Initial commit

agent.py

@@ -0,0 +1,148 @@
+"""
+agent.py — LiveKit voice agent (livekit-agents 1.7+)
+
+Requires an AgentServer at module level named 'agent'.
+"""
+
+import os, logging
+from dotenv import load_dotenv
+
+load_dotenv()
+
+from livekit.agents import (
+    Agent, AgentSession, AgentServer, JobContext, WorkerOptions, AutoSubscribe,
+    EndpointingOptions,
+)
+from livekit.plugins import openai as lk_openai, silero
+
+logger = logging.getLogger("agent")
+logging.basicConfig(level=logging.INFO)
+
+
+def make_tts():
+    backend = os.getenv("TTS_BACKEND", "kokoro")
+    if backend == "elevenlabs":
+        from livekit.plugins import elevenlabs
+        return elevenlabs.TTS(
+            api_key=os.environ["ELEVENLABS_API_KEY"],
+            voice_id=os.getenv("ELEVENLABS_VOICE_ID", "21m00Tcm4TlvDq8ikWAM"),
+        )
+    if backend == "piper":
+        return _make_piper_tts(os.getenv("PIPER_MODEL", "en_US-libritts-high"))
+    base_url = (
+        os.getenv("KOKORO_BASE_URL", "http://localhost:8880/v1")
+        if backend == "kokoro"
+        else os.getenv("OLLAMA_BASE_URL", "http://localhost:11434/v1")
+    )
+    voice = os.getenv("KOKORO_VOICE", "af_heart") if backend == "kokoro" else "alloy"
+    model = "kokoro" if backend == "kokoro" else "tts-1"
+    return lk_openai.TTS(model=model, voice=voice, base_url=base_url, api_key="x")
+
+
+def _make_piper_tts(model_name: str):
+    from piper.voice import PiperVoice
+    import wave, io, asyncio, pathlib, urllib.request
+    from livekit.agents import tts as lk_tts
+    from livekit import rtc
+
+    class PiperTTS(lk_tts.TTS):
+        def __init__(self):
+            super().__init__(streaming=False)
+            self._voice = None
+
+        def _load(self):
+            if self._voice: return
+            d = pathlib.Path.home() / ".local/share/piper"
+            d.mkdir(parents=True, exist_ok=True)
+            mp = d / f"{model_name}.onnx"
+            cp = d / f"{model_name}.onnx.json"
+            base = ("https://huggingface.co/rhasspy/piper-voices/resolve/main/"
+                    + "/".join(model_name.split("-")[:2]) + f"/medium/{model_name}")
+            for p, u in [(mp, base+".onnx"), (cp, base+".onnx.json")]:
+                if not p.exists():
+                    urllib.request.urlretrieve(u, p)
+            self._voice = PiperVoice.load(str(mp), config_path=str(cp))
+
+        def synthesize(self, text):
+            return _PiperStream(self, text)
+
+    class _PiperStream(lk_tts.ChunkedStream):
+        def __init__(self, tts, text):
+            super().__init__(tts=tts, input_text=text)
+
+        async def _run(self):
+            loop = asyncio.get_event_loop()
+            def _synth():
+                self._tts._load()
+                buf = io.BytesIO()
+                with wave.open(buf, "wb") as wf:
+                    self._tts._voice.synthesize(self._input_text, wf)
+                return buf.getvalue()
+            wav = await loop.run_in_executor(None, _synth)
+            buf = io.BytesIO(wav)
+            with wave.open(buf, "rb") as wf:
+                sr, nc, data = wf.getframerate(), wf.getnchannels(), wf.readframes(wf.getnframes())
+            self._event_ch.send_nowait(lk_tts.SynthesizedAudio(
+                request_id=self._request_id,
+                frame=rtc.AudioFrame(data=data, sample_rate=sr, num_channels=nc,
+                                     samples_per_channel=len(data)//(2*nc)),
+            ))
+
+    return PiperTTS()
+
+
+async def entrypoint(ctx: JobContext):
+    claims     = ctx.token_claims()
+    model      = (claims.attributes or {}).get("model") or os.getenv("DEFAULT_MODEL", "hf.co/mradermacher/Heretic-Dolphin3.0-Qwen2.5-3b-i1-GGUF:Q4_K_M")
+    prompt     = (claims.attributes or {}).get("prompt") or os.getenv("SYSTEM_PROMPT", "You are a helpful voice assistant. Keep your answers concise and conversational.")
+    ollama_url = os.getenv("OLLAMA_BASE_URL", "http://localhost:11434/v1")
+
+    logger.info(f"Starting agent: model={model} tts={os.getenv('TTS_BACKEND','kokoro')}")
+
+    await ctx.connect(auto_subscribe=AutoSubscribe.AUDIO_ONLY)
+
+    import openai as _openai
+    whisper_client = _openai.AsyncOpenAI(
+        base_url=os.getenv("WHISPER_BASE_URL", "http://localhost:8778/v1"),
+        api_key="x",
+        timeout=120.0,
+    )
+
+    session = AgentSession(
+        vad=silero.VAD.load(),
+        stt=lk_openai.STT(
+            model="Systran/faster-whisper-base",
+            client=whisper_client,
+        ),
+        llm=lk_openai.LLM(model=model, base_url=ollama_url, api_key="ollama"),
+        tts=make_tts(),
+        # Give whisper more time — it's slower than cloud STT
+        min_endpointing_delay=1.2,
+        max_endpointing_delay=6.0,
+    )
+
+    await session.start(
+        agent=Agent(instructions=prompt),
+        room=ctx.room,
+    )
+
+    await session.generate_reply(
+        instructions="Greet your creator briefly and casually"
+    )
+
+
+# AgentServer instance — required by livekit-agents 1.7+ module discovery
+agent = AgentServer(
+    ws_url=os.getenv("LIVEKIT_URL", "ws://localhost:7880"),
+    api_key=os.getenv("LIVEKIT_API_KEY", "devkey"),
+    api_secret=os.getenv("LIVEKIT_API_SECRET", "secret"),
+)
+
+@agent.rtc_session
+async def run_session(ctx: JobContext):
+    await entrypoint(ctx)
+
+if __name__ == "__main__":
+    import sys
+    from livekit.agents.__main__ import main
+    sys.exit(main())