Merge pull request #507 from compasspathways/fix/pass-vad-options

Fix: Allow vad options to be configurable by passing to FasterWhisperPipeline and merge_chunks.

whisperx/asr.py

@@ -111,6 +111,7 @@ def load_model(whisper_arch,
         tokenizer=tokenizer,
         language=language,
         suppress_numerals=suppress_numerals,
+        vad_params=default_vad_options,
     )
 
 class WhisperModel(faster_whisper.WhisperModel):
@@ -186,6 +187,7 @@ class FasterWhisperPipeline(Pipeline):
             self,
             model,
             vad,
+            vad_params: dict,
             options : NamedTuple,
             tokenizer=None,
             device: Union[int, str, "torch.device"] = -1,
@@ -218,6 +220,7 @@ class FasterWhisperPipeline(Pipeline):
 
         super(Pipeline, self).__init__()
         self.vad_model = vad
+        self._vad_params = vad_params
 
     def _sanitize_parameters(self, **kwargs):
         preprocess_kwargs = {}
@@ -266,7 +269,12 @@ class FasterWhisperPipeline(Pipeline):
                 yield {'inputs': audio[f1:f2]}
 
         vad_segments = self.vad_model({"waveform": torch.from_numpy(audio).unsqueeze(0), "sample_rate": SAMPLE_RATE})
-        vad_segments = merge_chunks(vad_segments, chunk_size)
+        vad_segments = merge_chunks(
+            vad_segments,
+            chunk_size,
+            onset=self._vad_params["vad_onset"],
+            offset=self._vad_params["vad_offset"],
+        )
         if self.tokenizer is None:
             language = language or self.detect_language(audio)
             task = task or "transcribe"

whisperx/vad.py

@@ -260,7 +260,12 @@ def merge_vad(vad_arr, pad_onset=0.0, pad_offset=0.0, min_duration_off=0.0, min_
     active_segs = pd.DataFrame([x['segment'] for x in active['content']])
     return active_segs
 
-def merge_chunks(segments, chunk_size):
+def merge_chunks(
+    segments,
+    chunk_size,
+    onset: float = 0.5,
+    offset: Optional[float] = None,
+):
     """
     Merge operation described in paper
     """
@@ -270,7 +275,7 @@ def merge_chunks(segments, chunk_size):
     speaker_idxs = []
 
     assert chunk_size > 0
-    binarize = Binarize(max_duration=chunk_size)
+    binarize = Binarize(max_duration=chunk_size, onset=onset, offset=offset)
     segments = binarize(segments)
     segments_list = []
     for speech_turn in segments.get_timeline():