Update util.py

util.py

@@ -172,6 +172,14 @@ class SpeakerManager:
             audio_array = audio_data
             sample_rate = 16000
 
+        # convert audio from int16 (gradio.audio() returns int16) to float32
+        audio_array = audio_array.astype(np.float32) / 32768.0
+
+        # make mono if stereo (gradio specific)! gradio returns waveform with shape (num_samples, num_channels) that not typical for torch paradigm
+        if audio_array.ndim == 2:
+            print('Make MONO from STEREO')
+            audio_array = audio_array.mean(axis=1)
+        
         # Generate embedding (SpeakerEmbedder will handle resampling if needed)
         embedding = self.embedder.embed_audio(audio_array, sample_rate=sample_rate)