Improvements from my branch

Makefile.am

@@ -22,6 +22,7 @@ librnnoise_la_SOURCES = \
 	src/denoise.c \
 	src/rnn.c \
 	src/rnn_data.c \
+	src/rnn_reader.c \
 	src/pitch.c \
 	src/kiss_fft.c \
 	src/celt_lpc.c
@@ -35,7 +36,7 @@ noinst_PROGRAMS = examples/rnnoise_demo
 endif
 
 examples_rnnoise_demo_SOURCES = examples/rnnoise_demo.c
-examples_rnnoise_demo_LDADD = librnnoise.la
+examples_rnnoise_demo_LDADD = librnnoise.la $(LIBM)
 
 pkgconfigdir = $(libdir)/pkgconfig
 pkgconfig_DATA = rnnoise.pc

README

@@ -1,4 +1,4 @@
-RNNoise is a noise suppression library based on a recurrent neural network
+RNNoise is a noise suppression library based on a recurrent neural network.
 
 To compile, just type:
 % ./autogen.sh
@@ -12,6 +12,6 @@ While it is meant to be used as a library, a simple command-line tool is
 provided as an example. It operates on RAW 16-bit (machine endian) mono
 PCM files sampled at 48 kHz. It can be used as:
 
-./examples/rnnoise_demo input.pcm output.pcm
+./examples/rnnoise_demo <number of channels> <maximum attenuation> [model file] < input.raw > output.raw
 
 The output is also a 16-bit raw PCM file.

TRAINING

@@ -0,0 +1,11 @@
+(1) cd src ; ./compile.sh
+
+(2) ./denoise_training signal.raw noise.raw count > training.f32
+
+    (note the matrix size and replace 500000 87 below)
+
+(3) cd training ; ./bin2hdf5.py ../src/training.f32 500000 87 training.h5
+
+(4) ./rnn_train.py
+
+(5) ./dump_rnn.py weights.hdf5 ../src/rnn_data.c ../src/rnn_data.h

examples/rnnoise_demo.c

@@ -1,4 +1,5 @@
-/* Copyright (c) 2017 Mozilla */
+/* Copyright (c) 2018 Gregor Richards
+ * Copyright (c) 2017 Mozilla */
 /*
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions
@@ -24,36 +25,82 @@
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
+#include <math.h>
 #include <stdio.h>
 #include "rnnoise.h"
+#include <stdlib.h>
+#include <sys/types.h>
+#include "rnnoise.h"
 
 #define FRAME_SIZE 480
 
 int main(int argc, char **argv) {
-  int i;
+  int i, ci;
   int first = 1;
+  int channels;
   float x[FRAME_SIZE];
-  FILE *f1, *fout;
-  DenoiseState *st;
-  st = rnnoise_create();
-  if (argc!=3) {
-    fprintf(stderr, "usage: %s <noisy speech> <output denoised>\n", argv[0]);
+  short *tmp;
+  RNNModel *model = NULL;
+  DenoiseState **sts;
+  float max_attenuation;
+  if (argc < 3) {
+    fprintf(stderr, "usage: %s <channels> <max attenuation dB> [model file]\n", argv[0]);
+    return 1;
+  }
+
+  channels = atoi(argv[1]);
+  if (channels < 1) channels = 1;
+  max_attenuation = pow(10, -atof(argv[2])/10);
+
+  if (argc >= 4) {
+    FILE *model_file = fopen(argv[3], "r");
+    if (!model_file) {
+      perror(argv[3]);
+      return 1;
+    }
+    model = rnnoise_model_from_file(model_file);
+    fprintf(stderr, "\n\n\n%p\n\n\n", model);
+    if (!model) {
+      perror(argv[3]);
+      return 1;
+    }
+    fclose(model_file);
+  }
+
+  sts = malloc(channels * sizeof(DenoiseState *));
+  if (!sts) {
+    perror("malloc");
     return 1;
   }
-  f1 = fopen(argv[1], "r");
-  fout = fopen(argv[2], "w");
+  tmp = malloc(channels * FRAME_SIZE * sizeof(short));
+  if (!tmp) {
+      perror("malloc");
+      return 1;
+  }
+  for (i = 0; i < channels; i++) {
+    sts[i] = rnnoise_create(model);
+    rnnoise_set_param(sts[i], RNNOISE_PARAM_MAX_ATTENUATION, max_attenuation);
+  }
+
   while (1) {
-    short tmp[FRAME_SIZE];
-    fread(tmp, sizeof(short), FRAME_SIZE, f1);
-    if (feof(f1)) break;
-    for (i=0;i<FRAME_SIZE;i++) x[i] = tmp[i];
-    rnnoise_process_frame(st, x, x);
-    for (i=0;i<FRAME_SIZE;i++) tmp[i] = x[i];
-    if (!first) fwrite(tmp, sizeof(short), FRAME_SIZE, fout);
+    fread(tmp, sizeof(short), channels * FRAME_SIZE, stdin);
+    if (feof(stdin)) break;
+
+    for (ci = 0; ci < channels; ci++) {
+        for (i=0;i<FRAME_SIZE;i++) x[i] = tmp[i*channels+ci];
+        rnnoise_process_frame(sts[ci], x, x);
+        for (i=0;i<FRAME_SIZE;i++) tmp[i*channels+ci] = x[i];
+    }
+
+    if (!first) fwrite(tmp, sizeof(short), channels * FRAME_SIZE, stdout);
     first = 0;
   }
-  rnnoise_destroy(st);
-  fclose(f1);
-  fclose(fout);
+
+  for (i = 0; i < channels; i++)
+    rnnoise_destroy(sts[i]);
+  free(tmp);
+  free(sts);
+  if (model)
+    rnnoise_model_free(model);
   return 0;
 }

include/rnnoise.h

@@ -1,4 +1,5 @@
-/* Copyright (c) 2017 Mozilla */
+/* Copyright (c) 2018 Gregor Richards
+ * Copyright (c) 2017 Mozilla */
 /*
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions
@@ -24,6 +25,12 @@
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */
 
+#ifndef RNNOISE_H
+#define RNNOISE_H 1
+
+#include <stdio.h>
+
+
 #ifndef RNNOISE_EXPORT
 # if defined(WIN32)
 #  if defined(RNNOISE_BUILD) && defined(DLL_EXPORT)
@@ -38,15 +45,26 @@
 # endif
 #endif
 
-
 typedef struct DenoiseState DenoiseState;
+typedef struct RNNModel RNNModel;
 
 RNNOISE_EXPORT int rnnoise_get_size();
 
-RNNOISE_EXPORT int rnnoise_init(DenoiseState *st);
+RNNOISE_EXPORT int rnnoise_init(DenoiseState *st, RNNModel *model);
 
-RNNOISE_EXPORT DenoiseState *rnnoise_create();
+RNNOISE_EXPORT DenoiseState *rnnoise_create(RNNModel *model);
 
 RNNOISE_EXPORT void rnnoise_destroy(DenoiseState *st);
 
 RNNOISE_EXPORT float rnnoise_process_frame(DenoiseState *st, float *out, const float *in);
+
+RNNOISE_EXPORT RNNModel *rnnoise_model_from_file(FILE *f);
+
+RNNOISE_EXPORT void rnnoise_model_free(RNNModel *model);
+
+/* Parameters to a denoise state */
+#define RNNOISE_PARAM_MAX_ATTENUATION           1
+
+RNNOISE_EXPORT void rnnoise_set_param(DenoiseState *st, int param, float value);
+
+#endif