her

app.py

@@ -6,17 +6,21 @@ from typing import Any, Awaitable, Callable
 
 from fastapi import FastAPI, WebSocket, WebSocketDisconnect
 from fastapi.responses import FileResponse, JSONResponse
+from fastapi.staticfiles import StaticFiles
 
 from supertonic_gateway import SuperTonicGateway
 from voice_rtc import WebRTCVoiceSession
 
 
 BASE_DIR = Path(__file__).resolve().parent
-INDEX_PATH = BASE_DIR / "static" / "index.html"
+STATIC_DIR = BASE_DIR / "static"
+INDEX_PATH = STATIC_DIR / "index.html"
 
 app = FastAPI(title="Nanobot SuperTonic Wisper Web")
 gateway = SuperTonicGateway()
 
+app.mount("/static", StaticFiles(directory=str(STATIC_DIR)), name="static")
+
 
 @app.get("/health")
 async def health() -> JSONResponse:
@@ -64,13 +68,15 @@ async def websocket_chat(websocket: WebSocket) -> None:
                 voice_session.set_push_to_talk_pressed(
                     bool(message.get("pressed", False))
                 )
+            elif msg_type == "user-message":
+                await gateway.send_user_message(str(message.get("text", "")))
             else:
                 await safe_send_json(
                     {
                         "role": "system",
                         "text": (
                             "Unknown message type. Use spawn, stop, rtc-offer, "
-                            "rtc-ice-candidate, or voice-ptt."
+                            "rtc-ice-candidate, voice-ptt, or user-message."
                         ),
                         "timestamp": "",
                     }

start.sh

@@ -36,7 +36,7 @@ export NANOBOT_SUPPRESS_NOISY_UI NANOBOT_OUTPUT_DEDUP_WINDOW_S
 # Host voice pipeline env vars (safe defaults).
 : "${HOST_STT_PROVIDER:=faster-whisper}"
 : "${HOST_STT_COMMAND:=}"
-: "${HOST_STT_MODEL:=base.en}"
+: "${HOST_STT_MODEL:=distil-large-v3}"
 : "${HOST_STT_DEVICE:=auto}"
 : "${HOST_STT_COMPUTE_TYPE:=int8}"
 : "${HOST_STT_LANGUAGE:=en}"
@@ -48,45 +48,39 @@ export NANOBOT_SUPPRESS_NOISY_UI NANOBOT_OUTPUT_DEDUP_WINDOW_S
 : "${HOST_STT_NO_SPEECH_THRESHOLD:=0.6}"
 : "${HOST_STT_COMPRESSION_RATIO_THRESHOLD:=2.4}"
 : "${HOST_STT_INITIAL_PROMPT:=Transcribe brief spoken English precisely. Prefer common words over sound effects.}"
+: "${HOST_STT_REPETITION_PENALTY:=1.0}"
+: "${HOST_STT_HALLUCINATION_SILENCE_THRESHOLD:=}"
 : "${HOST_TTS_PROVIDER:=supertonic}"
 : "${HOST_TTS_COMMAND:=}"
 : "${SUPERTONIC_MODEL:=supertonic-2}"
-: "${SUPERTONIC_VOICE_STYLE:=M1}"
+: "${SUPERTONIC_VOICE_STYLE:=F1}"
 : "${SUPERTONIC_LANG:=en}"
-: "${SUPERTONIC_TOTAL_STEPS:=5}"
-: "${SUPERTONIC_SPEED:=1.05}"
+: "${SUPERTONIC_TOTAL_STEPS:=8}"
+: "${SUPERTONIC_SPEED:=1.5}"
 : "${SUPERTONIC_INTRA_OP_THREADS:=1}"
 : "${SUPERTONIC_INTER_OP_THREADS:=1}"
 : "${SUPERTONIC_AUTO_DOWNLOAD:=1}"
 : "${HOST_STT_MIN_PTT_MS:=220}"
-: "${HOST_STT_MAX_PTT_MS:=12000}"
 : "${HOST_STT_SEGMENT_QUEUE_SIZE:=2}"
 : "${HOST_STT_BACKLOG_NOTICE_INTERVAL_S:=6.0}"
 : "${HOST_STT_SUPPRESS_DURING_TTS:=1}"
 : "${HOST_STT_SUPPRESS_MS_AFTER_TTS:=300}"
 : "${HOST_RTC_OUTBOUND_LEAD_IN_MS:=120}"
 : "${HOST_RTC_OUTBOUND_IDLE_S:=0.6}"
-: "${HOST_TTS_FLUSH_DELAY_S:=0.45}"
-: "${HOST_TTS_SENTENCE_FLUSH_DELAY_S:=0.15}"
-: "${HOST_TTS_MIN_CHARS:=10}"
-: "${HOST_TTS_MAX_WAIT_MS:=1800}"
-: "${HOST_TTS_MAX_CHUNK_CHARS:=140}"
 
 export HOST_STT_PROVIDER HOST_STT_COMMAND HOST_STT_MODEL HOST_STT_DEVICE
 export HOST_STT_COMPUTE_TYPE HOST_STT_LANGUAGE HOST_STT_BEAM_SIZE HOST_STT_BEST_OF HOST_STT_VAD_FILTER
 export HOST_STT_TEMPERATURE HOST_STT_LOG_PROB_THRESHOLD HOST_STT_NO_SPEECH_THRESHOLD
 export HOST_STT_COMPRESSION_RATIO_THRESHOLD
-export HOST_STT_INITIAL_PROMPT
+export HOST_STT_INITIAL_PROMPT HOST_STT_REPETITION_PENALTY HOST_STT_HALLUCINATION_SILENCE_THRESHOLD
 export HOST_TTS_PROVIDER HOST_TTS_COMMAND
 export SUPERTONIC_MODEL SUPERTONIC_VOICE_STYLE SUPERTONIC_LANG
 export SUPERTONIC_TOTAL_STEPS SUPERTONIC_SPEED
 export SUPERTONIC_INTRA_OP_THREADS SUPERTONIC_INTER_OP_THREADS SUPERTONIC_AUTO_DOWNLOAD
-export HOST_STT_MIN_PTT_MS HOST_STT_MAX_PTT_MS HOST_STT_SEGMENT_QUEUE_SIZE
+export HOST_STT_MIN_PTT_MS HOST_STT_SEGMENT_QUEUE_SIZE
 export HOST_STT_BACKLOG_NOTICE_INTERVAL_S
 export HOST_STT_SUPPRESS_DURING_TTS HOST_STT_SUPPRESS_MS_AFTER_TTS
 export HOST_RTC_OUTBOUND_LEAD_IN_MS HOST_RTC_OUTBOUND_IDLE_S
-export HOST_TTS_FLUSH_DELAY_S HOST_TTS_SENTENCE_FLUSH_DELAY_S
-export HOST_TTS_MIN_CHARS HOST_TTS_MAX_WAIT_MS HOST_TTS_MAX_CHUNK_CHARS
 
 : "${UVICORN_HOST:=0.0.0.0}"
 : "${UVICORN_PORT:=8000}"

supertonic_gateway.py

@@ -18,26 +18,35 @@ CONTROL_CHAR_RE = re.compile(r"[\x00-\x08\x0b-\x1f\x7f]")
 BRAILLE_SPINNER_RE = re.compile(r"[\u2800-\u28ff]")
 SPINNER_ONLY_RE = re.compile(r"^[\s|/\\\-]+$")
 BOX_DRAWING_ONLY_RE = re.compile(r"^[\s\u2500-\u257f]+$")
-THINKING_LINE_RE = re.compile(r"^(?:agent|nanobot)\s+is\s+thinking\b", re.IGNORECASE)
+THINKING_LINE_RE = re.compile(
+    r"\b(?:agent|nanobot|napbot)\b(?:\s+is)?\s+thinking\b",
+    re.IGNORECASE,
+)
+USER_ECHO_LINE_RE = re.compile(r"^(?:you|user)\s*:", re.IGNORECASE)
 TOOL_STREAM_LINE_RE = re.compile(
     r"^(?:tool(?:\s+call|\s+output)?|calling\s+tool|running\s+tool|executing\s+tool)\b",
     re.IGNORECASE,
 )
+LEADING_NON_WORD_RE = re.compile(r"^[^\w]+")
+WHITESPACE_RE = re.compile(r"\s+")
+AGENT_OUTPUT_PREFIX_RE = re.compile(
+    r"^(?:nanobot|napbot)\b\s*[:>\-]?\s*", re.IGNORECASE
+)
 EMOJI_RE = re.compile(
     "["  # Common emoji and pictograph blocks.
-    "\U0001F1E6-\U0001F1FF"
-    "\U0001F300-\U0001F5FF"
-    "\U0001F600-\U0001F64F"
-    "\U0001F680-\U0001F6FF"
-    "\U0001F700-\U0001F77F"
-    "\U0001F780-\U0001F7FF"
-    "\U0001F800-\U0001F8FF"
-    "\U0001F900-\U0001F9FF"
-    "\U0001FA00-\U0001FAFF"
-    "\u2600-\u26FF"
-    "\u2700-\u27BF"
-    "\uFE0F"
-    "\u200D"
+    "\U0001f1e6-\U0001f1ff"
+    "\U0001f300-\U0001f5ff"
+    "\U0001f600-\U0001f64f"
+    "\U0001f680-\U0001f6ff"
+    "\U0001f700-\U0001f77f"
+    "\U0001f780-\U0001f7ff"
+    "\U0001f800-\U0001f8ff"
+    "\U0001f900-\U0001f9ff"
+    "\U0001fa00-\U0001faff"
+    "\u2600-\u26ff"
+    "\u2700-\u27bf"
+    "\ufe0f"
+    "\u200d"
     "]"
 )
 
@@ -70,7 +79,10 @@ def _resolve_nanobot_command_and_workdir() -> tuple[str, Path]:
         if nanobot_venv_python.exists():
             if not workdir_override:
                 default_workdir = nanobot_dir
-            return f"{nanobot_venv_python} -m nanobot agent --no-markdown", default_workdir
+            return (
+                f"{nanobot_venv_python} -m nanobot agent --no-markdown",
+                default_workdir,
+            )
 
     return "nanobot agent --no-markdown", default_workdir
 
@@ -80,7 +92,11 @@ def _infer_venv_root(command_parts: list[str], workdir: Path) -> Path | None:
         return None
 
     binary = Path(command_parts[0]).expanduser()
-    if binary.is_absolute() and binary.name.startswith("python") and binary.parent.name == "bin":
+    if (
+        binary.is_absolute()
+        and binary.name.startswith("python")
+        and binary.parent.name == "bin"
+    ):
         return binary.parent.parent
 
     for candidate in (workdir / ".venv", workdir / "venv"):
@@ -89,7 +105,9 @@ def _infer_venv_root(command_parts: list[str], workdir: Path) -> Path | None:
     return None
 
 
-def _build_process_env(command_parts: list[str], workdir: Path) -> tuple[dict[str, str], Path | None]:
+def _build_process_env(
+    command_parts: list[str], workdir: Path
+) -> tuple[dict[str, str], Path | None]:
     env = os.environ.copy()
     env.pop("PYTHONHOME", None)
 
@@ -115,14 +133,18 @@ class NanobotTUIProcess:
         self._master_fd: int | None = None
         self._read_task: asyncio.Task[None] | None = None
         self._pending_output = ""
-        self._suppress_noisy_ui = os.getenv("NANOBOT_SUPPRESS_NOISY_UI", "1").strip() not in {
+        self._suppress_noisy_ui = os.getenv(
+            "NANOBOT_SUPPRESS_NOISY_UI", "1"
+        ).strip() not in {
             "0",
             "false",
             "False",
             "no",
             "off",
         }
-        self._dedup_window_s = max(0.2, float(os.getenv("NANOBOT_OUTPUT_DEDUP_WINDOW_S", "1.5")))
+        self._dedup_window_s = max(
+            0.2, float(os.getenv("NANOBOT_OUTPUT_DEDUP_WINDOW_S", "1.5"))
+        )
         self._recent_lines: deque[tuple[str, float]] = deque()
         self._last_tts_line = ""
 
@@ -132,14 +154,19 @@ class NanobotTUIProcess:
 
     async def start(self) -> None:
         if self.running:
-            await self._bus.publish(WisperEvent(role="system", text="Nanobot TUI is already running."))
+            await self._bus.publish(
+                WisperEvent(role="system", text="Nanobot TUI is already running.")
+            )
             return
 
         command_parts = [
-            os.path.expandvars(os.path.expanduser(part)) for part in shlex.split(self._command)
+            os.path.expandvars(os.path.expanduser(part))
+            for part in shlex.split(self._command)
         ]
         if not command_parts:
-            await self._bus.publish(WisperEvent(role="system", text="NANOBOT_COMMAND is empty."))
+            await self._bus.publish(
+                WisperEvent(role="system", text="NANOBOT_COMMAND is empty.")
+            )
             return
 
         if not self._workdir.exists():
@@ -152,7 +179,9 @@ class NanobotTUIProcess:
             return
 
         master_fd, slave_fd = pty.openpty()
-        child_env, child_venv_root = _build_process_env(command_parts=command_parts, workdir=self._workdir)
+        child_env, child_venv_root = _build_process_env(
+            command_parts=command_parts, workdir=self._workdir
+        )
         try:
             self._process = subprocess.Popen(
                 command_parts,
@@ -188,7 +217,9 @@ class NanobotTUIProcess:
         os.close(slave_fd)
         os.set_blocking(master_fd, False)
         self._master_fd = master_fd
-        self._read_task = asyncio.create_task(self._read_output(), name="nanobot-tui-reader")
+        self._read_task = asyncio.create_task(
+            self._read_output(), name="nanobot-tui-reader"
+        )
         await self._bus.publish(
             WisperEvent(
                 role="system",
@@ -206,14 +237,18 @@ class NanobotTUIProcess:
     async def send(self, text: str) -> None:
         if not self.running or self._master_fd is None:
             await self._bus.publish(
-                WisperEvent(role="system", text="Nanobot TUI is not running. Click spawn first.")
+                WisperEvent(
+                    role="system", text="Nanobot TUI is not running. Click spawn first."
+                )
             )
             return
         message = text.rstrip("\n") + "\n"
         try:
             os.write(self._master_fd, message.encode())
         except OSError as exc:
-            await self._bus.publish(WisperEvent(role="system", text=f"Failed to write to TUI: {exc}"))
+            await self._bus.publish(
+                WisperEvent(role="system", text=f"Failed to write to TUI: {exc}")
+            )
 
     async def stop(self) -> None:
         if self._read_task:
@@ -251,29 +286,40 @@ class NanobotTUIProcess:
         if self._master_fd is None:
             return
         while self.running:
+            if not await self._wait_for_fd_readable():
+                break
             try:
                 chunk = os.read(self._master_fd, 4096)
             except BlockingIOError:
-                await asyncio.sleep(0.05)
                 continue
             except OSError:
                 break
 
             if not chunk:
-                await asyncio.sleep(0.05)
+                if not self.running:
+                    break
+                await asyncio.sleep(0.01)
                 continue
 
             text = _clean_output(chunk.decode(errors="ignore"))
             if not text.strip():
                 continue
 
