feat(stt/voice): better recognition, sustained barge-in, fragment discard

STT (whisper_worker): add VAD tuning (speech_pad_ms so soft first/last words
aren't clipped), condition_on_previous_text=False + temperature fallback +
no_speech/logprob/compression thresholds to reject noisy/quiet decodes, drop
per-segment non-speech, and a hallucination guard that blanks Whisper's classic
Korean silence/noise boilerplate ("감사합니다" 등) when no_speech_prob is high.

Barge-in (bot.mjs): stop the bot's TTS only when the speaking (green ring) stays
on for >= WSAI_BARGE_IN_MS (default 700ms), not on the first blip — cancelled if
speaking stops in time. AfterSilence 800->1000ms so trailing soft words finish.

Fragment gate (dashboard): a lone connective filler ("그러면") carries no
answerable intent -> discard as [대기] instead of replying; hidden by the noise
filter like [잡음].

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

wsai/backends/whisper_worker.py

@@ -23,6 +23,31 @@ import os
 import sys
 import time
 
+# Whisper's classic Korean hallucinations on silence/noise/keyboard clatter —
+# it "hears" video-outro boilerplate. Drop these when the whole utterance is one
+# of them AND the segment looked like non-speech, so a real "감사합니다" survives.
+_HALLUCINATIONS = {
+    "감사합니다", "고맙습니다", "감사합니다.", "고맙습니다.",
+    "시청해주셔서 감사합니다", "시청해 주셔서 감사합니다", "끝까지 시청해주셔서 감사합니다",
+    "구독과 좋아요 부탁드립니다", "구독 좋아요 부탁드립니다", "다음 영상에서 만나요",
+    "다음 시간에 만나요", "안녕히 계세요",
+}
+
+
+def _norm(t: str) -> str:
+    return t.strip().rstrip(" .!?…~").strip()
+
+
+def _guard_hallucination(text: str, worst_no_speech: float) -> str:
+    """Blank out a lone known-hallucination phrase when the audio was probably
+    not speech (high no_speech_prob)."""
+    n = _norm(text)
+    if not n:
+        return ""
+    if n in {_norm(h) for h in _HALLUCINATIONS} and worst_no_speech > 0.5:
+        return ""
+    return text
+
 # Split protocol from library noise BEFORE importing anything heavy.
 _proto = os.fdopen(os.dup(1), "w", buffering=1)  # private copy of real stdout
 os.dup2(2, 1)  # fd1 -> stderr, so stray library prints don't hit the protocol
@@ -108,9 +133,31 @@ def main() -> None:
                 wav,
                 language=language,
                 beam_size=int(req.get("beam_size", 5)),
+                # VAD strips non-speech (keyboard clatter, room noise, silence)
+                # before decoding, which both improves accuracy and kills most
+                # hallucinations. speech_pad_ms keeps a little lead/trail so soft
+                # first/last words aren't clipped.
                 vad_filter=bool(req.get("vad_filter", True)),
+                vad_parameters=dict(min_silence_duration_ms=300, speech_pad_ms=250),
+                # Don't feed the previous text back in — that's what makes Whisper
+                # loop/hallucinate. Temperature fallback + thresholds reject
+                # low-confidence (noisy/quiet) decodes instead of inventing words.
+                condition_on_previous_text=False,
+                temperature=[0.0, 0.2, 0.4, 0.6],
+                no_speech_threshold=0.6,
+                log_prob_threshold=-1.0,
+                compression_ratio_threshold=2.4,
             )
-            text = "".join(seg.text for seg in segments).strip()
+            parts, worst_ns = [], 0.0
+            for seg in segments:
+                nsp = float(getattr(seg, "no_speech_prob", 0.0) or 0.0)
+                alp = float(getattr(seg, "avg_logprob", 0.0) or 0.0)
+                # Drop a segment that is almost certainly non-speech noise.
+                if nsp > 0.8 and alp < -0.4:
+                    continue
+                worst_ns = max(worst_ns, nsp)
+                parts.append(seg.text)
+            text = _guard_hallucination("".join(parts).strip(), worst_ns)
             ms = int((time.monotonic() - s) * 1000)
             _emit({"ok": True, "text": text, "language": info.language, "ms": ms})
         except Exception as exc:  # keep the worker alive across bad requests

wsai/dashboard.py

@@ -49,6 +49,20 @@ def _thought_summary(reply: str) -> str:
     return "감정 태그 없음 · 기본 톤으로 답변"
 
 
+# Lone connective/fillers that clearly expect more to come ("그러면…"). On their
+# own they carry no answerable intent, so we discard them instead of blurting a
+# confused reply. (A fuller version would buffer and wait for the continuation.)
+_FRAGMENT_FILLERS = {
+    "그러면", "그래서", "그런데", "근데", "그리고", "그러니까", "그럼", "그게",
+    "저기", "있잖아", "그", "음", "어", "저", "그러", "그니까",
+}
+
+
+def _is_fragment(text: str) -> bool:
+    t = (text or "").strip().rstrip(" .,!?…~").strip()
+    return t in _FRAGMENT_FILLERS
+
+
 def _make_handler(dash: "Dashboard"):
     monitor = dash.monitor
 
@@ -722,6 +736,13 @@ class Dashboard:
                 turn.replied("[잡음]")
                 turn.finish()
                 return {"heard": heard, "reply": "[잡음]", "wav": b""}
+            if _is_fragment(heard):
+                # A lone connective ("그러면") — no answerable intent yet. Discard
+                # rather than reply to a fragment.
+                turn.thought("불완전한 말(연결어) — 대답하지 않고 폐기")
+                turn.replied("[대기]")
+                turn.finish()
+                return {"heard": heard, "reply": "[대기]", "wav": b""}
             t_llm = time.monotonic()
             reply_text = self._think(heard)
             self._step(turn, "LLM" if self.brain else "echo", (time.monotonic() - t_llm) * 1000)
@@ -1305,7 +1326,7 @@ function turnFilter(){
     text:$('tfText').value.trim().toLowerCase() };
 }
 function isNoiseTurn(t){
-  return (t.heard==='(빈 결과)') || (t.reply==='[잡음]');
+  return (t.heard==='(빈 결과)') || (t.reply==='[잡음]') || (t.reply==='[대기]');
 }
 function turnMatches(t, f){
   if(hideNoise && isNoiseTurn(t)) return false;   // 잡음 로그 표시 안 함 토글