def draw_logmel(wav_file, label, feature_name, logmelband_nums=[4, 5]):
(y, sr) = librosa.load(wav_file)
rate, data = read(wav_file)
plt.figure()
plt.subplot(2, 1, 1)
D = librosa.amplitude_to_db(np.abs(librosa.stft(y)), ref=np.max)
librosa.display.specshow(D, y_axis='linear', cmap="viridis")
plt.colorbar(format='%+2.0f dB')
plt.title('Linear-frequency power spectrogram')
plt.ylim([0, 8192])
plt.subplot(2, 1, 2)
feat, _ = fbank(y, samplerate=rate, nfft=2048)
logfbank_energy = np.log(feat).T
logfbank_energy = logfbank(data, samplerate=rate, nfft=2048)
colors = ["r", "g", "b"]
for i in range(len(logmelband_nums)):
logmelband_num = logmelband_nums[i]
X = np.linspace(0, len(logfbank_energy[logmelband_num]),
len(logfbank_energy[logmelband_num]))
plt.plot(X,
logfbank_energy[logmelband_num],
'o',
markersize=5,
color=colors[i],
label="logMelFreqBands[{}]".format(logmelband_num))
#quantile_value=np.quantile(logfbank_energy[logmelband_num],0.25*2)
#plt.plot(X,[quantile_value]*len(X),markersize=2,color="r",label="quartile2")
#quantile_value=np.quantile(logfbank_energy[logmelband_num],0.25*3)
#plt.plot(X,[quantile_value]*len(X),markersize=2,color="g",label="quartile3")
plt.title('logMelFrequencyBands(de)')
plt.ylabel("Filterbank")
plt.xlabel("Frame Idx")
plt.legend(loc="upper right", prop={"size": 8})
plt.savefig("/home/jialu/voice_quality_plots/v2/logMelFreqBand/" + label +
"_" + feature_name + ".png")
def tiqu(path, weidu, logenergy, energy_1):
basedir = path
for mulu in os.listdir(basedir):
input_dir = os.path.join(basedir, mulu, "wav")
#语音文件的路劲
output_dir2 = os.path.join(basedir, mulu, 'log_yuan')
#log梅尔普系数
# output_dir3 =r"C:\Users\a7825\Desktop\工作空间\语音数据\UUDB\第一次实验\打标签\第三批\C063L\mfcc"
#mfcc
muluz.mkdir(output_dir2)
for ad_file in os.listdir(input_dir):
print(input_dir + "/" + ad_file)
(fs, audio) = wav.read(input_dir + "/" + ad_file)
if energy_1 == True:
feat, energy = fbank(audio, fs, nfilt=weidu)
np.savetxt(output_dir2 + "/" + ad_file + ".csv",
feat,
delimiter=',')
if logenergy == True:
log = logfbank(audio, fs, nfilt=weidu)
np.savetxt(output_dir2 + "/" + ad_file + ".csv",
log,
delimiter=',')
def logfbank_features(fname):
"""
Compute log Mel-filterbank energy features
"""
(rate, signal) = wavfile.read(fname)
fbank_beat = logfbank(signal, rate)
# take mean of all rows
features = fbank_beat.mean(axis=0)
return features
def logfbank(signal,
rate=default_rate,
filters_number=default_filters_number,
augmented=default_augmented):
logfbank_features = logfbank(signal, rate, nfilt=filters_number)
if not augmented:
return logfbank_features
d_logfbank_features = delta(logfbank_features, 2)
a_logfbank_features = delta(d_logfbank_features, 2)
concatenated_features = np.concatenate(
(logfbank_features, d_logfbank_features, a_logfbank_features),
axis=1)
return concatenated_features
def Features(self, data, rate, dim):
spec = np.abs(np.fft.rfft(data))
freq = np.fft.rfftfreq(len(data), d=1 / dim)
a = spec / spec.sum()
meaN = (freq * a).sum()
std = np.sqrt(np.sum(a * ((freq - meaN) ** 2)))
a_cumsum = np.cumsum(a)
mediaN = freq[len(a_cumsum[a_cumsum <= 0.5])]
modE = freq[a.argmax()]
q25 = freq[len(a_cumsum[a_cumsum <= 0.25])]
q75 = freq[len(a_cumsum[a_cumsum <= 0.75])]
IQR = q75 - q25
z = a - a.mean()
w = a.std()
skewnesS = ((z ** 3).sum() / (len(spec) - 1)) / w ** 3
kurtosiS = ((z ** 4).sum() / (len(spec) - 1)) / w ** 4
m = speech.mfcc(data,rate)
f = speech.fbank(data,rate)
l = speech.logfbank(data,rate)
s = speech.ssc(data,rate)
data = pd.DataFrame(data)
desc = data.describe()
mean = desc.loc["mean"].get(0)
mad = data.mad().get(0)
sd = desc.loc["std"].get(0)
median = data.median().get(0)
minimum = desc.loc["min"].get(0)
maximum = desc.loc["max"].get(0)
Q25 = desc.loc["25%"].get(0)
Q75 = desc.loc["75%"].get(0)
interquartileR = Q75 - Q25
skewness = data.skew().get(0)
kurtosis = data.kurtosis().get(0)
result = {
"Mean": mean, "Mad": mad, "deviation": sd, "Median": median, "Min": minimum, "Max": maximum,
"interquartileR": interquartileR, "Skewness": skewness, "Q25": Q25, "Q75": Q75, "Kurtosis": kurtosis,
"mfcc_mean": np.mean(m), "mfcc_max": np.max(m), "mfcc_min": np.min(m),
"fbank_mean": np.mean(f[0]), "fbank_max": np.max(f[0]), "fbank_min": np.min(f[0]),
"energy_mean": np.mean(f[1]), "energy_max": np.max(f[1]), "energy_min": np.min(f[1]),
"lfbank_mean": np.mean(l), "lfbank_max": np.max(l), "lfbank_min": np.min(l),
"ssc_mean": np.mean(s), "ssc_max": np.max(s), "ssc_min": np.min(s),
"meaN": meaN, "deviatioN": std, "mediaN": mediaN, "modE": modE, "IQR": IQR,
"skewnesS": skewnesS, "q25": q25, "q75": q75, "kurtosiS": kurtosiS}
return result
def get_fbank_feature(wavsignal, fs):
"""
输入为wav文件数学表示和采样频率，输出为语音的FBANK特征+一阶差分+二阶差分
:param wavsignal:
:param fs:
:return:
"""
feat_fbank = logfbank(wavsignal,
fs,
nfilt=40,
nfft=2048,
winstep=0.025,
winlen=0.05)
feat_fbank_d = delta(feat_fbank, 2)
feat_fbank_dd = delta(feat_fbank_d, 2)
wav_feature = np.column_stack((feat_fbank, feat_fbank_d, feat_fbank_dd))
return wav_feature
def get_mfcc(x):
y = np.concatenate([
mfcc(x, numcep=12, winlen=0.01, winstep=0.005),
logfbank(x, nfilt=1, winlen=0.01, winstep=0.005)
],
axis=-1)
derivatives = []
previousf = np.zeros((13, ))
for i in range(len(y)):
if (i + 1) == len(y):
nextf = np.zeros((13, ))
else:
nextf = y[i + 1]
derivatives.append(((nextf - previousf) / 2).reshape((1, 13)))
previousf = y[i]
derivatives = np.concatenate(derivatives, axis=0)
y = np.concatenate([y, derivatives], axis=1)
return y
def _compute_sample(self, file, sliceLength, segmentLength):
"""computes the feature and label vector for single audio slice"""
filename = self.dataset_path + file['audio_file_path']
fs, sound_data = wavfile.read(filename)
sound_data = sound_data.astype('float32')
label = file['is_hotword']
label_vec = np.full(int(sliceLength / segmentLength - 1),
0,
dtype='int8')
if label:
is_audio_subsampled = self._detect_audio(
sound_data
)[0::
segmentLength] #detect audio and subsample to single label per segment
num_segments = int(sliceLength / segmentLength)
if len(is_audio_subsampled) >= num_segments:
is_audio_subsampled = is_audio_subsampled[:
num_segments] #truncate if longer than target segments
else:
is_audio_subsampled = np.pad(
is_audio_subsampled,
(num_segments - len(is_audio_subsampled),
0)) #pad if shorter than target segments
label_vec = label_vec + is_audio_subsampled[
1:] # overlay frames where label is true
if len(sound_data) >= sliceLength:
sound_data = sound_data[:sliceLength]
else:
sound_data = np.pad(sound_data, (sliceLength - len(sound_data), 0))
#compute log mel filterbank energies
feature_vec = logfbank(sound_data,
samplerate=self.sampling_frequency,
winlen=self.window_size,
winstep=self.time_step,
nfilt=self.num_features)
return feature_vec, label_vec
def get_mfcc(x):
y = np.concatenate([
mfcc(x, numcep=12, winlen=0.01, winstep=0.005),
logfbank(x, nfilt=1, winlen=0.01, winstep=0.005)
],
axis=-1)
derivatives = []
previousf = np.zeros((13, ))
