.getOpenFileNames(self, "Select Wav Files", "", "Files (*.wav)")
print 'reco_files'
for f in fnames:
fs, sig = read_wav(f)
newsig = self.backend.filter(fs, sig)
label = self.backend.predict(fs, newsig)
print f, label
def enroll_file(self):
fname = QFileDialog.getOpenFileName(self, "Open Wav File", "", "Files (*.wav)")
if not fname:
return
self.status(fname)
self.enrollFileName.setText(fname)
fs, signal = read_wav(fname)
signal = monophonic(signal)
self.enrollWav = (fs, signal)
def reco_file(self):
fname = QFileDialog.getOpenFileName(self, "Open Wav File", "", "Files (*.wav)")
print 'reco_file'
if not fname:
return
self.status(fname)
fs, signal = read_wav(fname)
self.reco_do_predict(fs, signal)
def __init__(self):
self.backend = ModelInterface.load(self.INPUT_MODEL)
try:
fs, signal = read_wav(self.BG)
self.backend.init_noise(fs, signal)
except:
print "file not found!"
self.pub = rospy.Publisher('/speaker',String,queue_size = 10)
self.sub = rospy.Subscriber('/wav',numpy_msg(Floats),self.task_predict)
def __init__(self, parent=None):
QWidget.__init__(self, parent)
uic.loadUi("edytor.ui", self)
self.statusBar()
self.timer = QTimer(self)
self.timer.timeout.connect(self.timer_callback)
self.noiseButton.clicked.connect(self.noise_clicked)
self.recording_noise = False
self.loadNoise.clicked.connect(self.load_noise)
self.enrollRecord.clicked.connect(self.start_enroll_record)
self.stopEnrollRecord.clicked.connect(self.stop_enroll_record)
self.enrollFile.clicked.connect(self.enroll_file)
self.enroll.clicked.connect(self.do_enroll)
self.startTrain.clicked.connect(self.start_train)
self.dumpBtn.clicked.connect(self.dump)
self.loadBtn.clicked.connect(self.load)
self.recoRecord.clicked.connect(self.start_reco_record)
self.stopRecoRecord.clicked.connect(self.stop_reco_record)
# self.newReco.clicked.connect(self.new_reco)
self.recoFile.clicked.connect(self.reco_file)
self.recoInputFiles.clicked.connect(self.reco_files)
#UI.init
self.userdata =[]
self.loadUsers()
self.Userchooser.currentIndexChanged.connect(self.showUserInfo)
self.ClearInfo.clicked.connect(self.clearUserInfo)
self.UpdateInfo.clicked.connect(self.updateUserInfo)
self.UploadImage.clicked.connect(self.upload_avatar)
#movie test
self.movie = QMovie(u"image/recording.gif")
self.movie.start()
self.movie.stop()
self.Animation.setMovie(self.movie)
self.Animation_2.setMovie(self.movie)
self.Animation_3.setMovie(self.movie)
self.aladingpic = QPixmap(u"image/a_hello.png")
self.Alading.setPixmap(self.aladingpic)
self.Alading_conv.setPixmap(self.aladingpic)
#default user image setting
self.avatarname = "image/nouser.jpg"
self.defaultimage = QPixmap(self.avatarname)
self.Userimage.setPixmap(self.defaultimage)
self.recoUserImage.setPixmap(self.defaultimage)
self.convUserImage.setPixmap(self.defaultimage)
self.load_avatar('avatar/')
# Graph Window init
self.graphwindow = GraphWindow()
self.newname = ""
self.lastname = ""
self.Graph_button.clicked.connect(self.graphwindow.show)
self.convRecord.clicked.connect(self.start_conv_record)
self.convStop.clicked.connect(self.stop_conv)
