diff --git a/codes/scripts/audio/speech_synthesis_utils.py b/codes/scripts/audio/speech_synthesis_utils.py
index 046c1af6..af30b307 100644
--- a/codes/scripts/audio/speech_synthesis_utils.py
+++ b/codes/scripts/audio/speech_synthesis_utils.py
@@ -11,11 +11,11 @@ from trainer.injectors.base_injectors import TorchMelSpectrogramInjector
 from utils.audio import plot_spectrogram
 
 
-def wav_to_mel(wav):
+def wav_to_mel(wav, mel_norms_file='../experiments/clips_mel_norms.pth'):
     """
     Converts an audio clip into a MEL tensor that the vocoder, DVAE and GptTts models use whenever a MEL is called for.
     """
-    return TorchMelSpectrogramInjector({'in': 'wav', 'out': 'mel'},{})({'wav': wav})['mel']
+    return TorchMelSpectrogramInjector({'in': 'wav', 'out': 'mel', 'mel_norm_file': mel_norms_file},{})({'wav': wav})['mel']
 
 
 def convert_mel_to_codes(dvae_model, mel):
diff --git a/codes/trainer/injectors/base_injectors.py b/codes/trainer/injectors/base_injectors.py
index e909a31a..d82a8a84 100644
--- a/codes/trainer/injectors/base_injectors.py
+++ b/codes/trainer/injectors/base_injectors.py
@@ -632,7 +632,7 @@ class TorchMelSpectrogramInjector(Injector):
 
 def test_torch_mel_injector():
     a = load_audio('D:\\data\\audio\\libritts\\train-clean-100\\19\\198\\19_198_000000_000000.wav', 22050)
-    inj = TorchMelSpectrogramInjector({'in': 'in', 'out': 'out'}, {})
+    inj = TorchMelSpectrogramInjector({'in': 'in', 'out': 'out', 'mel_norm_file': '../experiments/clips_mel_norms.pth'}, {})
     f = inj({'in': a.unsqueeze(0)})['out']
     plot_spectrogram(f[0])
     inj = MelSpectrogramInjector({'in': 'in', 'out': 'out'}, {})