update

app/core/asr_engine.py

@@ -81,7 +81,6 @@ def transcribe_file_chunks(
     """
     if not wav_path:
         return []
-
     out = model(
         wav_path,
         chunk_length_s=chunk_length_s,
@@ -89,24 +88,36 @@ def transcribe_file_chunks(
         return_timestamps=True,
     )
 
+    # Pipeline output can vary across transformers versions/models:
+    # - some return `chunks` (with `timestamp` list),
+    # - others return `segments` (with `start`/`end`),
+    # so be permissive and handle both shapes.
+    raw_segments = out.get("chunks") or out.get("segments") or []
+
     chunks = []
-    for c in out.get("chunks", []) or []:
-        ts = c.get("timestamp") or [None, None]
-        start, end = ts if len(ts) == 2 else (None, None)
+    for c in raw_segments:
+        # try multiple timestamp shapes
+        start = None
+        end = None
+
+        if isinstance(c.get("timestamp"), (list, tuple)) and len(c.get("timestamp")) >= 2:
+            ts = c.get("timestamp")
+            start, end = ts[0], ts[1]
+        elif c.get("start") is not None and c.get("end") is not None:
+            start, end = c.get("start"), c.get("end")
 
         text = (c.get("text") or "").strip()
         if not text:
             continue
 
+        # If timestamps are missing, skip (we don't want chunks without timing)
         if start is None or end is None:
             continue
 
-        chunks.append(
-            {
-                "start": float(start),
-                "end": float(end),
-                "text": text,
-            }
-        )
+        try:
+            chunks.append({"start": float(start), "end": float(end), "text": text})
+        except Exception:
+            # be robust against unexpected types
+            continue
 
     return chunks