self.backend = ModelInterface()
# debug
QShortcut(QKeySequence("Ctrl+P"), self, self.printDebug)
#init
try:
fs, signal = read_wav("bg.wav")
self.backend.init_noise(fs, signal)
except:
pass
def load_noise(self):
fname = QFileDialog.getOpenFileName(self, "Open Data File:", "", "Wav File (*.wav)")
if fname:
fs, signal = read_wav(fname)
self.backend.init_noise(fs, signal)
for g in tqdm(GENRE):
# print(GENRE)
FILES = glob(DB + '/wav/' + g + '/*.wav')
label, pred_t1, pred_t2, P_score, ALOTC_score = list(), list(), list(), list(), list()
for f in FILES:
f = f.replace('\\', '/')
# print('FILE:', f)
# Read the labeled tempo(ground-truth tempo)
bpm = float(utils.read_tempofile(DB, f))
# print(bpm)
label.append(bpm)
# Compute local onset autocorrelation
sr, y = utils.read_wav(f)
hop_length = 512
onset_env = librosa.onset.onset_strength(y=y, sr=sr, hop_length=hop_length, n_fft=2048)
# tempogram = librosa.feature.tempogram(onset_envelope=onset_env, sr=sr, hop_length=hop_length)
tempogram = librosa.feature.tempogram(onset_envelope=onset_env, sr=sr, hop_length=hop_length)
# predict the tempo1(slower one), tempo2(faster one)
# tempo1, tempo2 = librosa.beat.tempo(onset_envelope=onset_env, sr=sr, hop_length=hop_length)
tempo1, tempo2 = utils.tempo(onset_envelope=onset_env, sr=sr, hop_length=hop_length)
pred_t1.append(tempo1)
pred_t2.append(tempo2)
# print(tempo1, tempo2)
# p score
s1 = tempo1/(tempo1+tempo2)
s2 = 1.0 - s1
def open_recording(self,
file_path,
speakers,
side=None,
silence_length=2.0):
"""Open combined recording and splits it into separate speaker-ear pairs.
Args:
file_path: Path to recording file.
speakers: Sequence of recorded speakers.
side: Which side (ear) tracks are contained in the file if only one. "left" or "right" or None for both.
silence_length: Length of silence used during recording in seconds.
Returns:
None
"""
if self.fs != self.estimator.fs:
raise ValueError(
'Refusing to open recording because HRIR\'s sampling rate doesn\'t match impulse response '
'estimator\'s sampling rate.')
fs, recording = read_wav(file_path, expand=True)
if fs != self.fs:
raise ValueError(
'Sampling rate of recording must match sampling rate of test signal.'
)
if silence_length * self.fs != int(silence_length * self.fs):
raise ValueError(
'Silence length must produce full samples with given sampling rate.'
)
silence_length = int(silence_length * self.fs)
# 2 tracks per speaker when side is not specified, only 1 track per speaker when it is
tracks_k = 2 if side is None else 1
# Number of speakers in each track
n_columns = round(len(speakers) / (recording.shape[0] // tracks_k))
# Crop out initial silence
recording = recording[:, silence_length:]
# Split sections in time to columns
columns = []
column_size = silence_length + len(self.estimator)
for i in range(n_columns):
columns.append(recording[:, i * column_size:(i + 1) * column_size])
# Split each track by columns
i = 0
while i < recording.shape[0]:
for j, column in enumerate(columns):
n = int(i // 2 * len(columns) + j)
speaker = speakers[n]
if speaker not in SPEAKER_NAMES:
# Skip non-standard speakers. Useful for skipping the other sweep in center channel recording.
continue
if speaker not in self.irs:
self.irs[speaker] = dict()
if side is None:
# Left first, right then
self.irs[speaker]['left'] = ImpulseResponse(
self.estimator.estimate(column[i, :]), self.fs,
column[i, :])
self.irs[speaker]['right'] = ImpulseResponse(
self.estimator.estimate(column[i + 1, :]), self.fs,
column[i + 1, :])
else:
# Only the given side
self.irs[speaker][side] = ImpulseResponse(
self.estimator.estimate(column[i, :]), self.fs,
column[i, :])
i += tracks_k
def task_predict(input_files, input_model):
m = ModelInterface.load(input_model)
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
label, score = m.predict(fs, signal)
print(f, '->', label, ", score->", score)
def __init__(self, parent=None):
QWidget.__init__(self, parent)
uic.loadUi("edytor.ui", self)
self.statusBar()
self.timer = QTimer(self)
self.timer.timeout.connect(self.timer_callback)
self.noiseButton.clicked.connect(self.noise_clicked)
self.recording_noise = False
self.loadNoise.clicked.connect(self.load_noise)
self.enrollRecord.clicked.connect(self.start_enroll_record)
self.stopEnrollRecord.clicked.connect(self.stop_enroll_record)
self.enrollFile.clicked.connect(self.enroll_file)
self.enroll.clicked.connect(self.do_enroll)
self.startTrain.clicked.connect(self.start_train)
self.dumpBtn.clicked.connect(self.dump)
self.loadBtn.clicked.connect(self.load)
self.recoRecord.clicked.connect(self.start_reco_record)
self.stopRecoRecord.clicked.connect(self.stop_reco_record)
# self.newReco.clicked.connect(self.new_reco)
self.recoFile.clicked.connect(self.reco_file)
self.recoInputFiles.clicked.connect(self.reco_files)
#UI.init
self.userdata = []
self.loadUsers()
self.Userchooser.currentIndexChanged.connect(self.showUserInfo)
self.ClearInfo.clicked.connect(self.clearUserInfo)
self.UpdateInfo.clicked.connect(self.updateUserInfo)
self.UploadImage.clicked.connect(self.upload_avatar)
#movie test
self.movie = QMovie(u"image/recording.gif")
self.movie.start()
self.movie.stop()
self.Animation.setMovie(self.movie)
self.Animation_2.setMovie(self.movie)
self.Animation_3.setMovie(self.movie)
self.aladingpic = QPixmap(u"image/a_hello.png")
self.Alading.setPixmap(self.aladingpic)
self.Alading_conv.setPixmap(self.aladingpic)
#default user image setting
self.avatarname = "image/nouser.jpg"
self.defaultimage = QPixmap(self.avatarname)
self.Userimage.setPixmap(self.defaultimage)
self.recoUserImage.setPixmap(self.defaultimage)
self.convUserImage.setPixmap(self.defaultimage)
self.load_avatar('avatar/')
# Graph Window init
self.graphwindow = GraphWindow()
self.newname = ""
self.lastname = ""
self.Graph_button.clicked.connect(self.graphwindow.show)
self.convRecord.clicked.connect(self.start_conv_record)
self.convStop.clicked.connect(self.stop_conv)
self.backend = ModelInterface()
# debug
QShortcut(QKeySequence("Ctrl+P"), self, self.printDebug)
#init
try:
fs, signal = read_wav("bg.wav")
self.backend.init_noise(fs, signal)
except:
pass
def load_noise(self):
fname = QFileDialog.getOpenFileName(self, "Open Data File:", "",
"Wav File (*.wav)")
if fname:
fs, signal = read_wav(fname)
self.backend.init_noise(fs, signal)
def reverberate_and_mix(out_folder,
sources_folder,
rir_folder,
mix_info,
scale_rirs=10.0,
part=0,
nparts=8,
chat=True):
"""Reverberate and mix sources."""
list_mix = sorted(mix_info.keys())
list_len = len(list_mix)
partsize = list_len // nparts
assert part < nparts
start = part * partsize
end = list_len if part == nparts - 1 else (part + 1) * partsize
if start == end:
raise ValueError('Not enough mixtures to generate. Part {} of {} to '
'generate a total of {} mixtures.'.format(
part, nparts, list_len))
print('Reverberating and mixing from {} to {} '
'out of {}.'.format(start, end, list_len))
for mix in list_mix[start:end]:
sources, rirs = mix_info[mix]
mix_to_data = []
max_src_len = -1
if chat:
print('--\n{} ='.format(mix))
for source, rir in zip(sources, rirs):
source_path = os.path.join(sources_folder, source)
src_data, samplerate_src = read_wav(source_path, always_2d=True)
rir_path = os.path.join(rir_folder, rir)
rir_data, samplerate_rir = read_wav(rir_path, always_2d=True)
assert samplerate_src == samplerate_rir
# Pick channel 0 of src_data.
src_data = src_data[:, 0]
# Pick channel 0 of rirs and scale it.
rir_data = scale_rirs * rir_data[:, 0]
rir_len = len(rir_data)
src_len = len(src_data)
rir_max = np.max(np.abs(rir_data))
src_max = np.max(np.abs(src_data))
max_src_len = np.maximum(src_len, max_src_len)
if chat:
print('+ {} [{}, {:1.2f}] * {} [{}, {:1.2f}]'.format(
source, src_len, src_max, rir, rir_len, rir_max))
mix_to_data.append([src_data, rir_data, source, rir])
mix_rev_sources = []
rir_paths_used = []
for data in mix_to_data:
src_data, rir_data, source_relpath, rir_relpath = data
rir_paths_used.append(rir_relpath)
src_len = len(src_data)
if src_len < max_src_len:
print('WARNING: original source data has {} samples '
'for source file {}, zero padding '
'to size {}.'.format(src_len, source_relpath,
max_src_len))
src_data = np.concatenate(
(src_data, np.zeros(max_src_len - src_len)), axis=0)
rev_src_data = np.convolve(src_data, rir_data, 'same')
# Write reverberated source data.
rev_src_path = os.path.join(out_folder, source_relpath)
os.makedirs(os.path.dirname(rev_src_path), exist_ok=True)
write_wav(rev_src_path, rev_src_data, samplerate_src)
mix_rev_sources.append(rev_src_data)
mixed_rev_data = np.sum(np.stack(mix_rev_sources, axis=0), axis=0)
mix_wav_path = os.path.join(out_folder, mix)
mix_wav_base = mix_wav_path.rstrip('.wav')
write_wav(mix_wav_path, mixed_rev_data, samplerate_src)
in_wav_path = os.path.join(sources_folder, mix)
in_wav_base = in_wav_path.rstrip('.wav')
if os.path.exists(in_wav_base + '.jams'):
shutil.copyfile(in_wav_base + '.jams', mix_wav_base + '.jams')
if os.path.exists(in_wav_base + '.txt'):
with open(in_wav_base + '.txt', 'r') as f:
lines = f.readlines()
with open(mix_wav_base + '.txt', 'w') as f:
f.write(''.join(lines))
f.write('\nroom impulse responses used:\n{}'.format(
'\n'.join(rir_paths_used)))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--file',
type=str,
required=True,
help='Path to HRIR or HeSuVi file.')
parser.add_argument(
'--track_order',
type=str,
required=True,
help='Track order in HRIR file. "hesuvi" or "hexadecagonal"')
parser.add_argument(
'--reverb',
type=str,
default=argparse.SUPPRESS,
help=
'Reverberation times for different channels in milliseconds. During this time the '
'reverberation tail will be reduced by 100 dB. A comma separated list of channel name and '
'reverberation time pairs, separated by colon. If only a single numeric value is given, '
'it is used for all channels. When some channel names are give but not all, the missing '
'channels are not affected. Must be at least 3 ms smaller than the HRIR length. '
'For example "--reverb=300" or '
'"--reverb=FL:500,FC:100,FR:500,SR:700,BR:700,BL:700,SL:700" or '
'"--reverb=FC:100".')
args = parser.parse_args()
file_path = args.file
track_order = args.track_order
reverb = dict()
try:
# Single float value
reverb = {ch: float(args.reverb) / 1000 for ch in SPEAKER_NAMES}
except ValueError:
# Channels separated
for ch_t in args.reverb.split(','):
reverb[ch_t.split(':')[0].upper()] = float(
ch_t.split(':')[1]) / 1000
fs, data = read_wav(file_path)
for ch, t in reverb.items():
print(f'{ch}: {t*1000:.0f}ms')
n_ones = int(fs * 0.003)
n_win = int(fs * t)
win = np.concatenate([
np.ones(n_ones),
signal.windows.hann(n_win * 2)[n_win:],
np.zeros(data.shape[1] - n_ones - n_win)
]) - 1.0
win *= 100 # 100 dB
win = 10**(win / 20) # Linear scale
if track_order == 'hesuvi':
tracks = [
i for i in range(len(HESUVI_TRACK_ORDER))
if ch in HESUVI_TRACK_ORDER[i]
]
elif track_order == 'hexadecagonal':
tracks = [
i for i in range(len(HEXADECAGONAL_TRACK_ORDER))
if ch in HEXADECAGONAL_TRACK_ORDER[i]
]
else:
raise ValueError(
f'Invalid track_order "{track_order}", allowed values are "hesuvi" and "hexadecagonal"'
)
for i in tracks:
data[i, :] *= win
# Write WAV
write_wav(os.path.join(DIR_PATH, 'cropped.wav'), fs, data)
def task_predict(input_files, input_model):
# 把输入的多个模型目录字符串分离为目录列表
input_models = [os.path.expanduser(k) for k in input_model.strip().split()]
# 把各个目录下的模型列表解压出来组合成一个迭代器
models = itertools.chain(*(glob.glob(m) for m in input_models))
# 生成并加载包括所有模型文件（skgmm.GMMSet object）的列表
models = [ModelInterface.load(m) for m in models]
if len(models) == 0:
print("No model file found in %s" % input_model)
sys.exit(1)
# 定义统计准确率的变量
right = 0
right1 = 0
wrong = 0
wrong1 = 0
num = 0
# 对每个预测音频文件提取特征并与每个模型匹配得到TOP结果
for f in glob.glob(os.path.expanduser(input_files)):
start_time = time.time()
fs, signal = read_wav(f)
print(f)
feat = get_feature(fs, signal)
#print("Get feature ", time.time() - start_time, " seconds")
predict_result = []
f_models = [(feat, m) for m in models]
#print(models)
# 每个音频文件分别匹配每个模型组并得出分数放到总列表
# for model in models:
# #start_time1 = time.time()
# #print(model)
# # 模型文件是一个元组：(label,gmm)
# score = model[1].score(feat)
# label=model[0]
# result=(label,score)
# #print(results)
# predict_result.append(result)
#print("Get one score ", time.time() - start_time1, " seconds")
pool = ThreadPool(2)
predict_result = pool.map(get_score, f_models)
pool.close()
pool.join()
#print(results)
#print("Get score ", time.time() - start_time, " seconds")
proba = GMMSet.softmax([i[1] for i in predict_result])
predict_result = [(predict_result[i][0], proba[i])
for i in range(len(proba))]
#print("predict_result:",predict_result)
# 对预测结果按分数作高到底排序
predict_result = sorted(predict_result,
key=operator.itemgetter(1),
reverse=True)
#print("sort_predict_result:", predict_result)
# 微信语音数据集的label格式
label = os.path.basename(f).split('_')[0] #[6:11]
#label=os.path.basename(f).split('(')[0]#[6:11]
# AISHELL数据集的label格式
# label=os.path.basename(f)[6:11]
predict = predict_result[0][0]
predict_score = predict_result[0][1]
print("Predict ", time.time() - start_time, " seconds")
# #print('Top:',predict_result[:10])
# 统计top1准确率
if label in predict:
right1 += 1
print('label:', label, ' predict:', predict, ' score:',
predict_score, ' top1 right')
else:
wrong1 += 1
print('label:', label, ' predict:', predict, ' score:',
predict_score, ' top1 wrong')
# 统计Top10准确率
predicts = []
predict_scores = []
for pre in predict_result[:10]:
predicts.append(pre[0])
predict_scores.append(pre[1])
if label in predicts:
right += 1
print('label:', label, ' predicts:', predicts, ' scores:',
predict_scores, ' top10 Right')
else:
wrong += 1
print('label:', label, ' predicts:', predicts, ' scores:',
predict_scores, ' top10 Wrong')
num += 1
print('top1:', num, ' right:', right1, ' wrong:', wrong1, ' top1 acc:',
right1 / num)
