shithub: opus

Info  •  Files  •  Log  •  Branches

Chopping silence from the training data

--- a/dnn/denoise.c

+++ b/dnn/denoise.c

@@ -633,6 +633,9 @@

   signed char iexc[FRAME_SIZE];

   short pred[FRAME_SIZE];

   short pcm[FRAME_SIZE];

+  short tmp[FRAME_SIZE] = {0};

+  float savedX[FRAME_SIZE] = {0};

+  int last_silent = 1;

   DenoiseState *st;

   st = rnnoise_create();

   if (argc!=4) {

@@ -649,18 +652,33 @@

     float Ln[NB_BANDS];

     float features[NB_FEATURES];

     float g[NB_BANDS];

-    short tmp[FRAME_SIZE];

     float E=0;

+    int silent;

+    for (i=0;i<FRAME_SIZE;i++) x[i] = tmp[i];

     fread(tmp, sizeof(short), FRAME_SIZE, f1);

     if (feof(f1)) {

       return 0;

     }

-    for (i=0;i<FRAME_SIZE;i++) x[i] = tmp[i];

-    for (i=0;i<FRAME_SIZE;i++) x[i] += rand()/(float)RAND_MAX - .5;

     for (i=0;i<FRAME_SIZE;i++) E += tmp[i]*(float)tmp[i];

+    silent = E < 5000 || (last_silent && E < 20000);

+    if (!last_silent && silent) {

+      for (i=0;i<FRAME_SIZE;i++) savedX[i] = x[i];

+    }

+    if (last_silent && !silent) {

+        for (i=0;i<FRAME_SIZE;i++) {

+          float f = (float)i/FRAME_SIZE;

+          tmp[i] = (int)floor(.5 + f*tmp[i] + (1-f)*savedX[i]);

+        }

+    }

+    if (last_silent) {

+      last_silent = silent;

+      continue;

+    }

+    last_silent = silent;

     biquad(x, mem_hp_x, x, b_hp, a_hp, FRAME_SIZE);

     preemphasis(x, &mem_preemph, x, PREEMPHASIS, FRAME_SIZE);

+    for (i=0;i<FRAME_SIZE;i++) x[i] += rand()/(float)RAND_MAX - .5;

     compute_frame_features(st, iexc, pred, pcm, X, P, Ex, Ep, Exp, features, x);

     fwrite(features, sizeof(float), NB_FEATURES, ffeat);

     fwrite(pcm, sizeof(short), FRAME_SIZE, fpcm);

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -136,7 +136,7 @@

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

 # dump models to disk as we go

-checkpoint = ModelCheckpoint('lpcnet9b_384_10_G16_{epoch:02d}.h5')

+checkpoint = ModelCheckpoint('lpcnet9c_384_10_G16_{epoch:02d}.h5')

 #model.load_weights('wavenet4f2_30.h5')

 model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])

-- 

⑨