-            displayable, tts_publishable = self._consume_output_chunk(text)
+            displayable, tts_publishable, saw_thinking = self._consume_output_chunk(
+                text
+            )
+            if saw_thinking:
+                await self._bus.publish(
+                    WisperEvent(role="agent-state", text="thinking")
+                )
             if displayable:
                 await self._bus.publish(WisperEvent(role="nanobot", text=displayable))
             if tts_publishable:
-                await self._bus.publish(WisperEvent(role="nanobot-tts", text=tts_publishable))
+                await self._bus.publish(
+                    WisperEvent(role="nanobot-tts", text=tts_publishable)
+                )
 
-        trailing_display, trailing_tts = self._consume_output_chunk("\n")
+        trailing_display, trailing_tts, _ = self._consume_output_chunk("\n")
         if trailing_display:
             await self._bus.publish(WisperEvent(role="nanobot", text=trailing_display))
         if trailing_tts:
@@ -282,10 +328,13 @@ class NanobotTUIProcess:
         if self._process is not None:
             exit_code = self._process.poll()
             await self._bus.publish(
-                WisperEvent(role="system", text=f"Nanobot TUI exited (code={exit_code}).")
+                WisperEvent(
+                    role="system", text=f"Nanobot TUI exited (code={exit_code})."
+                )
             )
 
-    def _consume_output_chunk(self, text: str) -> tuple[str, str]:
+    def _consume_output_chunk(self, text: str) -> tuple[str, str, bool]:
+        """Return (displayable, tts_publishable, saw_thinking)."""
         self._pending_output += text
 
         lines = self._pending_output.split("\n")
@@ -297,11 +346,16 @@ class NanobotTUIProcess:
 
         kept_lines: list[str] = []
         tts_lines: list[str] = []
+        saw_thinking = False
         for line in lines:
             normalized = self._normalize_line(line)
             if not normalized:
                 continue
             if self._suppress_noisy_ui and self._is_noisy_ui_line(normalized):
+                # Detect thinking lines even though they are filtered from display.
+                candidate = LEADING_NON_WORD_RE.sub("", normalized)
+                if THINKING_LINE_RE.search(candidate):
+                    saw_thinking = True
                 continue
             if normalized != self._last_tts_line:
                 tts_lines.append(normalized)
@@ -310,11 +364,16 @@ class NanobotTUIProcess:
                 continue
             kept_lines.append(normalized)
 
-        return "\n".join(kept_lines).strip(), "\n".join(tts_lines).strip()
+        return "\n".join(kept_lines).strip(), "\n".join(tts_lines).strip(), saw_thinking
 
     def _normalize_line(self, line: str) -> str:
         without_emoji = EMOJI_RE.sub(" ", line)
-        return re.sub(r"\s+", " ", without_emoji).strip()
+        normalized = WHITESPACE_RE.sub(" ", without_emoji).strip()
+        # Strip leading "nanobot:" prefix that the TUI echoes in its own output,
+        # since the frontend already labels lines with the role name and TTS
+        # should not read the agent's own name aloud.
+        normalized = AGENT_OUTPUT_PREFIX_RE.sub("", normalized)
+        return normalized
 
     def _is_noisy_ui_line(self, line: str) -> bool:
         if SPINNER_ONLY_RE.fullmatch(line):
@@ -322,18 +381,47 @@ class NanobotTUIProcess:
         if BOX_DRAWING_ONLY_RE.fullmatch(line):
             return True
 
-        candidate = re.sub(r"^[^\w]+", "", line)
-        if THINKING_LINE_RE.match(candidate):
+        candidate = LEADING_NON_WORD_RE.sub("", line)
+        if THINKING_LINE_RE.search(candidate):
             return True
         if TOOL_STREAM_LINE_RE.match(candidate):
             return True
+        if USER_ECHO_LINE_RE.match(candidate):
+            return True
         return False
 
+    async def _wait_for_fd_readable(self) -> bool:
+        if self._master_fd is None:
+            return False
+
+        loop = asyncio.get_running_loop()
+        ready: asyncio.Future[None] = loop.create_future()
+
+        def _mark_ready() -> None:
+            if not ready.done():
+                ready.set_result(None)
+
+        try:
+            loop.add_reader(self._master_fd, _mark_ready)
+        except (AttributeError, NotImplementedError, OSError, ValueError):
+            await asyncio.sleep(0.01)
+            return True
+
+        try:
+            await ready
+            return True
+        finally:
+            with contextlib.suppress(Exception):
+                loop.remove_reader(self._master_fd)
+
     def _is_recent_duplicate(self, line: str) -> bool:
         now = time.monotonic()
         normalized = line.lower()
 
-        while self._recent_lines and (now - self._recent_lines[0][1]) > self._dedup_window_s:
+        while (
+            self._recent_lines
+            and (now - self._recent_lines[0][1]) > self._dedup_window_s
+        ):
             self._recent_lines.popleft()
 
         for previous, _timestamp in self._recent_lines:
@@ -359,11 +447,15 @@ class SuperTonicGateway:
     async def spawn_tui(self) -> None:
         async with self._lock:
             if self._tui and self._tui.running:
-                await self.bus.publish(WisperEvent(role="system", text="Nanobot TUI is already running."))
+                await self.bus.publish(
+                    WisperEvent(role="system", text="Nanobot TUI is already running.")
+                )
                 return
 
             command, workdir = _resolve_nanobot_command_and_workdir()
-            self._tui = NanobotTUIProcess(bus=self.bus, command=command, workdir=workdir)
+            self._tui = NanobotTUIProcess(
+                bus=self.bus, command=command, workdir=workdir
+            )
             await self._tui.start()
 
     async def send_user_message(self, text: str) -> None:
@@ -374,7 +466,10 @@ class SuperTonicGateway:
         async with self._lock:
             if not self._tui:
                 await self.bus.publish(
-                    WisperEvent(role="system", text="Nanobot TUI is not running. Click spawn first.")
+                    WisperEvent(
+                        role="system",
+                        text="Nanobot TUI is not running. Click spawn first.",
+                    )
                 )
                 return
             await self._tui.send(message)

voice_rtc.py

@@ -68,8 +68,12 @@ SPEECH_FILTER_RE = re.compile(
     r"^(spawned nanobot tui|stopped nanobot tui|nanobot tui exited|websocket)",
     re.IGNORECASE,
 )
-THINKING_STATUS_RE = re.compile(r"\bnanobot is thinking\b", re.IGNORECASE)
+THINKING_STATUS_RE = re.compile(
+    r"\b(?:agent|nanobot|napbot)\b(?:\s+is)?\s+thinking\b",
+    re.IGNORECASE,
+)
 USER_PREFIX_RE = re.compile(r"^(?:you|user)\s*:\s*", re.IGNORECASE)
+AGENT_PREFIX_RE = re.compile(r"^(?:nanobot|napbot)\b\s*[:>\-]?\s*", re.IGNORECASE)
 VOICE_TRANSCRIPT_RE = re.compile(
     r"^(?:wisper\s*:\s*)?voice\s+transcript\s*:\s*",
     re.IGNORECASE,
@@ -77,7 +81,6 @@ VOICE_TRANSCRIPT_RE = re.compile(
 ANSI_ESCAPE_RE = re.compile(r"\x1B(?:[@-Z\\-_]|\[[0-?]*[ -/]*[@-~])")
 CONTROL_CHAR_RE = re.compile(r"[\x00-\x1f\x7f]")
 BRAILLE_SPINNER_RE = re.compile(r"[\u2800-\u28ff]")
-SENTENCE_END_RE = re.compile(r"[.!?]\s*$")
 TTS_ALLOWED_ASCII = set(
     "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
     "abcdefghijklmnopqrstuvwxyz"
@@ -141,6 +144,15 @@ if AIORTC_AVAILABLE:
             self._closed = False
             self._frame_duration_s = frame_ms / 1000.0
             self._last_recv_at = 0.0
+            self._playing = False
+            self._idle_frames = 0
+            # Number of consecutive silent frames before signalling idle.
+            # At 20ms per frame, 15 frames = 300ms grace period to avoid
+            # flickering between TTS synthesis chunks.
+            self._idle_grace_frames = max(
+                1, int(os.getenv("HOST_RTC_IDLE_GRACE_MS", "300")) // max(1, frame_ms)
+            )
+            self._on_playing_changed: Callable[[bool], None] | None = None
 
         async def enqueue_pcm(
             self, pcm: bytes, sample_rate: int, channels: int = 1
@@ -211,8 +223,24 @@ if AIORTC_AVAILABLE:
 
             try:
                 payload = self._queue.get_nowait()
+                has_audio = True
             except asyncio.QueueEmpty:
                 payload = b"\x00" * self._bytes_per_frame
+                has_audio = False
+
+            # Notify when playback state changes.
+            if has_audio:
+                self._idle_frames = 0
+                if not self._playing:
+                    self._playing = True
+                    if self._on_playing_changed:
+                        self._on_playing_changed(True)
+            elif self._playing:
+                self._idle_frames += 1
+                if self._idle_frames >= self._idle_grace_frames:
+                    self._playing = False
+                    if self._on_playing_changed:
+                        self._on_playing_changed(False)
 
             self._last_recv_at = loop.time()
 
@@ -233,6 +261,8 @@ if AIORTC_AVAILABLE:
 else:
 
     class QueueAudioTrack:  # pragma: no cover - used only when aiortc is unavailable
+        _on_playing_changed: Callable[[bool], None] | None = None
+
         async def enqueue_pcm(
             self, pcm: bytes, sample_rate: int, channels: int = 1
         ) -> None:
@@ -343,6 +373,15 @@ class FasterWhisperSpeechToText:
             ).strip()
             or None
         )
+        self._repetition_penalty = float(
+            os.getenv("HOST_STT_REPETITION_PENALTY", "1.0")
+        )
+        raw_hallucination_threshold = os.getenv(
+            "HOST_STT_HALLUCINATION_SILENCE_THRESHOLD", ""
+        ).strip()
+        self._hallucination_silence_threshold: float | None = (
+            float(raw_hallucination_threshold) if raw_hallucination_threshold else None
+        )
 
         self._model: Any = None
         self._init_error: str | None = None
@@ -429,6 +468,8 @@ class FasterWhisperSpeechToText:
                 log_prob_threshold=self._log_prob_threshold,
                 no_speech_threshold=self._no_speech_threshold,
                 compression_ratio_threshold=self._compression_ratio_threshold,
+                repetition_penalty=self._repetition_penalty,
+                hallucination_silence_threshold=self._hallucination_silence_threshold,
             )
             transcript_parts: list[str] = []
             for segment in segments:
@@ -456,6 +497,8 @@ class FasterWhisperSpeechToText:
                 log_prob_threshold=self._log_prob_threshold,
                 no_speech_threshold=self._no_speech_threshold,
                 compression_ratio_threshold=self._compression_ratio_threshold,
+                repetition_penalty=self._repetition_penalty,
+                hallucination_silence_threshold=self._hallucination_silence_threshold,
             )
             transcript_parts: list[str] = []
             for segment in segments:
@@ -541,11 +584,11 @@ class SupertonicTextToSpeech:
             os.getenv("SUPERTONIC_MODEL", "supertonic-2").strip() or "supertonic-2"
         )
         self._voice_style_name = (
-            os.getenv("SUPERTONIC_VOICE_STYLE", "M1").strip() or "M1"
+            os.getenv("SUPERTONIC_VOICE_STYLE", "F1").strip() or "F1"
         )
         self._lang = os.getenv("SUPERTONIC_LANG", "en").strip() or "en"
-        self._total_steps = int(os.getenv("SUPERTONIC_TOTAL_STEPS", "5"))
-        self._speed = float(os.getenv("SUPERTONIC_SPEED", "1.05"))
+        self._total_steps = int(os.getenv("SUPERTONIC_TOTAL_STEPS", "8"))
+        self._speed = float(os.getenv("SUPERTONIC_SPEED", "1.5"))
         self._intra_op_num_threads = _optional_int_env("SUPERTONIC_INTRA_OP_THREADS")
         self._inter_op_num_threads = _optional_int_env("SUPERTONIC_INTER_OP_THREADS")
         self._auto_download = os.getenv(
@@ -605,7 +648,7 @@ class SupertonicTextToSpeech:
             message = str(exc)
             if "unsupported character" not in message.lower():
                 raise
-            fallback_text = self._sanitize_text_for_supertonic(text)
+            fallback_text = _sanitize_tts_text(text)
             if not fallback_text or fallback_text == text:
                 raise
             wav, _duration = self._engine.synthesize(
@@ -655,9 +698,6 @@ class SupertonicTextToSpeech:
             channels=max(1, channels),
         )
 
-    def _sanitize_text_for_supertonic(self, text: str) -> str:
-        return _sanitize_tts_text(text)
-
     def _initialize_blocking(self) -> None:
         if self._engine is not None and self._voice_style is not None:
             return
@@ -853,23 +893,15 @@ class WebRTCVoiceSession:
             maxsize=self._stt_segment_queue_size
         )
 
-        self._tts_buffer = ""
+        self._tts_chunks: list[str] = []
         self._tts_flush_handle: asyncio.TimerHandle | None = None
         self._tts_flush_lock = asyncio.Lock()
         self._tts_buffer_lock = asyncio.Lock()
-        self._tts_flush_delay_s = max(
-            0.08, float(os.getenv("HOST_TTS_FLUSH_DELAY_S", "0.45"))
+        # How long to wait after the last incoming chunk before flushing the
+        # entire accumulated response to TTS in one go.
+        self._tts_response_end_delay_s = max(
+            0.1, float(os.getenv("HOST_TTS_RESPONSE_END_DELAY_S", "1.5"))
         )
-        self._tts_sentence_flush_delay_s = max(
-            0.06,
-            float(os.getenv("HOST_TTS_SENTENCE_FLUSH_DELAY_S", "0.15")),
-        )
-        self._tts_min_chars = max(1, int(os.getenv("HOST_TTS_MIN_CHARS", "10")))
-        self._tts_max_wait_ms = max(300, int(os.getenv("HOST_TTS_MAX_WAIT_MS", "1800")))
-        self._tts_max_chunk_chars = max(
-            60, int(os.getenv("HOST_TTS_MAX_CHUNK_CHARS", "140"))
-        )
-        self._tts_buffer_started_at = 0.0
 
         self._closed = False
         self._stt_unavailable_notice_sent = False
@@ -887,14 +919,7 @@ class WebRTCVoiceSession:
                 )
             ),
         )
-        self._stt_max_ptt_ms = max(
-            self._stt_min_ptt_ms,
-            int(
-                os.getenv(
-                    "HOST_STT_MAX_PTT_MS", os.getenv("HOST_STT_MAX_SEGMENT_MS", "12000")
-                )
-            ),
-        )
+
         self._stt_suppress_during_tts = os.getenv(
             "HOST_STT_SUPPRESS_DURING_TTS", "1"
         ).strip() not in {
@@ -924,22 +949,16 @@ class WebRTCVoiceSession:
         normalized_chunk = chunk.strip()
         if not normalized_chunk:
             return
-        flush_delay = (
-            self._tts_sentence_flush_delay_s
-            if SENTENCE_END_RE.search(normalized_chunk)
-            else self._tts_flush_delay_s
-        )
-        loop = asyncio.get_running_loop()
         async with self._tts_buffer_lock:
             if not self._pc or not self._outbound_track:
                 return
-            if not self._tts_buffer.strip():
-                self._tts_buffer_started_at = loop.time()
-                self._tts_buffer = normalized_chunk
-            else:
-                # Keep line boundaries between streamed chunks so line-based filters stay accurate.
-                self._tts_buffer = f"{self._tts_buffer}\n{normalized_chunk}"
-            self._schedule_tts_flush_after(flush_delay)
+            # Keep line boundaries between streamed chunks so line-based filters
+            # stay accurate while avoiding repeated full-string copies.
+            self._tts_chunks.append(normalized_chunk)
+            # Reset the flush timer on every incoming chunk so the entire
+            # response is accumulated before synthesis begins.  The timer
+            # fires once no new chunks arrive for the configured delay.
+            self._schedule_tts_flush_after(self._tts_response_end_delay_s)
 
     async def handle_offer(self, payload: dict[str, Any]) -> None:
         if not AIORTC_AVAILABLE or not RTCPeerConnection or not RTCSessionDescription:
@@ -965,6 +984,7 @@ class WebRTCVoiceSession:
         peer_connection = RTCPeerConnection()
         self._pc = peer_connection
         self._outbound_track = QueueAudioTrack()
+        self._outbound_track._on_playing_changed = self._on_track_playing_changed
         peer_connection.addTrack(self._outbound_track)
 
         @peer_connection.on("connectionstatechange")
@@ -1096,8 +1116,7 @@ class WebRTCVoiceSession:
         if self._tts_flush_handle:
             self._tts_flush_handle.cancel()
             self._tts_flush_handle = None
-        self._tts_buffer = ""
-        self._tts_buffer_started_at = 0.0
+        self._tts_chunks.clear()
 
         if self._incoming_audio_task:
             self._incoming_audio_task.cancel()
@@ -1136,43 +1155,27 @@ class WebRTCVoiceSession:
         async with self._tts_flush_lock:
             async with self._tts_buffer_lock:
                 self._tts_flush_handle = None
-                raw_text = self._tts_buffer
-                buffer_started_at = self._tts_buffer_started_at
-                self._tts_buffer = ""
-                self._tts_buffer_started_at = 0.0
+                raw_text = "\n".join(self._tts_chunks)
+                self._tts_chunks.clear()
             clean_text = self._clean_tts_text(raw_text)
             if not clean_text:
                 return
 
-            loop = asyncio.get_running_loop()
-            now = loop.time()
-            buffer_age_ms = int(max(0.0, now - buffer_started_at) * 1000)
-            should_wait_for_more = (
-                len(clean_text) < self._tts_min_chars
-                and not SENTENCE_END_RE.search(clean_text)
-                and buffer_age_ms < self._tts_max_wait_ms
-            )
-            if should_wait_for_more:
-                async with self._tts_buffer_lock:
-                    if self._tts_buffer.strip():
-                        self._tts_buffer = f"{clean_text}\n{self._tts_buffer}".strip()
-                    else:
-                        self._tts_buffer = clean_text
-                    if self._tts_buffer_started_at <= 0.0:
-                        self._tts_buffer_started_at = (
-                            buffer_started_at if buffer_started_at > 0.0 else now
-                        )
-                    self._schedule_tts_flush_after(self._tts_flush_delay_s)
-                return
-
             if not self._outbound_track:
                 return
 
-            for part in self._chunk_tts_text(clean_text):
             try:
-                    audio = await self._tts.synthesize(part)
+                audio = await self._tts.synthesize(clean_text)
+            except asyncio.CancelledError:
+                raise
             except Exception as exc:
-                    await self._publish_system(f"Host TTS failed: {exc}")
+                import traceback  # noqa: local import in exception handler
+
+                traceback.print_exc()
+                # Restore the lost text so a future flush can retry it.
+                async with self._tts_buffer_lock:
+                    self._tts_chunks.insert(0, clean_text)
+                await self._publish_system(f"TTS synthesis error: {exc}")
                 return
 
             if not audio:
@@ -1221,7 +1224,6 @@ class WebRTCVoiceSession:
         resample_state = None
         recording = False
         recording_started_at = 0.0
-        recording_truncated = False
         segment_ms = 0.0
         segment_buffer = bytearray()
 
@@ -1253,7 +1255,6 @@ class WebRTCVoiceSession:
                 ):
                     recording = False
                     recording_started_at = 0.0
-                    recording_truncated = False
                     segment_ms = 0.0
                     segment_buffer = bytearray()
                     continue
@@ -1262,20 +1263,11 @@ class WebRTCVoiceSession:
                     if not recording:
                         recording = True
                         recording_started_at = asyncio.get_running_loop().time()
-                        recording_truncated = False
                         segment_ms = 0.0
                         segment_buffer = bytearray()
 
-                    if not recording_truncated:
-                        next_segment_ms = segment_ms + frame_ms
-                        if next_segment_ms <= self._stt_max_ptt_ms:
                     segment_buffer.extend(pcm16)
-                            segment_ms = next_segment_ms
-                        else:
-                            recording_truncated = True
-                            await self._publish_system(
-                                "PTT max length reached; extra audio will be ignored until release."
-                            )
+                    segment_ms += frame_ms
                     continue
 
                 if recording:
@@ -1291,7 +1283,6 @@ class WebRTCVoiceSession:
                     )
                     recording = False
                     recording_started_at = 0.0
-                    recording_truncated = False
                     segment_ms = 0.0
                     segment_buffer = bytearray()
         except asyncio.CancelledError:
@@ -1456,10 +1447,21 @@ class WebRTCVoiceSession:
     async def _publish_system(self, text: str) -> None:
         await self._gateway.bus.publish(WisperEvent(role="system", text=text))
 
