live_test.py

import asyncio
import base64
import json
import os
import pyaudio
from websockets.asyncio.client import connect


class SimpleGeminiVoice:
    def __init__(self):
        self.audio_queue = asyncio.Queue()
        self.api_key = os.environ.get("GEMINI_API_KEY")
        self.model = "gemini-2.0-flash-exp"
        self.uri = f"wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent?key={self.api_key}"
        # Audio settings
        self.FORMAT = pyaudio.paInt16
        self.CHANNELS = 1
        self.CHUNK = 512
        self.RATE = 16000

    async def start(self):
        # Initialize websocket
        self.ws = await connect(
            self.uri, additional_headers={"Content-Type": "application/json"}
        )
        await self.ws.send(json.dumps({"setup": {"model": f"models/{self.model}"}}))
        await self.ws.recv(decode=False)
        print("Connected to Gemini, You can start talking now")
        # Start audio streaming
        async with asyncio.TaskGroup() as tg:
            tg.create_task(self.capture_audio())
            tg.create_task(self.stream_audio())
            tg.create_task(self.play_response())

    async def capture_audio(self):
        audio = pyaudio.PyAudio()
        stream = audio.open(
            format=self.FORMAT,
            channels=self.CHANNELS,
            rate=self.RATE,
            input=True,
            frames_per_buffer=self.CHUNK,
        )

        while True:
            data = await asyncio.to_thread(stream.read, self.CHUNK)
            await self.ws.send(
                json.dumps(
                    {
                        "realtime_input": {
                            "media_chunks": [
                                {
                                    "data": base64.b64encode(data).decode(),
                                    "mime_type": "audio/pcm",
                                }
                            ]
                        }
                    }
                )
            )

    async def stream_audio(self):
        async for msg in self.ws:
            response = json.loads(msg)
            try:
                audio_data = response["serverContent"]["modelTurn"]["parts"][0][
                    "inlineData"
                ]["data"]
                self.audio_queue.put_nowait(base64.b64decode(audio_data))
            except KeyError:
                pass
            try:
                turn_complete = response["serverContent"]["turnComplete"]
            except KeyError:
                pass
            else:
                if turn_complete:
                    # If you interrupt the model, it sends an end_of_turn. For interruptions to work, we need to empty out the audio queue
                    print("\nEnd of turn")
                    while not self.audio_queue.empty():
                        self.audio_queue.get_nowait()

    async def play_response(self):
        audio = pyaudio.PyAudio()
        stream = audio.open(
            format=self.FORMAT, channels=self.CHANNELS, rate=24000, output=True
        )
        while True:
            data = await self.audio_queue.get()
            await asyncio.to_thread(stream.write, data)


if __name__ == "__main__":
    client = SimpleGeminiVoice()
    asyncio.run(client.start())

no_queue_version.py

import asyncio
import base64
import json
import os
import pyaudio
from websockets.asyncio.client import connect


class SimpleGeminiVoice:
    def __init__(self):
        self.api_key = os.environ.get("GEMINI_API_KEY")
        self.model = "gemini-2.0-flash-exp"
        self.uri = f"wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent?key={self.api_key}"
        # Audio settings
        self.FORMAT = pyaudio.paInt16
        self.CHANNELS = 1
        self.CHUNK = 512

    async def start(self):
        # Initialize websocket
        self.ws = await connect(
            self.uri, additional_headers={"Content-Type": "application/json"}
        )
        await self.ws.send(json.dumps({"setup": {"model": f"models/{self.model}"}}))
        await self.ws.recv(decode=False)
        print("Connected to Gemini, You can start talking now")
        # Start audio streaming
        async with asyncio.TaskGroup() as tg:
            tg.create_task(self.send_user_audio())
            tg.create_task(self.recv_model_audio())

    async def send_user_audio(self):
        audio = pyaudio.PyAudio()
        stream = audio.open(
            format=self.FORMAT,
            channels=self.CHANNELS,
            rate="16000",
            input=True,
            frames_per_buffer=self.CHUNK,
        )

        while True:
            data = await asyncio.to_thread(stream.read, self.CHUNK)
            await self.ws.send(
                json.dumps(
                    {
                        "realtime_input": {
                            "media_chunks": [
                                {
                                    "data": base64.b64encode(data).decode(),
                                    "mime_type": "audio/pcm",
                                }
                            ]
                        }
                    }
                )
            )

    async def recv_model_audio(self):
        audio = pyaudio.PyAudio()
        stream = audio.open(
            format=self.FORMAT, channels=self.CHANNELS, rate=24000, output=True
        )
        async for msg in self.ws:
            response = json.loads(msg)
            try:
                audio_data = response["serverContent"]["modelTurn"]["parts"][0][
                    "inlineData"
                ]["data"]
                await asyncio.to_thread(stream.write, base64.b64decode(audio_data))
            except KeyError:
                pass


if __name__ == "__main__":
    client = SimpleGeminiVoice()
    asyncio.run(client.start())

Here for the system instruction shouldn't the role be system rather than user? @2187Nick

Im just winging it. Based it off of python-genai/google/genai/tests/live/live_test.py

But now the responses have slowed way down. Basically unusable until they get it back running faster.

I would like to use it in a voice channel through discord.py, but I can't seem to get it to work.

I've created a playground repository you can run and fork easily https://github.com/saharmor/gemini-multimodal-playground
You can change the system prompt and voice and enable/disable interruptions. cc @avinashgawali @Youmyluck @boozuk

I've created a playground repository you can run and fork easily https://github.com/saharmor/gemini-multimodal-playground You can change the system prompt and voice and enable/disable interruptions. cc @avinashgawali @Youmyluck @boozuk

Thank you

And he don't speak Russian language. He understand Russian, but answers Chinese or something like that instead of Russian.

Думаю русский добавят чуть позже, как было с LearnLM. Он поначалу тоже прикидывался в незнании русского.

Просто везде пишут, что он добавлен уже. И я подумал может это политика гугла такая.

Пока нет. Ждем 2025. Середина января/начало февраля.

Perhaps we can use VAD to filter the noise form the audio and only send voice? @philschmid

@jlia0 this is already implemented in the standalone script here https://github.com/saharmor/gemini-multimodal-playground

@jlia0 this is already implemented in the standalone script here https://github.com/saharmor/gemini-multimodal-playground

thanks so much 🫶