eden-emu/opus
15
0
Fork 0
mirror of https://github.com/xiph/opus.git synced 2025-05-22 11:28:31 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Computing signals in C

Browse source
This commit is contained in:
Jean-Marc Valin 2018-12-28 02:52:58 -05:00
parent d75a4aec72
commit ea02ef7e02
2 changed files with 53 additions and 64 deletions
Download patch file Download diff file Expand all files Collapse all files

74
dnn/train_lpcnet.py
View file

@ -66,85 +66,39 @@ pcm_chunk_size = frame_size*feature_chunk_size
# u for unquantised, load 16 bit PCM samples and convert to mu-law
udata = np.fromfile(pcm_file, dtype='int16')
data = lin2ulaw(udata)
nb_frames = len(data)//pcm_chunk_size
data = np.fromfile(pcm_file, dtype='uint8')
nb_frames = len(data)//(4*pcm_chunk_size)
features = np.fromfile(feature_file, dtype='float32')
# limit to discrete number of frames
data = data[:nb_frames*pcm_chunk_size]
udata = udata[:nb_frames*pcm_chunk_size]
data = data[:nb_frames*4*pcm_chunk_size]
features = features[:nb_frames*feature_chunk_size*nb_features]
# Noise injection: the idea is that the real system is going to be
# predicting samples based on previously predicted samples rather than
# from the original. Since the previously predicted samples aren't
# expected to be so good, I add noise to the training data. Exactly
# how the noise is added makes a huge difference
in_data = np.concatenate([data[0:1], data[:-1]]);
noise = np.concatenate([np.zeros((len(data)*1//5)), np.random.randint(-3, 3, len(data)*1//5), np.random.randint(-2, 2, len(data)*1//5), np.random.randint(-1, 1, len(data)*2//5)])
#noise = np.round(np.concatenate([np.zeros((len(data)*1//5)), np.random.laplace(0, 1.2, len(data)*1//5), np.random.laplace(0, .77, len(data)*1//5), np.random.laplace(0, .33, len(data)*1//5), np.random.randint(-1, 1, len(data)*1//5)]))
del data
in_data = in_data + noise
del noise
in_data = np.clip(in_data, 0, 255)
features = np.reshape(features, (nb_frames*feature_chunk_size, nb_features))
# Note: the LPC predictor output is now calculated by the loop below, this code was
# for an ealier version that implemented the prediction filter in C
upred = np.zeros((nb_frames*pcm_chunk_size,), dtype='float32')
# Use 16th order LPC to generate LPC prediction output upred[] and (in
# mu-law form) pred[]
pred_in = ulaw2lin(in_data)
for i in range(2, nb_frames*feature_chunk_size):
upred[i*frame_size:(i+1)*frame_size] = 0
for k in range(16):
upred[i*frame_size:(i+1)*frame_size] = upred[i*frame_size:(i+1)*frame_size] - \
pred_in[i*frame_size-k:(i+1)*frame_size-k]*features[i, nb_features-16+k]
del pred_in
pred = lin2ulaw(upred)
in_data = np.reshape(in_data, (nb_frames, pcm_chunk_size, 1))
in_data = in_data.astype('uint8')
# LPC residual, which is the difference between the input speech and
# the predictor output, with a slight time shift this is also the
# ideal excitation in_exc
out_data = lin2ulaw(udata-upred)
del upred
del udata
in_exc = np.concatenate([out_data[0:1], out_data[:-1]]);
out_data = np.reshape(out_data, (nb_frames, pcm_chunk_size, 1))
out_data = out_data.astype('uint8')
in_exc = np.reshape(in_exc, (nb_frames, pcm_chunk_size, 1))
in_exc = in_exc.astype('uint8')
sig = np.reshape(data[0::4], (nb_frames, pcm_chunk_size, 1))
pred = np.reshape(data[1::4], (nb_frames, pcm_chunk_size, 1))
in_exc = np.reshape(data[2::4], (nb_frames, pcm_chunk_size, 1))
out_exc = np.reshape(data[3::4], (nb_frames, pcm_chunk_size, 1))
del data
print("ulaw std = ", np.std(out_exc))
features = np.reshape(features, (nb_frames, feature_chunk_size, nb_features))
features = features[:, :, :nb_used_features]
features[:,:,18:36] = 0
pred = np.reshape(pred, (nb_frames, pcm_chunk_size, 1))
pred = pred.astype('uint8')
periods = (.1 + 50*features[:,:,36:37]+100).astype('int16')
in_data = np.concatenate([in_data, pred], axis=-1)
in_data = np.concatenate([sig, pred], axis=-1)
del sig
del pred
# dump models to disk as we go
checkpoint = ModelCheckpoint('lpcnet15_384_10_G16_{epoch:02d}.h5')
checkpoint = ModelCheckpoint('lpcnet18_384_10_G16_{epoch:02d}.h5')
#model.load_weights('lpcnet9b_384_10_G16_01.h5')
model.load_weights('lpcnet9b_384_10_G16_01.h5')
model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
model.fit([in_data, in_exc, features, periods], out_data, batch_size=batch_size, epochs=nb_epochs, validation_split=0.0, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, (0.1, 0.1, 0.1))])
model.fit([in_data, in_exc, features, periods], out_exc, batch_size=batch_size, epochs=nb_epochs, validation_split=0.0, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, (0.1, 0.1, 0.1))])