async queue pipeline

asr.py

@@ -11,7 +11,7 @@ import wave
 
 SAMPLE_RATE = 16000
 GROQ_TIMEOUT = 5.0
-VAD_THRESHOLD = 0.5
+VAD_THRESHOLD = 0.7
 
 GROQ_KEYS = [
     "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",
@@ -26,6 +26,17 @@ BLACKLIST = [
     "indonesia.",
     "terima kasih",
     "thank you",
+    "thanks",
+    "subscribe",
+    "like",
+    "comment",
+]
+
+ENGLISH_STARTERS = [
+    "so ", "we ", "the ", "this ", "that ", "i ", "you ", "it ",
+    "and ", "but ", "for ", "are ", "is ", "was ", "in ", "of ",
+    "a ", "an ", "to ", "with ", "my ", "your ", "our ", "they ",
+    "he ", "she ", "what ", "how ", "when ", "where ", "why ",
 ]
 
 def audio_to_wav_bytes(audio: np.ndarray, sample_rate: int = SAMPLE_RATE) -> bytes:
@@ -91,6 +102,10 @@ class ASRPipeline:
             confidences.append(conf)
         return max(confidences) if confidences else 0.0
 
+    def has_enough_energy(self, audio: np.ndarray) -> bool:
+        rms = np.sqrt(np.mean(audio ** 2))
+        return rms > 0.01
+
     def is_valid_text(self, text: str) -> bool:
         text_lower = text.lower().strip()
         if len(text_lower) < 3:
@@ -98,6 +113,9 @@ class ASRPipeline:
         for bl in BLACKLIST:
             if text_lower.startswith(bl):
                 return False
+        for starter in ENGLISH_STARTERS:
+            if text_lower.startswith(starter):
+                return False
         return True
 
     def groq_transcribe(self, audio: np.ndarray) -> str:
@@ -145,6 +163,9 @@ class ASRPipeline:
         if vad_conf <= VAD_THRESHOLD:
             return None
 
+        if not self.has_enough_energy(audio_chunk):
+            return None
+
         text = self.groq_transcribe(audio_chunk)
         if not text:
             return None