code

app.py

@@ -33,7 +33,8 @@ def transcribe(audio: tuple[int, np.ndarray], conversation: list[dict], gradio_c
     original_sr = audio[0]
     target_sr = 16000
 
-    audio_sr = librosa.resample(audio[1], orig_sr=original_sr, target_sr=target_sr)
+    audio_sr = librosa.resample(audio[1].astype(np.float32) / 32768.0,
+                                orig_sr=original_sr, target_sr=target_sr)
    
     output = pipe({"audio": audio_sr, "turns": conversation, "sampling_rate": target_sr},
                   max_new_tokens=512)