def read_wav_file(self, file_name):
audio, fs = librosa.core.load(file_name, sr=config.fs)
audio = np.float64(audio)
if len(audio.shape) == 2:
vocals = np.array((audio[:, 1] + audio[:, 0]) / 2)
else:
vocals = np.array(audio)
voc_stft = abs(
np.array(
utils.stft(audio,
hopsize=config.hopsize,
nfft=config.framesize,
fs=config.fs)))
feats = sig_process.get_world_feats(audio)
voc_stft = np.clip(voc_stft, 0.0, 1.0)
return feats, voc_stft
def get_feats(audio):
"""
Function to get acoustic input features, starting with STFT, needs to be extended to include MFCCs, will ask how many coefficients to use.
"""
# stft = librosa.core.stft(audio, n_fft = config.nfft, hop_length = config.hopsize, window = config.window).T
stft = utils.stft(audio,
window=config.window,
hopsize=config.hopsize,
nfft=config.nfft,
fs=config.fs)
assert abs(stft).max() <= 1.0
# voc_stft_mag = 2 * abs(voc_stft)/np.sum(config.window)
# voc_stft_phase = np.angle(voc_stft)
# cqt = librosa.core.cqt(audio, sr = config.fs, hop_length = config.hopsize, n_bins = config.cqt_bins, fmin = config.fmin, bins_per_octave = config.bins_per_octave).T
# hcqt = get_hcqt(audio)
# hcqt = np.swapaxes(hcqt, 0,1)
return stft
def extract_feature_wav(self, audio):
"""
Function to extract multi pitch from file. Currently supports only HDF5 files.
"""
stft = abs(np.array(utils.stft(audio, hopsize=config.hopsize, nfft=config.framesize, fs=config.fs)))
out_feats = self.extract_feature(stft, self.sess)
out_feats = out_feats[:stft.shape[0]]
return out_feats
def process_seg_yam(audio, audio_back):
"""
Process a segment of the audio.
Returns the world features, TONY annotated notes and the STFT.
"""
out_feats = sig_process.get_world_feats(audio)
#Test if the reverse works.
# audio_out = sig_process.feats_to_audio(out_feats)
traj = vamp_notes.extract_notes_pYIN_vamp(audio)
if traj.shape[0]<1 or len(out_feats)<=config.max_phr_len:
return None,None,None
else:
timestamps = np.arange(0, float(traj[-1][1]), config.hoptime)
out_notes = vamp_notes.note2traj(traj, timestamps)
out_notes_1 = sig_process.f0_to_hertz(out_notes[:,0])
out_notes_1[out_notes_1== -np.inf] = 0
out_notes[:,0] = out_notes_1
out_stft = abs(np.array(utils.stft(audio, hopsize=config.hopsize, nfft=config.framesize, fs=config.fs)))
back_stft = abs(np.array(utils.stft(audio_back, hopsize=config.hopsize, nfft=config.framesize, fs=config.fs)))
out_feats, out_notes, out_stft, back_stft = utils.match_time([out_feats, out_notes, out_stft, back_stft])
if len(out_feats)<=config.max_phr_len:
return None,None,None, None
else:
assert all(out_feats[:,-2]>0)
assert len(out_feats) == len(out_notes)
return out_feats, out_notes, out_stft, back_stft
def read_wav_file(self, file_name):
audio, fs = librosa.core.load(file_name, sr=config.fs)
audio = np.float64(audio)
if len(audio.shape) == 2:
vocals = np.array((audio[:, 1] + audio[:, 0]) / 2)
else:
vocals = np.array(audio)
voc_stft = abs(utils.stft(vocals))
feats = utils.stft_to_feats(vocals, fs)
voc_stft = np.clip(voc_stft, 0.0, 1.0)
return voc_stft, feats