diff --git a/dnn/lossgen.c b/dnn/lossgen.c
new file mode 100644
index 00000000..15505873
--- /dev/null
+++ b/dnn/lossgen.c
@@ -0,0 +1,69 @@
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <math.h>
+#include "lossgen.h"
+#include "os_support.h"
+#include "nnet.h"
+#include "lpcnet_private.h"
+
+int sample_loss(
+    LossGenState *st,
+    float percent_loss,
+    int arch
+    )
+{
+  float input[2];
+  float tmp[LOSSGEN_DENSE_IN_OUT_SIZE];
+  float out;
+  int loss;
+  LossGen *model = &st->model;
+  input[0] = st->last_loss;
+  input[1] = percent_loss;
+  compute_generic_dense(&model->lossgen_dense_in, tmp, input, ACTIVATION_TANH, arch);
+  compute_generic_gru(&model->lossgen_gru1_input, &model->lossgen_gru1_recurrent, st->gru1_state, tmp, arch);
+  compute_generic_gru(&model->lossgen_gru2_input, &model->lossgen_gru2_recurrent, st->gru2_state, st->gru1_state, arch);
+  compute_generic_dense(&model->lossgen_dense_out, &out, st->gru2_state, ACTIVATION_SIGMOID, arch);
+  loss = (float)rand()/RAND_MAX < out;
+  st->last_loss = loss;
+  return loss;
+}
+
+
+void lossgen_init(LossGenState *st)
+{
+  int ret;
+  OPUS_CLEAR(st, 1);
+#ifndef USE_WEIGHTS_FILE
+  ret = init_lossgen(&st->model, lossgen_arrays);
+#else
+  ret = 0;
+#endif
+  celt_assert(ret == 0);
+}
+
+int lossgen_load_model(LossGenState *st, const unsigned char *data, int len) {
+  WeightArray *list;
+  int ret;
+  parse_weights(&list, data, len);
+  ret = init_lossgen(&st->model, list);
+  opus_free(list);
+  if (ret == 0) return 0;
+  else return -1;
+}
+
+#if 0
+#include <stdio.h>
+int main(int argc, char **argv) {
+  int i, N;
+  float p;
+  LossGenState st;
+  lossgen_init(&st);
+  p = atof(argv[1]);
+  N = atoi(argv[2]);
+  for (i=0;i<N;i++) {
+    printf("%d\n", sample_loss(&st, p, 0));
+  }
+}
+#endif
diff --git a/dnn/lossgen.h b/dnn/lossgen.h
new file mode 100644
index 00000000..cb0460ae
--- /dev/null
+++ b/dnn/lossgen.h
@@ -0,0 +1,30 @@
+#ifndef LOSSGEN_H
+#define LOSSGEN_H
+
+
+#include "lossgen_data.h"
+
+#define PITCH_MIN_PERIOD 32
+#define PITCH_MAX_PERIOD 256
+
+#define NB_XCORR_FEATURES (PITCH_MAX_PERIOD-PITCH_MIN_PERIOD)
+
+
+typedef struct {
+  LossGen model;
+  float gru1_state[LOSSGEN_GRU1_STATE_SIZE];
+  float gru2_state[LOSSGEN_GRU2_STATE_SIZE];
+  int last_loss;
+} LossGenState;
+
+
+void lossgen_init(LossGenState *st);
+int lossgen_load_model(LossGenState *st, const unsigned char *data, int len);
+
+int sample_loss(
+    LossGenState *st,
+    float percent_loss,
+    int arch
+    );
+
+#endif
diff --git a/dnn/nnet.h b/dnn/nnet.h
index a2eaad82..7eb7a57b 100644
--- a/dnn/nnet.h
+++ b/dnn/nnet.h
@@ -147,6 +147,7 @@ extern const WeightArray rdovaedec_arrays[];
 extern const WeightArray fwgan_arrays[];
 extern const WeightArray fargan_arrays[];
 extern const WeightArray pitchdnn_arrays[];
+extern const WeightArray lossgen_arrays[];
 
 int linear_init(LinearLayer *layer, const WeightArray *arrays,
   const char *bias,
diff --git a/dnn/torch/lossgen/README.md b/dnn/torch/lossgen/README.md
index 26abc9eb..55c1b442 100644
--- a/dnn/torch/lossgen/README.md
+++ b/dnn/torch/lossgen/README.md
@@ -7,7 +7,7 @@ to build a generative model for packet loss.
 
 We use the training data provided for the Audio Deep Packet Loss Concealment Challenge, which is available at:
 
-http://plcchallenge2022pub.blob.core.windows.net/plcchallengearchive/test\_train.tar.gz
+http://plcchallenge2022pub.blob.core.windows.net/plcchallengearchive/test_train.tar.gz
 
 To create the training data, run: