カスタムLLMインテグレーション

pip install "elevenlabs" "aiohttp" "twilio" "python-dotenv"

from elevenlabs import AsyncElevenLabs
elevenlabs = AsyncElevenLabs(api_key=os.environ["ELEVENLABS_API_KEY"])
async def signed_url() -> str:
    response = await elevenlabs.conversational_ai.conversations.get_signed_url(
        agent_id=os.environ["SPEECH_ENGINE_ID"],
    )
    return response.signed_url

from aiohttp import web
from twilio.request_validator import RequestValidator
validator = RequestValidator(os.environ["TWILIO_AUTH_TOKEN"])
async def incoming_call(request: web.Request) -> web.Response:
    form = await request.post()
    signature = request.headers.get("X-Twilio-Signature", "")
    url = str(request.url)
    if not validator.validate(url, dict(form), signature):
        return web.Response(status=403, text="forbidden")
    host = request.headers.get("X-Forwarded-Host") or request.host
    twiml = (
        '<?xml version="1.0" encoding="UTF-8"?>'
        "<Response><Connect>"
        f'<Stream url="wss://{host}/media-stream"/>'
        "</Connect></Response>"
    )
    return web.Response(text=twiml, content_type="text/xml")

import asyncio
import json
import aiohttp
from aiohttp import web
async def media_stream(request: web.Request) -> web.WebSocketResponse:
    twilio_ws = web.WebSocketResponse()
    await twilio_ws.prepare(request)
    stream_sid: str | None = None
    el_session: aiohttp.ClientSession | None = None
    el_ws: aiohttp.ClientWebSocketResponse | None = None
    pump_task: asyncio.Task | None = None
    async def pump_el_to_twilio(el: aiohttp.ClientWebSocketResponse):
        async for msg in el:
            if msg.type != aiohttp.WSMsgType.TEXT:
                continue
            event = json.loads(msg.data)
            etype = event.get("type")
            if etype == "audio":
                await twilio_ws.send_str(json.dumps({
                    "event": "media",
                    "streamSid": stream_sid,
                    "media": {"payload": event["audio_event"]["audio_base_64"]},
                }))
            elif etype == "interruption":
                await twilio_ws.send_str(json.dumps({
                    "event": "clear",
                    "streamSid": stream_sid,
                }))
            elif etype == "ping":
                event_id = event.get("ping_event", {}).get("event_id")
                await el.send_str(json.dumps({
                    "type": "pong", "event_id": event_id,
                }))
    try:
        async for msg in twilio_ws:
            if msg.type != aiohttp.WSMsgType.TEXT:
                continue
            event = json.loads(msg.data)
            if event["event"] == "start":
                stream_sid = event["start"]["streamSid"]
                el_session = aiohttp.ClientSession()
                el_ws = await el_session.ws_connect(await signed_url())
                await el_ws.send_str(json.dumps({
                    "type": "conversation_initiation_client_data",
                }))
                pump_task = asyncio.create_task(pump_el_to_twilio(el_ws))
            elif event["event"] == "media" and el_ws is not None:
                await el_ws.send_str(json.dumps({
                    "user_audio_chunk": event["media"]["payload"],
                }))
            elif event["event"] == "stop":
                break
    finally:
        if pump_task:
            pump_task.cancel()
        if el_ws and not el_ws.closed:
            await el_ws.close()
        if el_session and not el_session.closed:
            await el_session.close()
    return twilio_ws

import os
from elevenlabs import AsyncElevenLabs
elevenlabs = AsyncElevenLabs(api_key=os.environ["ELEVENLABS_API_KEY"])
SHARED_SECRET = os.environ["SHARED_SECRET"]
async def brain_ws(request: web.Request) -> web.WebSocketResponse:
    if request.headers.get("x-api-key") != SHARED_SECRET:
        return web.Response(status=401, text="unauthorized")
    ws = web.WebSocketResponse()
    await ws.prepare(request)
    engine = await elevenlabs.speech_engine.get(os.environ["SPEECH_ENGINE_ID"])
    session = engine.create_session(ws)
    async def on_transcript(transcript):
        # Replace this with your own LLM call; see the quickstart.
        await session.send_response("Hello, you've reached the demo.")
    session.on("user_transcript", on_transcript)
    await session.run()
    return ws
def make_app() -> web.Application:
    app = web.Application()
    app.router.add_post("/incoming-call", incoming_call)
    app.router.add_get("/media-stream", media_stream)
    app.router.add_get("/ws", brain_ws)
    return app
if __name__ == "__main__":
    web.run_app(make_app(), port=3001)

await elevenlabs.speech_engine.update(
    speech_engine_id="seng_8k3m9xr4hjnfg983brhmhkd98n6",
    speech_engine={"ws_url": "wss://abc123.ngrok.io/ws"},
)