diff --git a/src/diffusers/pipelines/audioldm/pipeline_audioldm.py b/src/diffusers/pipelines/audioldm/pipeline_audioldm.py
index 86a8fd659046..c80e6ef99288 100644
--- a/src/diffusers/pipelines/audioldm/pipeline_audioldm.py
+++ b/src/diffusers/pipelines/audioldm/pipeline_audioldm.py
@@ -293,7 +293,7 @@ def mel_spectrogram_to_waveform(self, mel_spectrogram):
 
         waveform = self.vocoder(mel_spectrogram)
         # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
-        waveform = waveform.cpu()
+        waveform = waveform.cpu().float()
         return waveform
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs