nanobot-voice-interface/voice_rtc.py

import asyncio
import audioop
import contextlib
import io
import os
import re
import shlex
import shutil
import subprocess
import tempfile
import wave
from dataclasses import dataclass
from fractions import Fraction
from typing import TYPE_CHECKING, Any, Awaitable, Callable

from wisper import WisperEvent

if TYPE_CHECKING:
    from supertonic_gateway import SuperTonicGateway


try:
    import numpy as np

    NUMPY_AVAILABLE = True
except Exception:  # pragma: no cover - runtime fallback when numpy is unavailable
    np = None  # type: ignore[assignment]
    NUMPY_AVAILABLE = False


try:
    from supertonic import TTS as SupertonicTTS

    SUPERTONIC_TTS_AVAILABLE = True
except Exception:  # pragma: no cover - runtime fallback when supertonic is unavailable
    SupertonicTTS = None  # type: ignore[assignment]
    SUPERTONIC_TTS_AVAILABLE = False


try:
    from faster_whisper import WhisperModel

    FASTER_WHISPER_AVAILABLE = True
except (
    Exception
):  # pragma: no cover - runtime fallback when faster-whisper is unavailable
    WhisperModel = None  # type: ignore[assignment]
    FASTER_WHISPER_AVAILABLE = False


try:
    from aiortc import RTCPeerConnection, RTCSessionDescription
    from aiortc.mediastreams import MediaStreamTrack
    from aiortc.sdp import candidate_from_sdp
    from av import AudioFrame

    AIORTC_AVAILABLE = True
except Exception:  # pragma: no cover - runtime fallback when aiortc is unavailable
    RTCPeerConnection = None  # type: ignore[assignment]
    RTCSessionDescription = None  # type: ignore[assignment]
    MediaStreamTrack = object  # type: ignore[assignment,misc]
    candidate_from_sdp = None  # type: ignore[assignment]
    AudioFrame = None  # type: ignore[assignment]
    AIORTC_AVAILABLE = False


SPEECH_FILTER_RE = re.compile(
    r"^(spawned nanobot tui|stopped nanobot tui|nanobot tui exited|websocket)",
    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,
)
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]")
TTS_ALLOWED_ASCII = set(
    "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
    "abcdefghijklmnopqrstuvwxyz"
    "0123456789"
    " .,!?;:'\"()[]{}@#%&*+-_/<>|"
)


def _sanitize_tts_text(text: str) -> str:
    cleaned = ANSI_ESCAPE_RE.sub(" ", text)
    cleaned = BRAILLE_SPINNER_RE.sub(" ", cleaned)
    cleaned = cleaned.replace("\u00a0", " ")
    cleaned = cleaned.replace("•", " ")
    cleaned = CONTROL_CHAR_RE.sub(" ", cleaned)
    cleaned = "".join(
        ch if (ch in TTS_ALLOWED_ASCII or ch.isspace()) else " " for ch in cleaned
    )
    cleaned = re.sub(r"\s+", " ", cleaned).strip()
    return cleaned


def _optional_int_env(name: str) -> int | None:
    raw_value = os.getenv(name, "").strip()
    if not raw_value:
        return None
    return int(raw_value)


@dataclass(slots=True)
class PCMChunk:
    pcm: bytes
    sample_rate: int
    channels: int = 1


if AIORTC_AVAILABLE:

    class QueueAudioTrack(MediaStreamTrack):
        kind = "audio"

        def __init__(self, sample_rate: int = 48_000, frame_ms: int = 20) -> None:
            super().__init__()
            self._sample_rate = sample_rate
            self._frame_ms = max(1, frame_ms)
            self._samples_per_frame = max(1, (sample_rate * frame_ms) // 1000)
            self._bytes_per_frame = self._samples_per_frame * 2
            self._queue: asyncio.Queue[bytes] = asyncio.Queue()
            self._timestamp = 0
            self._resample_state = None
            self._resample_source_rate: int | None = None
            self._lead_in_ms = max(
                0, int(os.getenv("HOST_RTC_OUTBOUND_LEAD_IN_MS", "120"))
            )
            self._lead_in_frames = (
                self._lead_in_ms + self._frame_ms - 1
            ) // self._frame_ms
            self._lead_in_idle_s = max(
                0.1, float(os.getenv("HOST_RTC_OUTBOUND_IDLE_S", "0.6"))
            )
            self._last_enqueue_at = 0.0
            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
        ) -> None:
            if self._closed or not pcm:
                return

            now = asyncio.get_running_loop().time()
            should_add_lead_in = (
                self._lead_in_frames > 0
                and self._queue.empty()
                and (
                    self._last_enqueue_at <= 0.0
                    or (now - self._last_enqueue_at) >= self._lead_in_idle_s
                )
            )
            if should_add_lead_in:
                silence = b"\x00" * self._bytes_per_frame
                for _index in range(self._lead_in_frames):
                    await self._queue.put(silence)

            mono = pcm
            if channels > 1:
                mono = audioop.tomono(mono, 2, 0.5, 0.5)

            if sample_rate != self._sample_rate:
                # audioop rate conversion state is only valid when source/destination rates stay the same.
                if self._resample_source_rate != sample_rate:
                    self._resample_state = None
                    self._resample_source_rate = sample_rate
                mono, self._resample_state = audioop.ratecv(
                    mono,
                    2,
                    1,
                    sample_rate,
                    self._sample_rate,
                    self._resample_state,
                )
            else:
                self._resample_state = None
                self._resample_source_rate = None

            if not mono:
                return

            for start in range(0, len(mono), self._bytes_per_frame):
                chunk = mono[start : start + self._bytes_per_frame]
                if len(chunk) < self._bytes_per_frame:
                    chunk += b"\x00" * (self._bytes_per_frame - len(chunk))
                await self._queue.put(chunk)
            self._last_enqueue_at = now

        async def recv(self) -> AudioFrame:
            if self._closed:
                raise asyncio.CancelledError

