LiveKit OpenAI Plugin with DeepInfra Backend: API Returns Usage Logs but No Audio Output

Getting Started
,
1

When using the livekit-plugins-openai TTS class configured with a custom base_url to route requests to DeepInfra’s OpenAI-compatible endpoint, the API successfully processes the request and logs show usage/successful responses, but no audio output is produced or played back on the client side.

from livekit.plugins import assemblyai, google, silero

from livekit.plugins.openai import TTS

async def create_interview_session(

    tts: TTS, llm: google.LLM, stt: assemblyai.STT

) -> AgentSession:

    """Factory that creates the configured AgentSession"""


    return AgentSession(

        stt=stt,

        llm=llm,

        tts=tts,

        vad=silero.VAD.load(activation_threshold=0.7),

        turn_detection="stt",

        min_endpointing_delay=0,

    )

tts = TTS(

    base_url="https://api.deepinfra.com/v1",

    api_key=settings.TTS_API_KEY,

    model="hexgrad/Kokoro-82M",

    voice="af_bella",

)
2

Your base_url and key are fine the model name is what’s breaking it. livekit-plugins-openai only takes the raw-audio branch for tts-1/tts-1-hd (AUDIO_STREAM_MODELS (https://github.com/livekit/agents/blob/main/livekit-plugins/livekit-plugins-penai/livekit/plugins/openai/tts.py)); anything else, including hexgrad/Kokoro-82M, goes to SSEChunkedStream, which sends stream_format=“sse” and only emits audio from lines starting with data: carrying a speech.audio.delta. DeepInfra just returns plain MP3 bytes, so no line ever matches, nothing gets pushed, and you get a billed 200 with zero frames and no error.

Force the audio branch and it plays:

  from livekit.agents.types import DEFAULT_API_CONNECT_OPTIONS
  from livekit.plugins.openai.tts import TTS, AudioChunkedStream

  class DeepInfraTTS(TTS):
      def synthesize(self, text, *, conn_options=DEFAULT_API_CONNECT_OPTIONS):
          return AudioChunkedStream(tts=self, input_text=text, conn_options=conn_options)
3

First of all huge thanks you for help.
The current method works but its slow than usual.
Do you have any recommendation for cheap and fast alternative?

4

You can use fishaudio tts? its cheaper and faster i think, it also has s2.1 pro free model currently

5

Quick update this is fixed upstream. The plugin now picks the parser from the response content-type instead of the model name, so DeepInfra works with no workaround: fix(openai): read TTS response by content-type, not model name by abidullahdev · Pull Request #6930 · livekit/agents · GitHub

Should be in the next release. Until then, AUDIO_STREAM_MODELS.add(“hexgrad/Kokoro-82M”) at import time is the smallest patch if you want to stay on Kokoro.