From e23acf1555eba4039ef350b2e111b29804377ff6 Mon Sep 17 00:00:00 2001
From: jadssss <cheater.f0r.minecraft@gmail.com>
Date: Wed, 28 Oct 2026 09:47:27 +0300
Subject: [PATCH] stop passing vad_filter to whisperx transcribe

- retry only lowers batch_size: configured, 1, 1\n- keep vad_filter parameters for future model-load-level control\n- do not log success in staged finalize when no text was produced
---
 core/pipeline.py   |  5 ++++-
 core/transcribe.py | 24 +++++++++++-------------
 2 files changed, 15 insertions(+), 14 deletions(-)

diff --git a/core/pipeline.py b/core/pipeline.py
index e41d90b..823dc1a 100644
--- a/core/pipeline.py
+++ b/core/pipeline.py
@@ -1378,7 +1378,10 @@ def _run_staged_batch(
                 config, f, text, source,
                 metadata=(ident.metadata if ident else None),
             )
-            log(f"  ✓ {f.name}")
+            if text is not None or s["resolved"]:
+                log(f"  ✓ {f.name}")
+            else:
+                log(f"  ⊘ {f.name}: пропущен (нет текста)")
             mark_overall(is_complete(s))
 
     finally:
diff --git a/core/transcribe.py b/core/transcribe.py
index 3182c62..d8a1cd9 100644
--- a/core/transcribe.py
+++ b/core/transcribe.py
@@ -74,31 +74,29 @@ def _enable_tf32() -> None:
 
 
 def _model_transcribe_with_retry(
-    model_obj, audio, language, batch_size, vad_filter, name="<audio>",
+    model_obj, audio, language, batch_size, name="<audio>",
 ):
     """Try progressively lighter settings so one long file can still finish.
 
     The loaded model is intentionally kept in the cache between attempts:
     reloading it costs minutes and does not help with an OOM caused by a long
-    audio file. If even batch=1 without VAD fails, raise CudaOOMError so the
-    caller can skip this file and continue with the rest of the batch.
+    audio file. VAD is not touched here — WhisperX expects it at model-load
+    time (``vad_model``/``vad_options``), not per transcription call. If even
+    batch=1 fails, raise CudaOOMError so the caller can skip this file and
+    continue with the rest of the batch.
     """
-    attempts = [
-        (batch_size, vad_filter),   # as configured
-        (1, False),                 # minimum memory: no Pyannote VAD, batch=1
-        (1, False),                 # one more time after empty_cache()
-    ]
+    attempts = [batch_size, 1, 1]
     last: Optional[Exception] = None
-    for i, (bs, vad) in enumerate(attempts, 1):
+    for i, bs in enumerate(attempts, 1):
         try:
             return model_obj.transcribe(
-                audio, batch_size=bs, language=language, vad_filter=vad,
+                audio, batch_size=bs, language=language,
             )
         except Exception as e:  # noqa: BLE001
             if not _is_cuda_error(e):
                 raise
             last = e
-            print(f"  [ASR] OOM on {name} (attempt {i}/3, bs={bs}, vad={vad}): {e}")
+            print(f"  [ASR] OOM on {name} (attempt {i}/3, bs={bs}): {e}")
             _free_vram()
     raise CudaOOMError(f"OOM after 3 attempts on {name}: {last}") from last
 
@@ -213,7 +211,7 @@ def _transcribe_whisperx(
     model = _WhisperXModels.get(model_name, device, compute_type, local_files_only)
     audio = whisperx.load_audio(str(audio_path))
     result = _model_transcribe_with_retry(
-        model, audio, language, batch_size, vad_filter, name=audio_path.name,
+        model, audio, language, batch_size, name=audio_path.name,
     )
     text = " ".join(seg["text"].strip() for seg in result["segments"]).strip()
     del audio, result
@@ -247,7 +245,7 @@ def transcribe_aligned(
     model_obj = _WhisperXModels.get(model, device, compute_type, local_files_only)
     audio = whisperx.load_audio(str(audio_path))
     result = _model_transcribe_with_retry(
-        model_obj, audio, language, batch_size, vad_filter, name=audio_path.name,
+        model_obj, audio, language, batch_size, name=audio_path.name,
     )
 
     detected = result.get("language") or language or "en"