            # Pace frame delivery to real-time to prevent RTP burst sends.
            # Without pacing, when TTS enqueues audio faster than real-time,
            # aiortc sends RTP packets in a burst and the browser's jitter
            # buffer skips ahead, causing the user to only hear the tail end.
            loop = asyncio.get_running_loop()
            now = loop.time()
            if self._last_recv_at > 0.0:
                elapsed = now - self._last_recv_at
                remaining = self._frame_duration_s - elapsed
                if remaining > 0.001:
                    await asyncio.sleep(remaining)

            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()

            frame = AudioFrame(
                format="s16", layout="mono", samples=self._samples_per_frame
            )
            frame.planes[0].update(payload)
            frame.sample_rate = self._sample_rate
            frame.time_base = Fraction(1, self._sample_rate)
            frame.pts = self._timestamp
            self._timestamp += self._samples_per_frame
            return frame

        def stop(self) -> None:
            self._closed = True
            super().stop()

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:
            return

        def stop(self) -> None:
            return


def _write_temp_wav(pcm: bytes, sample_rate: int, channels: int) -> str:
    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_file:
        tmp_path = tmp_file.name
    with wave.open(tmp_path, "wb") as wav_file:
        wav_file.setnchannels(max(1, channels))
        wav_file.setsampwidth(2)
        wav_file.setframerate(sample_rate)
        wav_file.writeframes(pcm)
    return tmp_path


class CommandSpeechToText:
    def __init__(self) -> None:
        self._command_template = os.getenv("HOST_STT_COMMAND", "").strip()

    @property
    def enabled(self) -> bool:
        return bool(self._command_template)

    async def transcribe_pcm(
        self, pcm: bytes, sample_rate: int = 16_000, channels: int = 1
    ) -> str | None:
        if not self.enabled or not pcm:
            return None
        return await asyncio.to_thread(
            self._transcribe_blocking, pcm, sample_rate, channels
        )

    def unavailable_reason(self) -> str:
        if not self._command_template:
            return "HOST_STT_COMMAND is not configured."
        return "HOST_STT_COMMAND failed to produce transcript."

    def _transcribe_blocking(
        self, pcm: bytes, sample_rate: int, channels: int
    ) -> str | None:
        tmp_path: str | None = None
        try:
            tmp_path = _write_temp_wav(
                pcm=pcm, sample_rate=sample_rate, channels=channels
            )

            command = self._command_template
            if "{input_wav}" in command:
                command = command.replace("{input_wav}", shlex.quote(tmp_path))
            else:
                command = f"{command} {shlex.quote(tmp_path)}"

            result = subprocess.run(
                command,
                shell=True,
                capture_output=True,
                text=True,
                check=False,
            )
            if result.returncode != 0:
                stderr = result.stderr.strip() or "unknown error"
                raise RuntimeError(f"STT command failed: {stderr}")

            transcript = result.stdout.strip()
            return transcript or None
        finally:
            if tmp_path and os.path.exists(tmp_path):
                with contextlib.suppress(OSError):
                    os.unlink(tmp_path)


class FasterWhisperSpeechToText:
    def __init__(self) -> None:
        self._model_name = os.getenv("HOST_STT_MODEL", "base.en").strip() or "base.en"
        self._device = os.getenv("HOST_STT_DEVICE", "auto").strip() or "auto"
        self._compute_type = (
            os.getenv("HOST_STT_COMPUTE_TYPE", "int8").strip() or "int8"
        )
        self._language = os.getenv("HOST_STT_LANGUAGE", "en").strip()
        self._beam_size = max(1, int(os.getenv("HOST_STT_BEAM_SIZE", "2")))
        self._best_of = max(1, int(os.getenv("HOST_STT_BEST_OF", "2")))
        self._vad_filter = os.getenv("HOST_STT_VAD_FILTER", "0").strip() not in {
            "0",
            "false",
            "False",
            "no",
            "off",
        }
        self._temperature = float(os.getenv("HOST_STT_TEMPERATURE", "0.0"))
        self._log_prob_threshold = float(
            os.getenv("HOST_STT_LOG_PROB_THRESHOLD", "-1.0")
        )
        self._no_speech_threshold = float(
            os.getenv("HOST_STT_NO_SPEECH_THRESHOLD", "0.6")
        )
        self._compression_ratio_threshold = float(
            os.getenv("HOST_STT_COMPRESSION_RATIO_THRESHOLD", "2.4")
        )
        self._initial_prompt = (
            os.getenv(
                "HOST_STT_INITIAL_PROMPT",
                "Transcribe brief spoken English precisely. Prefer common words over sound effects.",
            ).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
        self._lock = asyncio.Lock()

    @property
    def enabled(self) -> bool:
        return FASTER_WHISPER_AVAILABLE and WhisperModel is not None

    @property
    def init_error(self) -> str | None:
        return self._init_error

    async def transcribe_pcm(
        self, pcm: bytes, sample_rate: int = 16_000, channels: int = 1
    ) -> str | None:
        if not self.enabled or not pcm:
            return None
        async with self._lock:
            return await asyncio.to_thread(
                self._transcribe_blocking, pcm, sample_rate, channels
            )

    async def warmup(self) -> None:
        if not self.enabled:
            return
        async with self._lock:
            await asyncio.to_thread(self._initialize_blocking)

    def _initialize_blocking(self) -> None:
        if self._model is not None:
            return
        if not self.enabled or WhisperModel is None:
            return

        try:
            self._model = WhisperModel(
                self._model_name,
                device=self._device,
                compute_type=self._compute_type,
            )
            self._init_error = None
        except Exception as exc:
            self._init_error = str(exc)
            self._model = None

    def _transcribe_blocking(
        self, pcm: bytes, sample_rate: int, channels: int
    ) -> str | None:
        self._initialize_blocking()
        if self._model is None:
            if self._init_error:
                raise RuntimeError(
                    f"faster-whisper initialization failed: {self._init_error}"
                )
            return None

        if NUMPY_AVAILABLE and np is not None:
            mono = pcm
            if channels > 1:
                mono = audioop.tomono(mono, 2, 0.5, 0.5)
            if sample_rate != 16_000:
                mono, _ = audioop.ratecv(
                    mono,
                    2,
                    1,
                    sample_rate,
                    16_000,
                    None,
                )
            audio = np.frombuffer(mono, dtype=np.int16).astype(np.float32) / 32768.0
            if audio.size == 0:
                return None
            segments, _info = self._model.transcribe(
                audio,
                language=self._language or None,
                beam_size=self._beam_size,
                best_of=self._best_of,
                vad_filter=self._vad_filter,
                condition_on_previous_text=False,
                without_timestamps=True,
                initial_prompt=self._initial_prompt,
                temperature=self._temperature,
                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:
                text = str(getattr(segment, "text", "")).strip()
                if text:
                    transcript_parts.append(text)
            transcript = " ".join(transcript_parts).strip()
            return transcript or None

        tmp_path: str | None = None
        try:
            tmp_path = _write_temp_wav(
                pcm=pcm, sample_rate=sample_rate, channels=channels
            )
            segments, _info = self._model.transcribe(
                tmp_path,
                language=self._language or None,
                beam_size=self._beam_size,
                best_of=self._best_of,
                vad_filter=self._vad_filter,
                condition_on_previous_text=False,
                without_timestamps=True,
                initial_prompt=self._initial_prompt,
                temperature=self._temperature,
                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:
                text = str(getattr(segment, "text", "")).strip()
                if text:
                    transcript_parts.append(text)
            transcript = " ".join(transcript_parts).strip()
            return transcript or None
        finally:
            if tmp_path and os.path.exists(tmp_path):
                with contextlib.suppress(OSError):
                    os.unlink(tmp_path)


class HostSpeechToText:
    def __init__(self) -> None:
        provider = (
            os.getenv("HOST_STT_PROVIDER", "faster-whisper").strip() or "faster-whisper"
        ).lower()
        if provider not in {"faster-whisper", "command", "auto"}:
            provider = "auto"
        self._provider = provider
        self._faster_whisper = FasterWhisperSpeechToText()
        self._command = CommandSpeechToText()

    @property
    def enabled(self) -> bool:
        if self._provider == "faster-whisper":
            return self._faster_whisper.enabled
        if self._provider == "command":
            return self._command.enabled
        return self._faster_whisper.enabled or self._command.enabled

    async def transcribe_pcm(
        self, pcm: bytes, sample_rate: int = 16_000, channels: int = 1
    ) -> str | None:
        if self._provider in {"faster-whisper", "auto"}:
            transcript = await self._faster_whisper.transcribe_pcm(
                pcm=pcm,
                sample_rate=sample_rate,
                channels=channels,
            )
            if transcript:
                return transcript
            if self._provider == "faster-whisper":
                return None

        if self._provider in {"command", "auto"}:
            return await self._command.transcribe_pcm(
                pcm=pcm,
                sample_rate=sample_rate,
                channels=channels,
            )

        return None

    async def warmup(self) -> None:
        if self._provider in {"faster-whisper", "auto"}:
            await self._faster_whisper.warmup()

    def unavailable_reason(self) -> str:
        if self._provider == "faster-whisper":
            if not self._faster_whisper.enabled:
                return "faster-whisper package is not available."
            if self._faster_whisper.init_error:
                return f"faster-whisper initialization failed: {self._faster_whisper.init_error}"
            return "faster-whisper did not return transcript."
        if self._provider == "command":
            return self._command.unavailable_reason()

        if self._faster_whisper.init_error:
            return f"faster-whisper initialization failed: {self._faster_whisper.init_error}"
        if self._command.enabled:
            return "HOST_STT_COMMAND failed to produce transcript."
        if not self._faster_whisper.enabled:
            return "faster-whisper package is not available."
        return "No STT provider is configured."


class SupertonicTextToSpeech:
    def __init__(self) -> None:
        self._model = (
            os.getenv("SUPERTONIC_MODEL", "supertonic-2").strip() or "supertonic-2"
        )
        self._voice_style_name = (
            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", "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(
            "SUPERTONIC_AUTO_DOWNLOAD", "1"
        ).strip() not in {
            "0",
            "false",
            "False",
            "no",
            "off",
        }

        self._engine: Any = None
        self._voice_style: Any = None
        self._init_error: str | None = None
        self._lock = asyncio.Lock()

    @property
    def enabled(self) -> bool:
        return (
            SUPERTONIC_TTS_AVAILABLE and SupertonicTTS is not None and NUMPY_AVAILABLE
        )

    @property
    def init_error(self) -> str | None:
        return self._init_error

    async def synthesize(self, text: str) -> PCMChunk | None:
        if not self.enabled:
            return None

        clean_text = " ".join(text.split())
        if not clean_text:
            return None

        async with self._lock:
            return await asyncio.to_thread(self._synthesize_blocking, clean_text)

    def _synthesize_blocking(self, text: str) -> PCMChunk | None:
        self._initialize_blocking()
        if self._engine is None or self._voice_style is None or np is None:
            return None

        text = _sanitize_tts_text(text)
        if not text:
            return None

        try:
            wav, _duration = self._engine.synthesize(
                text,
                voice_style=self._voice_style,
                lang=self._lang,
                total_steps=self._total_steps,
                speed=self._speed,
            )
        except ValueError as exc:
            message = str(exc)
            if "unsupported character" not in message.lower():
                raise
            fallback_text = _sanitize_tts_text(text)
            if not fallback_text or fallback_text == text:
                raise
            wav, _duration = self._engine.synthesize(
                fallback_text,
                voice_style=self._voice_style,
                lang=self._lang,
                total_steps=self._total_steps,
                speed=self._speed,
            )

        samples = np.asarray(wav)
        if samples.size == 0:
            return None

        channels = 1
        if samples.ndim == 0:
            samples = samples.reshape(1)
        elif samples.ndim == 1:
            channels = 1
        elif samples.ndim == 2:
            # Normalize to frames x channels so PCM bytes are correctly interleaved.
            dim0, dim1 = int(samples.shape[0]), int(samples.shape[1])
            if dim0 <= 2 and dim1 > dim0:
                channels = dim0
                samples = samples.T
            elif dim1 <= 2 and dim0 > dim1:
                channels = dim1
            else:
                channels = 1
                samples = samples.reshape(-1)
        else:
            channels = 1
            samples = samples.reshape(-1)

        if np.issubdtype(samples.dtype, np.floating):
            samples = np.clip(samples, -1.0, 1.0)
            samples = (samples * 32767.0).astype(np.int16)
        else:
            if samples.dtype != np.int16:
                samples = samples.astype(np.int16)

        pcm = samples.tobytes()

        return PCMChunk(
            pcm=pcm,
            sample_rate=int(getattr(self._engine, "sample_rate", 24_000)),
            channels=max(1, channels),
        )

    def _initialize_blocking(self) -> None:
        if self._engine is not None and self._voice_style is not None:
            return
        if not self.enabled or SupertonicTTS is None:
            return

        try:
            engine = SupertonicTTS(
                model=self._model,
                auto_download=self._auto_download,
                intra_op_num_threads=self._intra_op_num_threads,
                inter_op_num_threads=self._inter_op_num_threads,
            )
            voice_style = engine.get_voice_style(self._voice_style_name)
        except Exception as exc:
            self._init_error = str(exc)
            return

        self._engine = engine
        self._voice_style = voice_style
        self._init_error = None


class HostTextToSpeech:
    def __init__(self) -> None:
        provider = (
            os.getenv("HOST_TTS_PROVIDER", "supertonic").strip() or "supertonic"
        ).lower()
        if provider not in {"supertonic", "command", "espeak", "auto"}:
            provider = "auto"
        self._provider = provider
        self._supertonic = SupertonicTextToSpeech()
        self._command_template = os.getenv("HOST_TTS_COMMAND", "").strip()
        self._espeak = shutil.which("espeak")

    @property
    def enabled(self) -> bool:
        if self._provider == "supertonic":
            return self._supertonic.enabled
        if self._provider == "command":
            return bool(self._command_template)
        if self._provider == "espeak":
            return bool(self._espeak)
        return self._supertonic.enabled or bool(self._command_template or self._espeak)

    async def synthesize(self, text: str) -> PCMChunk | None:
        clean_text = " ".join(text.split())
        if not clean_text:
            return None

        if self._provider in {"supertonic", "auto"}:
            audio = await self._supertonic.synthesize(clean_text)
            if audio:
                return audio
            if self._provider == "supertonic":
                return None

        if self._provider in {"command", "auto"} and self._command_template:
            return await asyncio.to_thread(self._synthesize_with_command, clean_text)
        if self._provider == "command":
            return None

        if self._provider in {"espeak", "auto"} and self._espeak:
            return await asyncio.to_thread(self._synthesize_with_espeak, clean_text)

        return None

    def unavailable_reason(self) -> str:
        if self._provider == "supertonic":
            if not self._supertonic.enabled:
                return "supertonic package is not available."
            if self._supertonic.init_error:
                return (
                    f"supertonic initialization failed: {self._supertonic.init_error}"
                )
            return "supertonic did not return audio."
        if self._provider == "command":
            return "HOST_TTS_COMMAND is not configured."
        if self._provider == "espeak":
            return "espeak binary is not available."

        if self._supertonic.init_error:
            return f"supertonic initialization failed: {self._supertonic.init_error}"
        if self._command_template:
            return "HOST_TTS_COMMAND failed to produce audio."
        if self._espeak:
            return "espeak failed to produce audio."
        return "No TTS provider is configured."

    def _synthesize_with_command(self, text: str) -> PCMChunk | None:
        command = self._command_template
        if "{text}" in command:
            command = command.replace("{text}", shlex.quote(text))
        else:
            command = f"{command} {shlex.quote(text)}"

        if "{output_wav}" in command:
            tmp_path: str | None = None
            try:
                with tempfile.NamedTemporaryFile(
                    suffix=".wav", delete=False
                ) as tmp_file:
                    tmp_path = tmp_file.name
                command_with_output = command.replace(
                    "{output_wav}", shlex.quote(tmp_path)
                )
                result = subprocess.run(
                    command_with_output,
                    shell=True,
                    capture_output=True,
                    text=True,
                    check=False,
                )
                if result.returncode != 0:
                    stderr = result.stderr.strip() or "unknown error"
                    raise RuntimeError(f"TTS command failed: {stderr}")
                return self._read_wav_file(tmp_path)
            finally:
                if tmp_path and os.path.exists(tmp_path):
                    with contextlib.suppress(OSError):
                        os.unlink(tmp_path)

        result = subprocess.run(
            command,
            shell=True,
            capture_output=True,
            check=False,
        )
        if result.returncode != 0:
            stderr = result.stderr.decode(errors="ignore").strip() or "unknown error"
            raise RuntimeError(f"TTS command failed: {stderr}")
        return self._decode_wav_bytes(result.stdout)

    def _synthesize_with_espeak(self, text: str) -> PCMChunk | None:
        if not self._espeak:
            return None

        result = subprocess.run(
            [self._espeak, "--stdout", text],
            capture_output=True,
            check=False,
        )
        if result.returncode != 0:
            stderr = result.stderr.decode(errors="ignore").strip() or "unknown error"
            raise RuntimeError(f"espeak failed: {stderr}")
        return self._decode_wav_bytes(result.stdout)

    def _read_wav_file(self, path: str) -> PCMChunk | None:
        try:
            with open(path, "rb") as wav_file:
                return self._decode_wav_bytes(wav_file.read())
        except OSError:
            return None

    def _decode_wav_bytes(self, payload: bytes) -> PCMChunk | None:
        if not payload:
            return None

        with wave.open(io.BytesIO(payload), "rb") as wav_file:
            channels = wav_file.getnchannels()
            sample_width = wav_file.getsampwidth()
            sample_rate = wav_file.getframerate()
            pcm = wav_file.readframes(wav_file.getnframes())

        if sample_width != 2:
            pcm = audioop.lin2lin(pcm, sample_width, 2)

        return PCMChunk(pcm=pcm, sample_rate=sample_rate, channels=max(1, channels))


SendJsonCallable = Callable[[dict[str, Any]], Awaitable[None]]


class WebRTCVoiceSession:
    def __init__(
        self, gateway: "SuperTonicGateway", send_json: SendJsonCallable
    ) -> None:
        self._gateway = gateway
        self._send_json = send_json

        self._pc: RTCPeerConnection | None = None
        self._outbound_track: QueueAudioTrack | None = None
        self._incoming_audio_task: asyncio.Task[None] | None = None
        self._stt_worker_task: asyncio.Task[None] | None = None
        self._stt_warmup_task: asyncio.Task[None] | None = None

        self._stt = HostSpeechToText()
        self._tts = HostTextToSpeech()
        self._stt_segment_queue_size = max(
            1, int(os.getenv("HOST_STT_SEGMENT_QUEUE_SIZE", "2"))
        )
        self._stt_segments: asyncio.Queue[bytes] = asyncio.Queue(
            maxsize=self._stt_segment_queue_size
        )

        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()
        # 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._closed = False
        self._stt_unavailable_notice_sent = False
        self._tts_unavailable_notice_sent = False
        self._audio_seen_notice_sent = False
        self._audio_format_notice_sent = False
        self._stt_first_segment_notice_sent = False
        self._ptt_timing_correction_notice_sent = False

        self._stt_min_ptt_ms = max(
            120,
            int(
                os.getenv(
                    "HOST_STT_MIN_PTT_MS", os.getenv("HOST_STT_MIN_SEGMENT_MS", "220")
                )
            ),
        )

        self._stt_suppress_during_tts = os.getenv(
            "HOST_STT_SUPPRESS_DURING_TTS", "1"
        ).strip() not in {
            "0",
            "false",
            "False",
            "no",
            "off",
        }
        self._stt_suppress_ms_after_tts = max(
            0,
            int(os.getenv("HOST_STT_SUPPRESS_MS_AFTER_TTS", "300")),
        )
        self._stt_suppress_until = 0.0
        self._stt_backlog_notice_interval_s = max(
            2.0,
            float(os.getenv("HOST_STT_BACKLOG_NOTICE_INTERVAL_S", "6.0")),
        )
        self._last_stt_backlog_notice_at = 0.0
        self._pending_ice_candidates: list[dict[str, Any] | None] = []
        self._ptt_pressed = False

    def set_push_to_talk_pressed(self, pressed: bool) -> None:
        self._ptt_pressed = bool(pressed)

    async def queue_output_text(self, chunk: str) -> None:
        normalized_chunk = chunk.strip()
        if not normalized_chunk:
            return
        async with self._tts_buffer_lock:
            if not self._pc or not self._outbound_track:
                return
            # 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:
            await self._send_json(
                {
                    "type": "rtc-error",
                    "message": "WebRTC backend unavailable on host (aiortc is not installed).",
                }
            )
            return

        sdp = str(payload.get("sdp", "")).strip()
        rtc_type = str(payload.get("rtcType", "offer")).strip() or "offer"
        if not sdp:
            await self._send_json(
                {"type": "rtc-error", "message": "Missing SDP offer payload."}
            )
            return

        await self._close_peer_connection()
        self._ptt_pressed = False

        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")
        def on_connectionstatechange() -> None:
            asyncio.create_task(
                self._send_json(
                    {
                        "type": "rtc-state",
                        "state": peer_connection.connectionState,
                    }
                )
            )

        @peer_connection.on("track")
        def on_track(track: MediaStreamTrack) -> None:
            if track.kind != "audio":
                return
            if self._incoming_audio_task:
                self._incoming_audio_task.cancel()
            self._incoming_audio_task = asyncio.create_task(
                self._consume_audio_track(track),
                name="voice-inbound-track",
            )

        await peer_connection.setRemoteDescription(
            RTCSessionDescription(sdp=sdp, type=rtc_type)
        )
        await self._drain_pending_ice_candidates(peer_connection)
        answer = await peer_connection.createAnswer()
        await peer_connection.setLocalDescription(answer)
        await self._wait_for_ice_gathering(peer_connection)

        local_description = peer_connection.localDescription
        sdp_answer = str(local_description.sdp or "")
        if sdp_answer:
            sdp_answer = (
                sdp_answer.replace("\r\n", "\n")
                .replace("\r", "\n")
                .strip()
                .replace("\n", "\r\n")
                + "\r\n"
            )
        await self._send_json(
            {
                "type": "rtc-answer",
                "sdp": sdp_answer,
                "rtcType": local_description.type,
            }
        )

        if self._stt.enabled and not self._stt_worker_task:
            self._stt_worker_task = asyncio.create_task(
                self._stt_worker(), name="voice-stt-worker"
            )
        if self._stt.enabled and (
            self._stt_warmup_task is None or self._stt_warmup_task.done()
        ):
            self._stt_warmup_task = asyncio.create_task(
                self._warmup_stt(), name="voice-stt-warmup"
            )
        elif not self._stt.enabled and not self._stt_unavailable_notice_sent:
            self._stt_unavailable_notice_sent = True
            await self._publish_system(
                f"Voice input backend unavailable. {self._stt.unavailable_reason()}"
            )

    async def handle_ice_candidate(self, payload: dict[str, Any]) -> None:
        if not AIORTC_AVAILABLE:
            return

        raw_candidate = payload.get("candidate")
        candidate_payload: dict[str, Any] | None
        if raw_candidate in (None, ""):
            candidate_payload = None
        elif isinstance(raw_candidate, dict):
            candidate_payload = raw_candidate
        else:
            return

        if not self._pc or self._pc.remoteDescription is None:
            self._pending_ice_candidates.append(candidate_payload)
            return

        await self._apply_ice_candidate(self._pc, candidate_payload)

    async def _drain_pending_ice_candidates(
        self,
        peer_connection: RTCPeerConnection,
    ) -> None:
        if not self._pending_ice_candidates:
            return

        pending = list(self._pending_ice_candidates)
        self._pending_ice_candidates.clear()
        for pending_candidate in pending:
            await self._apply_ice_candidate(peer_connection, pending_candidate)

    async def _apply_ice_candidate(
        self,
        peer_connection: RTCPeerConnection,
        raw_candidate: dict[str, Any] | None,
    ) -> None:
        if raw_candidate is None:
            with contextlib.suppress(Exception):
                await peer_connection.addIceCandidate(None)
            return

        candidate_sdp = str(raw_candidate.get("candidate", "")).strip()
        if not candidate_sdp or not candidate_from_sdp:
            return

        if candidate_sdp.startswith("candidate:"):
            candidate_sdp = candidate_sdp[len("candidate:") :]

        try:
            candidate = candidate_from_sdp(candidate_sdp)
            candidate.sdpMid = raw_candidate.get("sdpMid")
            line_index = raw_candidate.get("sdpMLineIndex")
            candidate.sdpMLineIndex = (
                int(line_index) if line_index is not None else None
            )
            await peer_connection.addIceCandidate(candidate)
        except Exception as exc:
            await self._publish_system(f"Failed to add ICE candidate: {exc}")

    async def close(self) -> None:
        self._closed = True
        self._ptt_pressed = False
        if self._tts_flush_handle:
            self._tts_flush_handle.cancel()
            self._tts_flush_handle = None
        self._tts_chunks.clear()

        if self._incoming_audio_task:
            self._incoming_audio_task.cancel()
            with contextlib.suppress(asyncio.CancelledError):
                await self._incoming_audio_task
            self._incoming_audio_task = None

        if self._stt_worker_task:
            self._stt_worker_task.cancel()
            with contextlib.suppress(asyncio.CancelledError):
                await self._stt_worker_task
            self._stt_worker_task = None

        if self._stt_warmup_task:
            self._stt_warmup_task.cancel()
            with contextlib.suppress(asyncio.CancelledError):
                await self._stt_warmup_task
            self._stt_warmup_task = None

        await self._close_peer_connection()

    def _schedule_tts_flush(self) -> None:
        if self._closed:
            return
        asyncio.create_task(self._flush_tts(), name="voice-tts-flush")

    def _schedule_tts_flush_after(self, delay_s: float) -> None:
        if self._tts_flush_handle:
            self._tts_flush_handle.cancel()
        loop = asyncio.get_running_loop()
        self._tts_flush_handle = loop.call_later(
            max(0.05, delay_s), self._schedule_tts_flush
        )

    async def _flush_tts(self) -> None:
        async with self._tts_flush_lock:
            async with self._tts_buffer_lock:
                self._tts_flush_handle = None
                raw_text = "\n".join(self._tts_chunks)
                self._tts_chunks.clear()
            clean_text = self._clean_tts_text(raw_text)
            if not clean_text:
                return

            if not self._outbound_track:
                return

            try:
                audio = await self._tts.synthesize(clean_text)
            except asyncio.CancelledError:
                raise
            except Exception as 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:
                if not self._tts_unavailable_notice_sent:
                    self._tts_unavailable_notice_sent = True
                    await self._publish_system(
                        f"Host TTS backend is unavailable. {self._tts.unavailable_reason()}"
                    )
                return

            if not self._outbound_track:
                return
            self._extend_stt_suppression(audio)
            await self._outbound_track.enqueue_pcm(
                pcm=audio.pcm,
                sample_rate=audio.sample_rate,
                channels=audio.channels,
            )

    def _extend_stt_suppression(self, audio: PCMChunk) -> None:
        if not self._stt_suppress_during_tts:
            return

        channels = max(1, int(audio.channels))
        sample_rate = max(1, int(audio.sample_rate))
        sample_count = len(audio.pcm) // (2 * channels)
        if sample_count <= 0:
            return

        duration_s = sample_count / float(sample_rate)
        cooldown_s = float(self._stt_suppress_ms_after_tts) / 1000.0
        now = asyncio.get_running_loop().time()
        base = max(now, self._stt_suppress_until)
        self._stt_suppress_until = base + duration_s + cooldown_s

    async def _consume_audio_track(self, track: MediaStreamTrack) -> None:
        if not self._stt.enabled:
            try:
                while True:
                    await track.recv()
            except asyncio.CancelledError:
                raise
            except Exception:
                return

        resample_state = None
        recording = False
        recording_started_at = 0.0
        segment_ms = 0.0
        segment_buffer = bytearray()

        try:
            while True:
                frame = await track.recv()
                pcm16, frame_ms, resample_state = self._frame_to_pcm16k_mono(
                    frame, resample_state
                )
                if not pcm16:
                    continue

                if not self._audio_seen_notice_sent:
                    self._audio_seen_notice_sent = True
                    await self._publish_system("Receiving microphone audio on host.")

                if not self._audio_format_notice_sent:
                    self._audio_format_notice_sent = True
                    await self._publish_system(
                        "Inbound audio frame stats: "
                        f"sample_rate={int(getattr(frame, 'sample_rate', 0) or 0)}, "
                        f"samples={int(getattr(frame, 'samples', 0) or 0)}, "
                        f"time_base={getattr(frame, 'time_base', None)}."
                    )

                if (
                    self._stt_suppress_during_tts
                    and asyncio.get_running_loop().time() < self._stt_suppress_until
                ):
                    recording = False
                    recording_started_at = 0.0
                    segment_ms = 0.0
                    segment_buffer = bytearray()
                    continue

                if self._ptt_pressed:
                    if not recording:
                        recording = True
                        recording_started_at = asyncio.get_running_loop().time()
                        segment_ms = 0.0
                        segment_buffer = bytearray()

                    segment_buffer.extend(pcm16)
                    segment_ms += frame_ms
                    continue

                if recording:
                    observed_duration_ms = max(
                        1.0,
                        (asyncio.get_running_loop().time() - recording_started_at)
                        * 1000.0,
                    )
                    await self._finalize_ptt_segment(
                        bytes(segment_buffer),
                        segment_ms,
                        observed_duration_ms=observed_duration_ms,
                    )
                    recording = False
                    recording_started_at = 0.0
                    segment_ms = 0.0
                    segment_buffer = bytearray()
        except asyncio.CancelledError:
            raise
        except Exception as exc:
            details = str(exc).strip()
            if details:
                await self._publish_system(
                    f"Voice input stream ended ({exc.__class__.__name__}): {details}"
                )
            else:
                await self._publish_system(
                    f"Voice input stream ended ({exc.__class__.__name__})."
                )
        finally:
            if recording and segment_ms >= self._stt_min_ptt_ms:
                observed_duration_ms = max(
                    1.0,
                    (asyncio.get_running_loop().time() - recording_started_at) * 1000.0,
                )
                await self._finalize_ptt_segment(
                    bytes(segment_buffer),
                    segment_ms,
                    observed_duration_ms=observed_duration_ms,
                )

    async def _finalize_ptt_segment(
        self,
        pcm16: bytes,
        duration_ms: float,
        observed_duration_ms: float | None = None,
    ) -> None:
        if not pcm16 or duration_ms <= 0.0:
            return

        normalized_pcm = pcm16
        normalized_duration_ms = duration_ms
        if observed_duration_ms is not None and observed_duration_ms > 0.0:
            duration_ratio = duration_ms / observed_duration_ms
            if duration_ratio < 0.70 or duration_ratio > 1.40:
                estimated_source_rate = int(round(16_000 * duration_ratio))
                estimated_source_rate = max(8_000, min(96_000, estimated_source_rate))
                candidate_rates = [
                    8_000,
                    12_000,
                    16_000,
                    24_000,
                    32_000,
                    44_100,
                    48_000,
                ]
                nearest_source_rate = min(
                    candidate_rates,
                    key=lambda candidate: abs(candidate - estimated_source_rate),
                )
                if nearest_source_rate != 16_000:
                    normalized_pcm, _state = audioop.ratecv(
                        pcm16,
                        2,
                        1,
                        nearest_source_rate,
                        16_000,
                        None,
                    )
                    normalized_duration_ms = (len(normalized_pcm) / 2 / 16_000) * 1000.0
                    if not self._ptt_timing_correction_notice_sent:
                        self._ptt_timing_correction_notice_sent = True
                        await self._publish_system(
                            "Corrected PTT timing mismatch "
                            f"(estimated source={nearest_source_rate}Hz)."
                        )

        await self._enqueue_stt_segment(
            pcm16=normalized_pcm, duration_ms=normalized_duration_ms
        )

    async def _enqueue_stt_segment(self, pcm16: bytes, duration_ms: float) -> None:
        if duration_ms < self._stt_min_ptt_ms:
            return

        if self._stt_segments.full():
            with contextlib.suppress(asyncio.QueueEmpty):
                self._stt_segments.get_nowait()

            now = asyncio.get_running_loop().time()
            if (
                now - self._last_stt_backlog_notice_at
            ) >= self._stt_backlog_notice_interval_s:
                self._last_stt_backlog_notice_at = now
                await self._publish_system(
                    "Voice input backlog detected; dropping stale segment."
                )

        with contextlib.suppress(asyncio.QueueFull):
            self._stt_segments.put_nowait(pcm16)

    async def _stt_worker(self) -> None:
        while True:
            pcm16 = await self._stt_segments.get()
            if not self._stt_first_segment_notice_sent:
                self._stt_first_segment_notice_sent = True
                await self._publish_system(
                    "Push-to-talk audio captured. Running host STT..."
                )
            try:
                transcript = await self._stt.transcribe_pcm(
                    pcm=pcm16,
                    sample_rate=16_000,
                    channels=1,
                )
            except asyncio.CancelledError:
                raise
            except Exception as exc:
                await self._publish_system(f"Host STT failed: {exc}")
                continue

            if not transcript:
                continue

            transcript = transcript.strip()
            if not transcript:
                continue

            await self._gateway.bus.publish(
                WisperEvent(role="wisper", text=f"voice transcript: {transcript}")
            )
            await self._gateway.send_user_message(transcript)

    async def _close_peer_connection(self) -> None:
        if self._outbound_track:
            self._outbound_track.stop()
            self._outbound_track = None

        if self._pc:
            await self._pc.close()
            self._pc = None

        self._pending_ice_candidates.clear()

    async def _wait_for_ice_gathering(self, peer_connection: RTCPeerConnection) -> None:
        if peer_connection.iceGatheringState == "complete":
            return

        completed = asyncio.Event()

        @peer_connection.on("icegatheringstatechange")
        def on_icegatheringstatechange() -> None:
            if peer_connection.iceGatheringState == "complete":
                completed.set()

        with contextlib.suppress(asyncio.TimeoutError):
            await asyncio.wait_for(completed.wait(), timeout=3)

    async def _warmup_stt(self) -> None:
        try:
            await self._stt.warmup()
        except asyncio.CancelledError:
            raise
        except Exception:
            return

    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 = [
            AGENT_PREFIX_RE.sub("", line)
            for line in lines
            if not SPEECH_FILTER_RE.match(line)
            and not THINKING_STATUS_RE.search(line)
            and not USER_PREFIX_RE.match(line)
            and not VOICE_TRANSCRIPT_RE.match(line)
        ]
        return _sanitize_tts_text(" ".join(useful_lines))

    def _frame_to_pcm16k_mono(
        self, frame: AudioFrame, resample_state: tuple[Any, ...] | None
    ) -> tuple[bytes, float, tuple[Any, ...] | None]:
        try:
            pcm = frame.to_ndarray(format="s16")
        except TypeError:
            pcm = frame.to_ndarray()

        if (
            NUMPY_AVAILABLE
            and np is not None
            and getattr(pcm, "dtype", None) is not None
        ):
            if pcm.dtype != np.int16:
                if np.issubdtype(pcm.dtype, np.floating):
                    pcm = np.clip(pcm, -1.0, 1.0)
                    pcm = (pcm * 32767.0).astype(np.int16)
                else:
                    pcm = pcm.astype(np.int16)

        if pcm.ndim == 1:
            mono = pcm.tobytes()
        elif pcm.ndim == 2:
            expected_channels = 0
            if getattr(frame, "layout", None) is not None:
                with contextlib.suppress(Exception):
                    expected_channels = len(frame.layout.channels)

            rows = int(pcm.shape[0])
            cols = int(pcm.shape[1])

            # Normalize to [frames, channels] to avoid accidental channel mis-detection.
            if expected_channels > 0:
                if rows == expected_channels:
                    frames_channels = pcm.T
                elif cols == expected_channels:
                    frames_channels = pcm
                else:
                    frames_channels = pcm.reshape(-1, 1)
            else:
                if rows == 1:
                    frames_channels = pcm.T
                elif cols == 1:
                    frames_channels = pcm
                elif rows <= 8 and cols > rows:
                    frames_channels = pcm.T
                elif cols <= 8 and rows > cols:
                    frames_channels = pcm
                else:
                    frames_channels = pcm.reshape(-1, 1)

            channel_count = (
                int(frames_channels.shape[1]) if frames_channels.ndim == 2 else 1
            )
            if channel_count <= 1:
                mono = frames_channels.reshape(-1).tobytes()
            elif NUMPY_AVAILABLE and np is not None:
                mixed = frames_channels.astype(np.int32).mean(axis=1)
                mono = np.clip(mixed, -32768, 32767).astype(np.int16).tobytes()
            elif channel_count == 2:
                interleaved = frames_channels.reshape(-1).tobytes()
                mono = audioop.tomono(interleaved, 2, 0.5, 0.5)
            else:
                mono = frames_channels[:, 0].reshape(-1).tobytes()
        else:
            return b"", 0.0, resample_state

        source_rate = int(
            getattr(frame, "sample_rate", 0) or getattr(frame, "rate", 0) or 0
        )

        time_base = getattr(frame, "time_base", None)
        tb_rate = 0
        if time_base is not None:
            with contextlib.suppress(Exception):
                numerator = int(getattr(time_base, "numerator", 0))
                denominator = int(getattr(time_base, "denominator", 0))
                if numerator == 1 and denominator > 0:
                    tb_rate = denominator

        samples_per_channel = int(getattr(frame, "samples", 0) or 0)
        if samples_per_channel > 0:
            candidate_rates = [8_000, 16_000, 24_000, 32_000, 44_100, 48_000]
            inferred_rate = min(
                candidate_rates,
                key=lambda rate: abs((samples_per_channel / float(rate)) - 0.020),
            )
            inferred_frame_ms = (samples_per_channel / float(inferred_rate)) * 1000.0
            # If metadata suggests implausibly long frames, trust the inferred rate instead.
            if (
                source_rate <= 0
                or (samples_per_channel / float(max(1, source_rate))) * 1000.0 > 40.0
            ):
                source_rate = inferred_rate
            elif abs(inferred_frame_ms - 20.0) <= 2.5 and source_rate not in {
                inferred_rate,
                tb_rate,
            }:
                source_rate = inferred_rate

        if tb_rate > 0 and (source_rate <= 0 or abs(tb_rate - source_rate) > 2_000):
            source_rate = tb_rate
        if source_rate <= 0:
            source_rate = 48_000

        if source_rate != 16_000:
            mono, resample_state = audioop.ratecv(
                mono,
                2,
                1,
                source_rate,
                16_000,
                resample_state,
            )

        if not mono:
            return b"", 0.0, resample_state

        sample_count = len(mono) // 2
        duration_ms = (sample_count / 16_000) * 1000
        return mono, duration_ms, resample_state