Retrained speech/music RNN

Larger network, using 8-bit weights

Author: jmvalin

src/analysis.c

@@ -867,7 +867,7 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt
     info->activity_probability = frame_probs[1];
     /* It seems like the RNN tends to have a bias towards speech and this
        warping of the probabilities compensates for it. */
-    info->music_prob = frame_probs[0] * (2 - frame_probs[0]);
+    info->music_prob = MAX16(1-10*(1-frame_probs[0]), MIN16(10*frame_probs[0], .12+.69*frame_probs[0]*(2-frame_probs[0])));
 
     /*printf("%f %f %f\n", frame_probs[0], frame_probs[1], info->music_prob);*/
 #ifdef MLP_TRAINING

src/mlp.h

@@ -29,22 +29,22 @@
 
 #include "opus_types.h"
 
-#define WEIGHTS_SCALE (1.f/8192)
+#define WEIGHTS_SCALE (1.f/128)
 
-#define MAX_NEURONS 20
+#define MAX_NEURONS 32
 
 typedef struct {
-  const opus_int16 *bias;
-  const opus_int16 *input_weights;
+  const opus_int8 *bias;
+  const opus_int8 *input_weights;
   int nb_inputs;
   int nb_neurons;
   int sigmoid;
 } DenseLayer;
 
 typedef struct {
-  const opus_int16 *bias;
-  const opus_int16 *input_weights;
-  const opus_int16 *recurrent_weights;
+  const opus_int8 *bias;
+  const opus_int8 *input_weights;
+  const opus_int8 *recurrent_weights;
   int nb_inputs;
   int nb_neurons;
 } GRULayer;