From c545916ac1bba20112a98c3a75b3280b3042186e Mon Sep 17 00:00:00 2001
From: Pinapelz <yukais@pinapelz.com>
Date: Sun, 26 Apr 2026 21:31:26 -0700
Subject: add filtering for common whisper hallucination words

---
 config.py | 34 ++++++++++++++++++++++++++++++++++
 server.py | 15 ++++++++-------
 2 files changed, 42 insertions(+), 7 deletions(-)

diff --git a/config.py b/config.py
index 1101085..676e8e1 100644
--- a/config.py
+++ b/config.py
@@ -22,3 +22,37 @@ _SYSTEM_PROMPT: str = (
     "     output an empty string and nothing else.\n"
     "  7. Output ONLY the final subtitle text. No labels, no explanations."
 )
+
+_LLM_EMPTY_SENTINELS: frozenset = frozenset({
+    "empty string", "empty", "(empty)", "[empty]",
+    "(empty string)", "[empty string]", "(none)", "none", "n/a",
+})
+
+_HALLUCINATION_PHRASES: frozenset = frozenset({
+    "empty string",
+    "thank you for watching",
+    "thanks for watching",
+    "thank you for watching this video",
+    "thanks for watching this video",
+    "thank you for watching and ill see you next time",
+    "thank you for watching ill see you in the next video",
+    "thanks for watching ill see you next time",
+    "thank you so much for watching",
+    "thanks for watching and ill see you in the next one",
+    "see you next time",
+    "see you in the next video",
+    "ill see you in the next video",
+    "ill see you next time",
+    "see you later",
+    "bye bye",
+    "please subscribe",
+    "like and subscribe",
+    "dont forget to like and subscribe",
+    "please like and subscribe",
+    "subscribe to my channel",
+    "if you enjoyed this video please like and subscribe",
+    "you",
+    "uh",
+    "um",
+    "hmm",
+})
diff --git a/server.py b/server.py
index 9d0b396..cea2c8a 100644
--- a/server.py
+++ b/server.py
@@ -16,7 +16,7 @@ import sounddevice as sd
 from faster_whisper import WhisperModel
 from gui import select_settings, prompt_input_sample_rate
 from routes import register_routes
-from config import _SYSTEM_PROMPT
+from config import _SYSTEM_PROMPT, _LLM_EMPTY_SENTINELS, _HALLUCINATION_PHRASES
 
 TARGET_SAMPLE_RATE: int = 16000
 CAPTURE_SAMPLE_RATE: int = 0
@@ -47,7 +47,7 @@ DEFAULT_SETTINGS: Dict[str, Any] = {
     "use_ollama_cleanup": True,
     "ollama_device": "GPU",
     "ollama_context_window": 5,
-    "ollama_raw_batch_size": 2,
+    "ollama_raw_batch_size": 1,
 }
 
 MODEL_CHOICES: List[str] = ["tiny", "base", "small", "medium", "large-v2", "large-v3", "distil-large-v3"]
@@ -173,11 +173,6 @@ def ensure_ollama_ready() -> None:
     except Exception as exc:
         raise RuntimeError(f"Ollama warm-up failed: {exc}") from exc
 
-_LLM_EMPTY_SENTINELS: frozenset = frozenset({
-    "empty string", "empty", "(empty)", "[empty]",
-    "(empty string)", "[empty string]", "(none)", "none", "n/a",
-})
-
 
 def normalize_llm_output(text: str) -> str:
     if text.strip().lower().rstrip(".") in _LLM_EMPTY_SENTINELS:
@@ -214,6 +209,12 @@ def is_hallucination(text: str) -> bool:
     if count >= 4 and count / len(clean) > 0.40:
         print(f"🔴 Hallucination (\'{top}\' x{count}, {count/len(clean):.0%}): {text[:60]!r}")
         return True
+
+    normalized = re.sub(r"[^\w\s]", "", text.lower()).strip()
+    if normalized in _HALLUCINATION_PHRASES:
+        print(f"🔴 Hallucination (blocked phrase): {text!r}")
+        return True
+
     return False
 
 
-- 
cgit v1.2.3