代码来自https://github.com/rosrad/asvspoof2017
def trim_silence(audio, threshold=0.1, frame_length=2048):
if audio.size < frame_length:
frame_length = audio.size
energy = librosa.feature.rmse(audio, frame_length=frame_length)
frames = np.nonzero(energy > threshold)
indices = librosa.core.frames_to_samples(frames)[1]
return audio[indices[0]:indices[-1]] if indices.size else audio[0:0]
def extract_imfcc(wav_path):
audio, sr = librosa.load(wav_path, sr=16000)
S = np.abs(librosa.core.stft(audio, n_fft=n_fft, hop_length=hop_length)) ** 2.0
mel_basis = librosa.filters.mel(sr, n_fft)
mel_basis = np.linalg.pinv(mel_basis).T
mel = np.dot(mel_basis, S)
S = librosa.power_to_db(mel)
imfcc = np.dot(librosa.filters.dct(n_imfcc, S.shape[0]), S)
imfcc_delta = librosa.feature.delta(imfcc)
imfcc_delta_delta = librosa.feature.delta(imfcc)
feature = np.concatenate((imfcc, imfcc_delta, imfcc_delta_delta), axis=0)
return feature
def extract_mfcc(wav_path):
audio, sr = librosa.load(wav_path, sr=16000)
y = trim_silence(audio)
if y.size == 0:
y =audio
mfcc = librosa.feature.mfcc(y, sr, n_mfcc=n_mfcc, n_fft=n_fft, hop_length=hop_length)
mfcc_delta = librosa.feature.delta(mfcc)
mfcc_delta_delta = librosa.feature.delta(mfcc)
feature = np.concatenate((mfcc, mfcc_delta, mfcc_delta_delta), axis=0)
return feature
def extract_cqt(wav_path):
audio, sr = librosa.load(wav_path, sr=16000)
y = trim_silence(audio)
if y.size == 0:
y =audio
cqt = librosa.feature.chroma_cqt(y, sr, hop_length=hop_length, fmin=f_min, n_chroma=n_cqt, n_octaves=5)
return cqt
def extract_spect(wav_path):
audio, sr = librosa.load(wav_path, sr=16000)
# audio = trim_silence(audio, 0.01)
S, _ = librosa.core.spectrum._spectrogram(audio, hop_length=100, n_fft=1000, power=2)
return librosa.power_to_db(S)
def extract_fft(wav_path):
p_preemphasis = 0.97
min_level_db = -100
num_freq = 1025
ref_level_db = 20
frame_length_ms = 20
frame_shift_ms = 10
def _normalize(S):
return np.clip((S - min_level_db) / -min_level_db, 0, 1)
def preemphasis(x):
return signal.lfilter([1, -p_preemphasis], [1], x)
def _stft(y):
n_fft, hop_length, win_length = _stft_parameters()
return librosa.stft(y=y, n_fft=n_fft, hop_length=hop_length)
def _stft_parameters():
# n_fft = (num_freq - 1) * 2
n_fft = 1800
hop_length = 150
# hop_length = int(frame_shift_ms / 1000 * sample_rate)
# win_length = int(frame_length_ms / 1000 * sample_rate)
win_length = 1500
return n_fft, hop_length, win_length
def _amp_to_db(x):
return 20 * np.log10(np.maximum(1e-5, x))
y = librosa.core.load(wav_path, sr=sample_rate)[0]
D = _stft(preemphasis(y))
S = _amp_to_db(np.abs(D)) - ref_level_db
return _normalize(S)
def extract_db4(wav_path):
audio, sr = librosa.load(wav_path, sr=16000)
S, _ = librosa.core.spectrum._spectrogram(audio, hop_length=150, n_fft=1500, power=2)
S =librosa.power_to_db(S)
cA, cD = pywt.dwt(S, 'db4')
return cA
def extract_db8(wav_path):
audio, sr = librosa.load(wav_path, sr=16000)
S, _ = librosa.core.spectrum._spectrogram(audio, hop_length=150, n_fft=1500, power=2)
S =librosa.power_to_db(S)
cA, cD = pywt.dwt(S, 'db8')
return cA
def extract_raw(wav_path):
audio, sr = librosa.load(wav_path, sr=16000)
# y = trim_silence(audio, threshold=0.05)
# if y.size == 0:
# y = audio
return audio
