shithub: opus

Info  •  Files  •  Log  •  Branches

LPC generation from the cepstral coefficients

--- a/dnn/denoise.c

+++ b/dnn/denoise.c

@@ -325,12 +325,29 @@

   RNN_COPY(st->analysis_mem, in, FRAME_SIZE);

   RNN_COPY(x0, x, WINDOW_SIZE);

   apply_window(x);

+  forward_transform(X, x);

+#if TRAINING

+  for (i=lowpass;i<FREQ_SIZE;i++)

+    X[i].r = X[i].i = 0;

+#endif

+  compute_band_energy(Ex, X);

   {

     float e;

     float g_1;

-    _celt_autocorr(x, ac, NULL, 0, LPC_ORDER, WINDOW_SIZE);

+    if (0) {

+      _celt_autocorr(x, ac, NULL, 0, LPC_ORDER, WINDOW_SIZE);

+    } else {

+      float Xr[FREQ_SIZE];

+      kiss_fft_cpx X_auto[FREQ_SIZE];

+      float x_auto[FRAME_SIZE];

+      interp_band_gain(Xr, Ex);

+      RNN_CLEAR(X_auto, FREQ_SIZE);

+      for (i=0;i<160;i++) X_auto[i].r = Xr[i];

+      inverse_transform(x_auto, X_auto);

+      for (i=0;i<LPC_ORDER+1;i++) ac[i] = x_auto[i];

+    }

     /* -40 dB noise floor. */

-    ac[0] += ac[0]*1e-4 + 320/12;

+    ac[0] += ac[0]*1e-4 + 320/12/38.;

     /* Lag windowing. */

     for (i=1;i<LPC_ORDER+1;i++) ac[i] *= (1 - 6e-5*i*i);

     e = _celt_lpc(lpc, rc, ac, LPC_ORDER);

@@ -341,7 +358,7 @@

     printf("\n");

 #endif

 #if 0

-    printf("%f 1 ", e);

+    printf("1 ");

     for(i=0;i<LPC_ORDER;i++) printf("%f ", lpc[i]);

     printf("\n");

 #endif

@@ -363,12 +380,6 @@

 #endif

     }

   }

-  forward_transform(X, x);

-#if TRAINING

-  for (i=lowpass;i<FREQ_SIZE;i++)

-    X[i].r = X[i].i = 0;

-#endif

-  compute_band_energy(Ex, X);

   return g;

 }

--- a/dnn/lpcnet.py

+++ b/dnn/lpcnet.py

@@ -11,8 +11,8 @@

 import h5py

 import sys

-rnn_units1=512

-rnn_units2=32

+rnn_units1=384

+rnn_units2=16

 pcm_bits = 8

 embed_size = 128

 pcm_levels = 2**pcm_bits

--- a/dnn/test_wavenet_audio.py

+++ b/dnn/test_wavenet_audio.py

@@ -42,7 +42,7 @@

-model.load_weights('wavenet5p0_30.h5')

+model.load_weights('wavenet5_384_10_G16np_50.h5')

 order = 16

--- a/dnn/train_wavenet_audio.py

+++ b/dnn/train_wavenet_audio.py

@@ -120,8 +120,8 @@

 in_data = np.concatenate([in_data, pred], axis=-1)

 # dump models to disk as we go

-checkpoint = ModelCheckpoint('lpcnet5_512_10_G32np_{epoch:02d}.h5')

+checkpoint = ModelCheckpoint('lpcnet9_384_10_G16_{epoch:02d}.h5')

 #model.load_weights('wavenet4f2_30.h5')

-model.compile(optimizer=Adam(0.0005, 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=60, validation_split=0.2, callbacks=[checkpoint, lpcnet.Sparsify(1000, 20000, 200, 0.1)])

+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=120, validation_split=0.0, callbacks=[checkpoint, lpcnet.Sparsify(2000, 40000, 400, 0.1)])

-- 

⑨