def get_world_feats(vocals):
feats = pw.wav2world(vocals, config.fs, frame_period=config.hoptime * 1000)
ap = feats[2].reshape([feats[1].shape[0],
feats[1].shape[1]]).astype(np.float32)
ap = 10. * np.log10(ap**2)
harm = 10 * np.log10(feats[1].reshape(
[feats[2].shape[0], feats[2].shape[1]]))
f0 = feats[0]
# f0 = pitch.extract_f0_sac(vocals, fs, config.hoptime)
y = 69 + 12 * np.log2(f0 / 440)
# import pdb;pdb.set_trace()
# y = hertz_to_new_base(f0)
nans, x = utils.nan_helper(y)
naners = np.isinf(y)
y[nans] = np.interp(x(nans), x(~nans), y[~nans])
# y=[float(x-(min_note-1))/float(max_note-(min_note-1)) for x in y]
y = np.array(y).reshape([len(y), 1])
guy = np.array(naners).reshape([len(y), 1])
y = np.concatenate((y, guy), axis=-1)
if config.comp_mode == 'mfsc':
harmy = sp_to_mfsc(harm, 60, 0.45)
apy = sp_to_mfsc(ap, 4, 0.45)
elif config.comp_mode == 'mgc':
harmy = sp_to_mgc(harm, 60, 0.45)
apy = sp_to_mgc(ap, 4, 0.45)
out_feats = np.concatenate((harmy, apy, y.reshape((-1, 2))), axis=1)
return out_feats
def stft_to_feats(vocals, fs, mode=config.comp_mode):
if len(vocals.shape)>1:
vocals = vocals[:,0]
vocals = np.ascontiguousarray(vocals)
feats=pw.wav2world(vocals,fs,frame_period=5.80498866)
ap = feats[2].reshape([feats[1].shape[0],feats[1].shape[1]]).astype(np.float32)
ap = 10.*np.log10(ap**2)
harm=10*np.log10(feats[1].reshape([feats[2].shape[0],feats[2].shape[1]]))
f0 = feats[0]
# f0_1 = pitch.extract_f0_sac(vocals, fs, 0.00580498866)
# import pdb;pdb.set_trace()
y=69+12*np.log2(f0/440)
# y = hertz_to_new_base(f0)
nans, x= nan_helper(y)
naners=np.isinf(y)
y[nans]= np.interp(x(nans), x(~nans), y[~nans])
# y=[float(x-(min_note-1))/float(max_note-(min_note-1)) for x in y]
y=np.array(y).reshape([len(y),1])
guy=np.array(naners).reshape([len(y),1])
y=np.concatenate((y,guy),axis=-1)
if mode == 'mfsc':
harmy=sp_to_mfsc(harm,60,0.45)
apy=sp_to_mfsc(ap,4,0.45)
elif mode == 'mgc':
harmy=sp_to_mgc(harm,60,0.45)
apy=sp_to_mgc(ap,4,0.45)
# import pdb;pdb.set_trace()
out_feats=np.concatenate((harmy,apy,y.reshape((-1,2))),axis=1)
# harm_in=mgc_to_sp(harmy, 1025, 0.45)
# ap_in=mgc_to_sp(apy, 1025, 0.45)
return out_feats
def stft_to_feats(vocals, fs = config.fs):
if len(vocals.shape)>1:
vocals = vocals[:,0]
vocals = np.ascontiguousarray(vocals)
feats=pw.wav2world(vocals,fs = config.fs,frame_period=config.hoptime*1000)
ap = feats[2].reshape([feats[1].shape[0],feats[1].shape[1]]).astype(np.float32)
ap = 10.*np.log10(ap**2)
harm=10 * np.log10(feats[1].reshape([feats[2].shape[0],feats[2].shape[1]]))
harm = harm - 20
f0 = feats[0]
is_voiced = f0 > 0.0
if not np.any(is_voiced):
pass # all unvoiced, do nothing
else:
for k in range(ap.shape[1]):
ap[~is_voiced, k] = np.interp(np.where(~is_voiced)[0], np.where(is_voiced)[0], ap[is_voiced, k])
# f0_1 = pitch.extract_f0_sac(vocals, fs, 0.00580498866)
y=69+12*np.log2(f0/440)
# y = hertz_to_new_base(f0)
nans, x= nan_helper(y)
naners=np.isinf(y)
y[nans]= np.interp(x(nans), x(~nans), y[~nans])
# y=[float(x-(min_note-1))/float(max_note-(min_note-1)) for x in y]
y=np.array(y).reshape([len(y),1])
guy=np.array(naners).reshape([len(y),1])
y=np.concatenate((y,guy),axis=-1)
harm = np.nan_to_num(harm)
ap = np.nan_to_num(ap)
harmy= sp_to_mfsc(harm+1e-12,60,0.45)
apy= ap_to_wbap(ap+1e-12,4,config.fs)
out_feats=np.concatenate((harmy,apy,y.reshape((-1,2))),axis=1)
# import pdb;pdb.set_trace()
# audio_out = feats_to_audio(out_feats)
# sf.write('./test_mfsc.wav', audio_out, config.fs)
# import pdb;pdb.set_trace()
# harm_in=mgc_to_sp(harmy, 1025, 0.45)
# ap_in= wbap_to_ap(apy, 1025, config.fs)
# harm_in = 10**((harm_in + 20)/10)
# ap_in = np.clip(10**(ap_in/20), 0.0, 1.0)
# audio_out = pw.synthesize(f0 , np.ascontiguousarray(harm_in).astype('double') , np.ascontiguousarray(ap).astype('double'),config.fs,config.hoptime*1000)
# sf.write('./test.wav', audio_out, config.fs)
return out_feats, f0