diff --git a/dnn/training_tf2/fec_encoder.py b/dnn/training_tf2/fec_encoder.py
index 43941aba..f755d560 100644
--- a/dnn/training_tf2/fec_encoder.py
+++ b/dnn/training_tf2/fec_encoder.py
@@ -121,7 +121,7 @@ quant_gru_state_dec = pvq_quantize(gru_state_dec, 30)
 
 # rate estimate
 hard_distr_embed = tf.math.sigmoid(quant_embed_dec[:, :, 4 * nsymbols : ]).numpy()
-rate_input = np.concatenate((symbols, hard_distr_embed, enc_lambda), axis=-1)
+rate_input = np.concatenate((qsymbols, hard_distr_embed, enc_lambda), axis=-1)
 rates = sq_rate_metric(None, rate_input, reduce=False).numpy()
 
 # run decoder
@@ -133,7 +133,7 @@ packet_sizes = []
 
 for i in range(offset, num_frames):
     print(f"processing frame {i - offset}...")
-    features = decoder.predict([symbols[:, i - 2 * input_length + 2 : i + 1 : 2, :], quant_embed_dec[:, i - 2 * input_length + 2 : i + 1 : 2, :], quant_gru_state_dec[:, i, :]])
+    features = decoder.predict([qsymbols[:, i - 2 * input_length + 2 : i + 1 : 2, :], quant_embed_dec[:, i - 2 * input_length + 2 : i + 1 : 2, :], quant_gru_state_dec[:, i, :]])
     packets.append(features)
     packet_size = 8 * int((np.sum(rates[:, i - 2 * input_length + 2 : i + 1 : 2]) + 7) / 8) + 64
     packet_sizes.append(packet_size)