realtime_voice_service/main.py

from __future__ import annotations

import asyncio
import logging
import os
import uuid
from contextlib import asynccontextmanager

from fastapi import FastAPI, WebSocket

from realtime_voice_service.core.session import CallSession
from realtime_voice_service.core.vad import SileroVADDetector
from realtime_voice_service.providers.llm import OpenAILLM
from realtime_voice_service.providers.stt import ElevenLabsSTT
from realtime_voice_service.providers.tts import ElevenLabsTTS
from realtime_voice_service.transports.audiosocket import AudioSocketServer
from realtime_voice_service.transports.base import BaseMediaTransport
from realtime_voice_service.transports.websocket import WebSocketMediaTransport


LOGGER = logging.getLogger("uvicorn.error")


def _int_env(name: str, default: int) -> int:
    raw = os.getenv(name)
    if raw is None:
        return default
    try:
        return int(str(raw).strip())
    except ValueError:
        return default


def _float_env(name: str, default: float) -> float:
    raw = os.getenv(name)
    if raw is None:
        return default
    try:
        return float(str(raw).strip())
    except ValueError:
        return default


def _optional_float_env(name: str) -> float | None:
    raw = os.getenv(name)
    if raw is None:
        return None
    normalized = str(raw).strip()
    if not normalized:
        return None
    try:
        return float(normalized)
    except ValueError:
        return None


def _bool_env(name: str, default: bool = False) -> bool:
    raw = os.getenv(name)
    if raw is None:
        return default
    return str(raw).strip().lower() in {"1", "true", "yes", "on"}


def _audiosocket_host() -> str:
    return str(os.getenv("REALTIME_VOICE_AUDIOSOCKET_HOST", "0.0.0.0") or "0.0.0.0").strip()


def _audiosocket_port() -> int:
    return max(_int_env("REALTIME_VOICE_AUDIOSOCKET_PORT", 9092), 1)


def _http_host() -> str:
    return str(os.getenv("REALTIME_VOICE_HTTP_HOST", "0.0.0.0") or "0.0.0.0").strip()


def _http_port() -> int:
    return max(_int_env("REALTIME_VOICE_HTTP_PORT", 8000), 1)


class RealtimeVoiceService:
    def __init__(self) -> None:
        self._stt = ElevenLabsSTT()
        self._llm = OpenAILLM()
        self._tts = ElevenLabsTTS()
        self._audiosocket_server = AudioSocketServer(
            host=_audiosocket_host(),
            port=_audiosocket_port(),
            session_handler=self._run_transport_session,
        )
        self._active_sessions: dict[str, CallSession] = {}
        self._session_lock = asyncio.Lock()

    @property
    def active_session_count(self) -> int:
        return len(self._active_sessions)

    async def start(self) -> None:
        await self._audiosocket_server.start()

    async def stop(self) -> None:
        await self._audiosocket_server.stop()
        sessions = list(self._active_sessions.values())
        for session in sessions:
            await session.stop()
        self._active_sessions.clear()

    async def handle_websocket(self, websocket: WebSocket, *, client_id: str | None = None) -> None:
        await websocket.accept()
        transport = WebSocketMediaTransport(
            websocket=websocket,
            transport_id=client_id or str(uuid.uuid4()),
        )
        await self._run_transport_session(transport)

    async def _run_transport_session(self, transport: BaseMediaTransport) -> None:
        session = self._build_session(transport)
        async with self._track_session(session):
            LOGGER.info(
                "starting realtime session %s via %s",
                session.session_id,
                transport.protocol,
            )
            await session.run()

    def _build_session(self, transport: BaseMediaTransport) -> CallSession:
        return CallSession(
            session_id=transport.transport_id,
            transport=transport,
            vad=SileroVADDetector(
                sample_rate_hz=transport.sample_rate_hz,
                threshold=_float_env("VAD_THRESHOLD", 0.5),
                negative_threshold=_optional_float_env("VAD_NEGATIVE_THRESHOLD"),
                speech_end_silence_ms=_int_env("VAD_SILENCE_TIMEOUT_MS", 1600),
                speech_pad_ms=_int_env("VAD_SPEECH_PAD_MS", 64),
                min_speech_duration_ms=_int_env("VAD_MIN_SPEECH_DURATION_MS", 0),
                use_onnx=_bool_env("VAD_USE_ONNX", False),
            ),
            stt=self._stt,
            llm=self._llm,
            tts=self._tts,
        )

    @asynccontextmanager
    async def _track_session(self, session: CallSession):
        async with self._session_lock:
            self._active_sessions[session.session_id] = session
        try:
            yield
        finally:
            async with self._session_lock:
                self._active_sessions.pop(session.session_id, None)


service = RealtimeVoiceService()


@asynccontextmanager
async def lifespan(_: FastAPI):
    await service.start()
    try:
        yield
    finally:
        await service.stop()


app = FastAPI(title="realtime-voice-service", version="0.1.0", lifespan=lifespan)


@app.get("/health")
async def health() -> dict[str, object]:
    return {
        "status": "ok",
        "service": "realtime-voice-service",
        "active_sessions": service.active_session_count,
        "audiosocket_port": _audiosocket_port(),
    }


@app.websocket("/ws/{client_id}")
async def websocket_media(websocket: WebSocket, client_id: str) -> None:
    await service.handle_websocket(websocket, client_id=client_id)


if __name__ == "__main__":
    import uvicorn

    uvicorn.run(
        "realtime_voice_service.main:app",
        host=_http_host(),
        port=_http_port(),
        reload=False,
    )