Merge pull request #96 from smly/fix-batch-processing

FIX: Assertion error in batch processing

whisperx/transcribe.py

@@ -365,8 +365,6 @@ def transcribe_with_vad_parallel(
     if mel is None:
         mel = log_mel_spectrogram(audio)
 
-    output = {"segments": []}
-
     vad_segments = vad_pipeline(audio)
     # merge segments to approx 30s inputs to make whisper most appropraite
     vad_segments = merge_chunks(vad_segments)
@@ -428,7 +426,9 @@ def transcribe_with_vad_parallel(
     language = kwargs["language"]
     task = kwargs["task"]
     tokenizer = get_tokenizer(model.is_multilingual, language=language, task=task)
-    result_segments = post_process_results(
+
+    output = post_process_results(
+        vad_segments,
         decode_result, 
         duration_list, 
         offset_list,
@@ -438,29 +438,11 @@ def transcribe_with_vad_parallel(
         no_speech_threshold=no_speech_threshold,
         logprob_threshold=logprob_threshold,
         verbose=verbose)
-
-    # post processing: collect outputs
-    assert len(result_segments) == len(vad_segments)
-    for sdx, (seg_t, result) in enumerate(zip(vad_segments, result_segments)):
-        seg_t["text"] = result["text"]
-        output["segments"].append(
-            {
-                "start": seg_t["start"],
-                "end": seg_t["end"],
-                "language": result["language"],
-                "text": result["text"],
-                "seg-text": [x["text"] for x in result["segments"]],
-                "seg-start": [x["start"] for x in result["segments"]],
-                "seg-end": [x["end"] for x in result["segments"]],
-                }
-            )
-
-    output["language"] = output["segments"][0]["language"]
-
     return output
 
 
 def post_process_results(
+    vad_segments,
     result_list, 
     duration_list, 
     offset_list,
@@ -478,7 +460,7 @@ def post_process_results(
     )  # time per output token: 0.02 (seconds)
     all_tokens = []
     all_segments = []
-    outputs = []
+    output = {"segments": []}
 
     def add_segment(
         *, start: float, end: float, text_tokens: torch.Tensor, result: DecodingResult
@@ -505,7 +487,7 @@ def add_segment(
             print(f"[{format_timestamp(start)} --> {format_timestamp(end)}] {text}")
 
     # process the output
-    for result, segment_duration, timestamp_offset in zip(result_list, duration_list, offset_list):
+    for seg_t, result, segment_duration, timestamp_offset in zip(vad_segments, result_list, duration_list, offset_list):
         all_tokens = []
         all_segments = []
 
@@ -568,9 +550,22 @@ def add_segment(
             seek += segment_shape
             all_tokens.extend(tokens.tolist())
 
-        outputs.append(dict(text=tokenizer.decode(all_tokens), segments=all_segments, language=language))
+        result = dict(text=tokenizer.decode(all_tokens), segments=all_segments, language=language)
+        output["segments"].append(
+            {
+                "start": seg_t["start"],
+                "end": seg_t["end"],
+                "language": result["language"],
+                "text": result["text"],
+                "seg-text": [x["text"] for x in result["segments"]],
+                "seg-start": [x["start"] for x in result["segments"]],
+                "seg-end": [x["end"] for x in result["segments"]],
+                }
+            )
+
+    output["language"] = output["segments"][0]["language"]
 
-    return outputs
+    return output
 
 
 def cli():