print('top10:', num, ' right:', right, ' wrong:', wrong, ' top10 acc:',
right / num)
def seg_ditail(fid, trained_model, mt_size, mt_step, st_win):
"""segment ditail"""
st_step = st_win
results = {} ###
fs, signal = read_wav(fid)
[_, st_features] = mt_feature_extraction(signal, fs, mt_size * fs, mt_step * fs,
round(fs * st_win))
# VAD：
segments = vad(st_features, st_step, smooth_window=0.5, weight=0)
i = 0
delta_t = 0.4
for seg in segments:
if seg[1] - seg[0] > 2*delta_t:
start_seg = seg[0]
end_seg = seg[0] + delta_t
while start_seg < end_seg:
label = trained_model.predict(fs, signal[int(start_seg * fs):int(end_seg * fs)])
print(fid, '--', [start_seg, end_seg], '->', label)
# # ***********
# write_wav(os.path.join(os.path.pardir, "result", "result_wav",
# os.path.basename(fid)[:-3] + "-" + str(start_seg) + "-" +
# str(end_seg) + "-" + label + ".wav"),
# fs, signal[int(start_seg * fs):int(end_seg * fs)])
results[i] = {"label": label, "start": start_seg, "end": end_seg}
i = i + 1
start_seg = end_seg
end_seg = start_seg + delta_t if start_seg + 2*delta_t < seg[1] else seg[1]
else:
label = trained_model.predict(fs, signal[int(seg[0] * fs):int(seg[1] * fs)])
print(fid, '--', seg, '->', label)
results[i] = {"label": label, "start": seg[0], "end": seg[1]}
i = i + 1
# # ***********
# write_wav(os.path.join(os.path.pardir, "result", "result_wav",
# os.path.basename(fid)[:-3] + "-" + str(last) + "-" +
# str(seg[0]) + "-静音.wav"),
# fs, signal[int(last * fs):int(seg[0] * fs)])
# write_wav(os.path.join(os.path.pardir, "result", "result_wav",
# os.path.basename(fid)[:-3] + "-" + str(seg[0]) + "-" +
# str(seg[1]) + "-" + label + ".wav"),
# fs, signal[int(seg[0] * fs):int(seg[1] * fs)])
# last = seg[1]
data = {"video_info": {}, "results": []}
min_duration = 0.5
start_seg = results[0]["start"]
end_seg = results[0]["end"]
label = results[0]["label"]
# k = 0 ###
# test = {}###
# last = 0 ###
for j in range(1, i-1):
if results[j]["start"] - end_seg < min_duration \
and results[j]["label"] == label:
end_seg = results[j]["end"]
else:
if end_seg - start_seg >= 2*min_duration:
data["results"].append({"start": start_seg, "end": end_seg, "speaker_id": label})
# write_wav(os.path.join(os.path.pardir, "result", "result_wav",
# str(int(k/10))+str(k%10)+os.path.basename(fid)[:-3] + "-" + str(last) + "-" +
# str(start_seg) + "-静音.wav"),
# fs, signal[int(last * fs):int(start_seg *
# write_wav(os.path.join(os.path.pardir, "result", "result_wav",
# str(int(k / 10)) + str(k % 10) +os.path.basename(fid)[:-3] +
# "-" + str(start_seg) + "-" +
# str(end_seg) + "-" + label + ".wav"),
# fs, signal[int(start_seg * fs):int(end_seg * fs)])
# if start_seg - last > 0.5:
# test[k] = {"label": "无人声", "start": last, "end": start_seg} ###
# k = k + 1
# if end_seg - start_seg > 0.5:
# test[k] = {"label": label, "start": start_seg, "end": end_seg} ###
# k = k + 1
# last = end_seg
start_seg = results[j]["start"]
end_seg = results[j]["end"]
label = results[j]["label"]
# test[k] = {"start": start_seg, "end": end_seg, "label": label}
# with open("D:\\pro_file\\untitled\\Amber_SpeechSeparation\\test\\result_wav\\example.json",
# 'w', encoding='utf-8') as fid_exam:
# json.dump(test, fid_exam, ensure_ascii=False)
data["results"].append({"start": start_seg, "end": end_seg, "speaker_id": label})
write_wav(os.path.join(os.path.pardir, "result", "result_wav",
os.path.basename(fid)[:-4] + "-" + str(start_seg) + "-" +
str(end_seg) + "-" + label + ".wav"),
fs, signal[int(start_seg * fs):int(end_seg * fs)])
with open(os.path.join(os.path.pardir, "result", "test_json",
os.path.basename(fid)[:-3] + "json"),
'w', encoding='utf-8') as json_file:
print("..\\result\\test_json\\" + os.path.basename(fid)[:-3] + "json->Generated")
json.dump(data, json_file, ensure_ascii=False)
