diff --git a/dnn/nnet.c b/dnn/nnet.c
index 3513a606..ea64e3cf 100644
--- a/dnn/nnet.c
+++ b/dnn/nnet.c
@@ -283,7 +283,7 @@ void compute_gru2(const GRULayer *gru, float *state, const float *input)
    sgemv_accum8x4(zrh, gru->input_weights, 3*N, M, stride, input);
    for (i=0;i<3*N;i++)
       recur[i] = gru->bias[3*N + i];
-   sgemv_accum(recur, gru->recurrent_weights, 3*N, N, stride, state);
+   sgemv_accum8x4(recur, gru->recurrent_weights, 3*N, N, stride, state);
    for (i=0;i<2*N;i++)
       zrh[i] += recur[i];
    compute_activation(zrh, zrh, 2*N, ACTIVATION_SIGMOID);
@@ -326,7 +326,7 @@ void compute_gruB(const GRULayer *gru, const float* gru_b_condition, float *stat
    sparse_sgemv_accum8x4(zrh, gru->input_weights, 3*N, M, gru->input_weights_idx, input);
    for (i=0;i<3*N;i++)
       recur[i] = gru->bias[3*N + i];
-   sgemv_accum(recur, gru->recurrent_weights, 3*N, N, stride, state);
+   sgemv_accum8x4(recur, gru->recurrent_weights, 3*N, N, stride, state);
    for (i=0;i<2*N;i++)
       zrh[i] += recur[i];
    compute_activation(zrh, zrh, 2*N, ACTIVATION_SIGMOID);
@@ -361,7 +361,7 @@ void compute_gru3(const GRULayer *gru, float *state, const float *input)
    RNN_COPY(zrh, input, 3*N);
    for (i=0;i<3*N;i++)
       recur[i] = gru->bias[3*N + i];
-   sgemv_accum(recur, gru->recurrent_weights, 3*N, N, stride, state);
+   sgemv_accum8x4(recur, gru->recurrent_weights, 3*N, N, stride, state);
    for (i=0;i<2*N;i++)
       zrh[i] += recur[i];
    compute_activation(zrh, zrh, 2*N, ACTIVATION_SIGMOID);
diff --git a/dnn/nnet.h b/dnn/nnet.h
index 0c06280d..e0504e53 100644
--- a/dnn/nnet.h
+++ b/dnn/nnet.h
@@ -59,7 +59,7 @@ typedef struct {
   const float *subias;
   const qweight *input_weights;
   const int *input_weights_idx;
-  const float *recurrent_weights;
+  const qweight *recurrent_weights;
   int nb_inputs;
   int nb_neurons;
   int activation;
diff --git a/dnn/training_tf2/dump_lpcnet.py b/dnn/training_tf2/dump_lpcnet.py
index 083dc3ed..8eac0db7 100755
--- a/dnn/training_tf2/dump_lpcnet.py
+++ b/dnn/training_tf2/dump_lpcnet.py
@@ -138,7 +138,14 @@ def dump_grub(self, f, hf, gru_a_size):
     print("printing layer " + name + " of type " + self.__class__.__name__)
     weights = self.get_weights()
     qweight = printSparseVector(f, weights[0][:gru_a_size, :], name + '_weights', have_diag=False)
+
+    f.write('#ifdef DOT_PROD\n')
+    qweight = np.clip(np.round(128.*weights[1]).astype('int'), -128, 127)
+    printVector(f, qweight, name + '_recurrent_weights', dotp=True, dtype='qweight')
+    f.write('#else /*DOT_PROD*/\n')
     printVector(f, weights[1], name + '_recurrent_weights')
+    f.write('#endif /*DOT_PROD*/\n')
+
     printVector(f, weights[-1], name + '_bias')
     subias = weights[-1].copy()
     subias[0,:] = subias[0,:] - np.sum(qweight*(1./128.),axis=0)
diff --git a/dnn/training_tf2/lpcnet.py b/dnn/training_tf2/lpcnet.py
index 2f14ecd3..7b7bbf1d 100644
--- a/dnn/training_tf2/lpcnet.py
+++ b/dnn/training_tf2/lpcnet.py
@@ -259,12 +259,12 @@ def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features = 20, train
         rnn = CuDNNGRU(rnn_units1, return_sequences=True, return_state=True, name='gru_a',
               recurrent_constraint = constraint, recurrent_regularizer=quant)
         rnn2 = CuDNNGRU(rnn_units2, return_sequences=True, return_state=True, name='gru_b',
-               kernel_constraint=constraint, kernel_regularizer=quant)
+               kernel_constraint=constraint, recurrent_constraint = constraint, kernel_regularizer=quant, recurrent_regularizer=quant)
     else:
         rnn = GRU(rnn_units1, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_a',
               recurrent_constraint = constraint, recurrent_regularizer=quant)
         rnn2 = GRU(rnn_units2, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_b',
-               kernel_constraint=constraint, kernel_regularizer=quant)
+               kernel_constraint=constraint, recurrent_constraint = constraint, kernel_regularizer=quant, recurrent_regularizer=quant)
 
     rnn_in = Concatenate()([cpcm, rep(cfeat)])
     md = MDense(pcm_levels, activation='sigmoid', name='dual_fc')