for i in range(len(y)):
if (i + 1) == len(y):
nextf = np.zeros((13, ))
else:
nextf = y[i + 1]
derivatives.append(((nextf - previousf) / 2).reshape((1, 13)))
previousf = y[i]
derivatives = np.concatenate(derivatives, axis=0)
y = np.concatenate([y, derivatives], axis=1)
ynoise = np.random.normal(0, 0.6, y.shape)
orig_len = len(y)
pad = [np.zeros((1, 26))] * (3150 - y.shape[0])
ypad = np.concatenate([y] + pad, axis=0)
noisepad = np.concatenate([ynoise] + pad, axis=0)
return orig_len, ypad, ypad + noisepad
import scipy.io.wavfile as wav
from python_speech_features.base import mfcc, fbank, logfbank
import numpy as np
import os
input_dir = r"C:\Users\a7825\Desktop\工作空间\跑代码\打标签\F1\B\wav_16"
output_dir1 = r"C:\Users\a7825\Desktop\工作空间\跑代码\打标签\F1\B\melbank"
output_dir2 = r"C:\Users\a7825\Desktop\工作空间\跑代码\打标签\F1\B\log"
if __name__ == "__main__":
for ad_file in os.listdir(input_dir):
(fs, audio) = wav.read(input_dir + "/" + ad_file)
feature_m, feature_n = fbank(audio, fs, winfunc=np.hamming)
log = logfbank(audio, fs)
np.savetxt(output_dir1 + "/" + ad_file + ".fbank.csv",
feature_m,
delimiter=',')
np.savetxt(output_dir2 + "/" + ad_file + ".log.csv",
log,
delimiter=',')
# 拼接字符串，把单引号改成双引号居然好使
# fs, audio = wav.read(r"C:\Users\a7825\Desktop\新しいフォルダー/a.wav")
# # feature_mfcc = mfcc(audio, samplerate=fs, numcep=40, nfilt=40)
# feature_m,feature_n = fbank(audio, fs)
# feature_log = logfbank(audio, fs)
# # print(feature_mel[0].shape)
# # print(feature_m)
# # print(feature_n.shape)
# np.savetxt('fbank.csv', feature_m, delimiter = ',')
# np.savetxt('energy.csv', feature_n, delimiter= ',')
import matplotlib.pyplot as plt
import librosa as li
import librosa.display as ds
import librosa
from python_speech_features.base import mfcc, logfbank
import scipy.io.wavfile as wav
import numpy as np
import os
indir = r'C:\Users\a7825\Desktop\工作空间\セミナー\语音\wav/C001L_061.wav'
# indir_1 =r'C:\Users\a7825\Desktop\工作空间\杂物\临时\这个就对了'
# 显示メルフィルタバンク的图
x, fs = li.load(indir, sr=16000)
# (fs, x) = wav.read(indir)
log = logfbank(x, fs)
# np.savetxt(indir_1 + ".csv", log, delimiter=',')
# print(log.shape)
# os.system('pause')
# for e in range(11,21):
# for i in range(26):
# log[e][i] = log[e][i]+3.5#为了让图变得亮一点
# ig, ax = plt.subplots()
# plt.xlim(0,130)#设置x的范围
plt.ylim(0, 25.6) #设置y的范围
# new_ticks = np.linspace(0.00,5.00,4)
# plt.xticks(new_ticks)
import matplotlib.pyplot as plt import librosa as li import librosa.display as ds import librosa from python_speech_features.base import mfcc, logfbank, fbank import scipy.io.wavfile as wav import numpy as np import zhengguihua import os indir = r'C:\Users\a7825\Desktop\工作空间\セミナー\语音\wav/C001L_061.wav' # indir_1 =r'C:\Users\a7825\Desktop\工作空间\杂物\临时\这个就对了' # 显示logメルフィルタバンク的图 (fs, x) = wav.read(indir) log = logfbank(x, fs, nfilt=40) np.savetxt(indir + ".csv", log, delimiter=',') ig, ax = plt.subplots() log = np.swapaxes(log, 0, 1) cax = ax.imshow(log, interpolation='nearest', origin='lower', aspect='auto') plt.show()
def calcAcousticFeatures(sound,
fs,
featureMode,
speakerType='male',
tmpDir=".",
speech_sound_type='vowel',
octave_binding=None):
"""
Calculates acoustic features with given featureMode for sound
with audio sampling rate fs.
sound: 1D np.array or 2D array of horizontal vector.
Returns a tuple containing a 2D np.array of numTimeSteps x numFeatureParams
and a 2 x numFeatureParams array that contains scaling factors that
can be used to ensure equal contribution of each feature type.
"""
# how many Hz may change in the first formant in mergeFactor ms – few variation for vowels required
if speech_sound_type == 'vowel':
maxFormantChange = 50
elif speech_sound_type == 'syllable':
maxFormantChange = 800
if np.ndim(sound) == 2:
sound = sound[0, :]
if featureMode == 'formants':
formants = getPraatFormantsMean(sound, fs, speakerType, tmpDir)
return (np.array(formants).reshape((1, -1)), None)
elif featureMode == 'formants_full':
(timePos, formants) = getPraatFormants(sound, fs, speakerType, tmpDir)
# downsample
mergeFactor = 10 # how many time steps (=ms) should be merged to one value
newFormants = np.array(np.mean(formants[0:mergeFactor, :], 0), ndmin=2)
for t in range(mergeFactor, len(timePos), mergeFactor):
new = np.mean(formants[t:t + mergeFactor, :], 0)
if abs(newFormants[-1, 0] - new[0]) > maxFormantChange:
# TODO: this is dangerous if the first detected formant is incorrect!
pass
else:
newFormants = np.vstack((newFormants, new))
return (newFormants, None)
elif featureMode == 'mfcc':
# sound as 1d
if sound.ndim == 2:
sound = np.reshape(sound, (-1))
# returns (numFrames x numCeps) np array
window_length = 0.02 # 0.025 * 22050 = ca. 551 frames
window_step = 0.01 # 0.01 * 22050 = ca. 221 frames
num_cepstrals = 13
features = mfcc(sound, fs, window_length, window_step, num_cepstrals)
return (features, None)
elif featureMode == 'mfcc_formants':
# sound as 1d
if sound.ndim == 2:
sound = np.reshape(sound, (-1))
# returns (numFrames x numCeps) np array
window_length = 0.02 # 0.025 * 22050 = ca. 551 frames
window_step = 0.005 # 0.01 * 22050 = ca. 221 frames
num_cepstrals = 13
features = mfcc(sound, fs, window_length, window_step, num_cepstrals)
(timePos, formants) = getPraatFormants(sound, fs, speakerType, tmpDir)
# downsample
mergeFactor = 10 # how many time steps (=ms) should be merged to one value
# get a good estimate for initial formants (ignoring initial perturbations):
initialFormants = np.median(formants[0:5, :], axis=0)
newFormants = None
i = 0
while not newFormants:
if abs(formants[i, 0] - initialFormants[0]) < maxFormantChange:
newFormants = formants[i, :]
break
else:
i += 1
newFormants = np.array(np.mean(formants[0:mergeFactor, :], 0), ndmin=2)
for t in range(mergeFactor, len(timePos), mergeFactor):
new = np.mean(formants[t:t + mergeFactor, :], 0)
if abs(newFormants[-1, 0] - new[0]) > maxFormantChange:
pass
else:
newFormants = np.vstack((newFormants, new))
# resample formants according to mfccs
# TODO Warning: interp just copies the last element to make trajectories longer!!!
# alternative: https://docs.scipy.org/doc/scipy/reference/tutorial/interpolate.html
resampledFormants = np.zeros(
(np.shape(features)[0], np.shape(newFormants)[1]))
for i in range(np.shape(newFormants)[1]):
resampledFormants[:,
i] = np.interp(range(np.shape(features)[0]),
range(np.shape(newFormants)[0]),
newFormants[:, i])
minmax = np.array([
np.concatenate((np.repeat([-1 / np.shape(resampledFormants)[1]],
np.shape(resampledFormants)[1]),
np.repeat([-1 / np.shape(features)[1]],
np.shape(features)[1]))),
np.concatenate((np.repeat([1 / np.shape(resampledFormants)[1]],
np.shape(resampledFormants)[1]),
np.repeat([1 / np.shape(features)[1]],
np.shape(features)[1])))
])
return (np.concatenate((resampledFormants, features), axis=1), minmax)
elif featureMode == "fbank":
# sound as 1d
if sound.ndim == 2:
sound = np.reshape(sound, (-1))
fbank_feat = logfbank(sound, fs, nfft=1024)
return (fbank_feat,
np.concatenate(([-1 * np.ones(fbank_feat.shape[1])],
[np.ones(fbank_feat.shape[1])])))
elif featureMode == "logfbank":
# sound as 1d
if sound.ndim == 2:
sound = np.reshape(sound, (-1))
fbank_feat = logfbank(sound, fs, nfft=1024)
return (fbank_feat,
np.concatenate(([-1 * np.ones(fbank_feat.shape[1])],
[np.ones(fbank_feat.shape[1])])))
elif featureMode == 'gbfb': # Gabor filter bank features, requires Octave installed
# scaledAudio = np.int16(copiedAudio/maxAmplitude * 32767)
soundNorm = sound / 32767
#features = octave_binding.gbfb_feature_extraction(soundNorm, fs)
features = octave_binding.heq(
octave_binding.gbfb(
octave_binding.log_mel_spectrogram(soundNorm, fs)))
features = features.transpose()
return (features,
np.concatenate(([-1 * np.ones(features.shape[1])],
[np.ones(features.shape[1])])))
else:
print("Feature mode " + featureMode +
" not yet defined in calcAcousticFeatures()!")
return None