+    async def _publish_agent_state(self, state: str) -> None:
+        await self._gateway.bus.publish(WisperEvent(role="agent-state", text=state))
+
+    def _on_track_playing_changed(self, playing: bool) -> None:
+        """Called from QueueAudioTrack.recv() when audio playback starts or stops."""
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            return
+        loop.create_task(self._publish_agent_state("speaking" if playing else "idle"))
+
     def _clean_tts_text(self, raw_text: str) -> str:
         lines = [line.strip() for line in raw_text.splitlines() if line.strip()]
         useful_lines = [
-            line
+            AGENT_PREFIX_RE.sub("", line)
             for line in lines
             if not SPEECH_FILTER_RE.match(line)
             and not THINKING_STATUS_RE.search(line)
@@ -1468,71 +1470,6 @@ class WebRTCVoiceSession:
         ]
         return _sanitize_tts_text(" ".join(useful_lines))
 
-    def _chunk_tts_text(self, text: str) -> list[str]:
-        clean_text = " ".join(text.split())
-        if not clean_text:
-            return []
-
-        max_chars = max(60, int(self._tts_max_chunk_chars))
-        if len(clean_text) <= max_chars:
-            return [clean_text]
-
-        sentence_parts = [
-            part.strip()
-            for part in re.split(r"(?<=[.!?])\s+", clean_text)
-            if part.strip()
-        ]
-        if not sentence_parts:
-            sentence_parts = [clean_text]
-
-        chunks: list[str] = []
-        for part in sentence_parts:
-            if len(part) <= max_chars:
-                chunks.append(part)
-            else:
-                chunks.extend(self._chunk_tts_words(part, max_chars))
-        return chunks
-
-    def _chunk_tts_words(self, text: str, max_chars: int) -> list[str]:
-        words = [word for word in text.split() if word]
-        if not words:
-            return []
-
-        chunks: list[str] = []
-        current_words: list[str] = []
-        current_len = 0
-
-        for word in words:
-            if len(word) > max_chars:
-                if current_words:
-                    chunks.append(" ".join(current_words))
-                    current_words = []
-                    current_len = 0
-                start = 0
-                while start < len(word):
-                    end = min(start + max_chars, len(word))
-                    piece = word[start:end]
-                    if len(piece) == max_chars:
-                        chunks.append(piece)
-                    else:
-                        current_words = [piece]
-                        current_len = len(piece)
-                    start = end
-                continue
-
-            extra = len(word) if not current_words else (1 + len(word))
-            if current_words and (current_len + extra) > max_chars:
-                chunks.append(" ".join(current_words))
-                current_words = [word]
-                current_len = len(word)
-            else:
-                current_words.append(word)
-                current_len += extra
-
-        if current_words:
-            chunks.append(" ".join(current_words))
-        return chunks
-
     def _frame_to_pcm16k_mono(
         self, frame: AudioFrame, resample_state: tuple[Any, ...] | None
     ) -> tuple[bytes, float, tuple[Any, ...] | None]:

wisper.py

@@ -1,8 +1,20 @@
 import asyncio
+import contextlib
+import os
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 
 
+def _positive_int_env(name: str, default: int) -> int:
+    raw_value = os.getenv(name, "").strip()
+    if not raw_value:
+        return default
+    try:
+        return max(1, int(raw_value))
+    except ValueError:
+        return default
+
+
 @dataclass(slots=True)
 class WisperEvent:
     role: str
@@ -19,19 +31,29 @@ class WisperBus:
     def __init__(self) -> None:
         self._subscribers: set[asyncio.Queue[WisperEvent]] = set()
         self._lock = asyncio.Lock()
+        self._subscriber_snapshot: tuple[asyncio.Queue[WisperEvent], ...] = ()
+        self._subscriber_queue_size = _positive_int_env(
+            "WISPER_SUBSCRIBER_QUEUE_SIZE", 512
+        )
 
     async def subscribe(self) -> asyncio.Queue[WisperEvent]:
-        queue: asyncio.Queue[WisperEvent] = asyncio.Queue()
+        queue: asyncio.Queue[WisperEvent] = asyncio.Queue(
+            maxsize=self._subscriber_queue_size
+        )
         async with self._lock:
             self._subscribers.add(queue)
+            self._subscriber_snapshot = tuple(self._subscribers)
         return queue
 
     async def unsubscribe(self, queue: asyncio.Queue[WisperEvent]) -> None:
         async with self._lock:
             self._subscribers.discard(queue)
+            self._subscriber_snapshot = tuple(self._subscribers)
 
     async def publish(self, event: WisperEvent) -> None:
-        async with self._lock:
-            subscribers = list(self._subscribers)
-        for queue in subscribers:
+        for queue in self._subscriber_snapshot:
+            if queue.full():
+                with contextlib.suppress(asyncio.QueueEmpty):
+                    queue.get_nowait()
+            with contextlib.suppress(asyncio.QueueFull):
                 queue.put_nowait(event)