Update inference_gradio.py

inference_gradio.py

@@ -396,6 +396,7 @@ class WhisperxModel:
     def __init__(self, model_name):
         from pathlib import Path
         import whisperx.vad as wx_vad
+        import whisperx.asr as wx_asr
         from whisperx import load_model
         prompt = None  # "This might be a blend of Simplified Chinese and English speech, do not translate, only transcription be allowed."
 
@@ -408,8 +409,9 @@ class WhisperxModel:
             )
             vad_fp = None
         else:
-            # Monkey-patch whisperx.vad.load_vad_model so it loads our local
-            # segmentation model without enforcing the baked-in SHA256 check.
+            # Monkey-patch whisperx *before* constructing the pipeline so it
+            # loads our local segmentation model without enforcing the
+            # baked-in SHA256 checksum.
             def _patched_load_vad_model(device, vad_onset=0.500, vad_offset=0.363, use_auth_token=None, model_fp=None):
                 import torch
                 from pyannote.audio import Model
@@ -430,7 +432,10 @@ class WhisperxModel:
                 vad_pipeline.instantiate(hyperparameters)
                 return vad_pipeline
 
+            # asr.py does `from .vad import load_vad_model`, so we must patch
+            # both the `vad` module and the alias in `asr`.
             wx_vad.load_vad_model = _patched_load_vad_model
+            wx_asr.load_vad_model = _patched_load_vad_model
 
         self.model = load_model(
             model_name,