Big noise in the wave start about lpcnet

[bpotard]

This is a small bug with initialising the µlaw variables. They shouldn't be initialised to 0, but to 128.

--- a/src/test_lpcnet.py
+++ b/src/test_lpcnet.py
@@ -81,6 +81,8 @@ fout = open(out_file, 'wb')
 skip = order + 1
 for c in range(0, nb_frames):
     cfeat = enc.predict([features[c:c+1, :, :nb_used_features], periods[c:c+1, :, :]])
+    fexc[0,0,0] = 128 # 0 mulaw
+    iexc[0,0,0] = 128
     for fr in range(0, feature_chunk_size):
         f = c*feature_chunk_size + fr
         a = features[c, fr, nb_features-order:]
--- a/src/lpcnet.c
+++ b/src/lpcnet.c
@@ -110,6 +110,7 @@ LPCNetState *lpcnet_create()
 {
     LPCNetState *lpcnet;
     lpcnet = (LPCNetState *)calloc(sizeof(LPCNetState), 1);
+    lpcnet->last_exc = 128;
     return lpcnet;
 }

[attitudechunfeng]

I didn't met this problem before. An irrelevant question, what's the player used in the picture?

[hdmjdp]

@candlewill I have the same, but not only in the start, have you solved this problem?

[ohleo]

@candlewill @hdmjdp
I also have the same, synthesized by C on CPU inference.
In Python-GPU inference, there is no noise in the wave start.

LJ050-0269.zip

[hdmjdp]

@ohleo have you found the reason?

[hdmjdp]

@bpotard I think you are right. but why sometimes it did not have the noise in the start even that they are initialised to 0.

[candlewill]

I think the problem is solved by @bpotard 's method, and it's better to be merged into master.