Allow users to set output format for TTS in livekit-plugins-openai

[zhanghx0905]

Hello, I am using edgetts to simulate the OpenAI TTS service for a local, free voice agent. However, edgetts only supports returning audio in MP3 format. In livekit-plugins-openai version 0.10.9 and later, the request format has been set to PCM.

Proposed Solution

Add a configuration option or parameter to livekit.plugins.openai.tts.TTS to allow users to specify the desired output format (e.g., output_format='mp3' or output_format='pcm').

Temporary Patch

This mismatch requires me to dynamically modify livekit.plugins.openai.tts.ChunkedStream at runtime to handle the MP3 format, which is not ideal.

import httpx
from livekit.agents import (
    APIConnectionError,
    APIStatusError,
    APITimeoutError,
    tts,
    utils,
)

import openai
from livekit.plugins.openai.tts import (
    OPENAI_TTS_SAMPLE_RATE,
    OPENAI_TTS_CHANNELS,
    ChunkedStream,
)


async def _run(self: ChunkedStream):
    print(f"TTS debug: {self.input_text}")
    oai_stream = self._client.audio.speech.with_streaming_response.create(
        input=self.input_text,
        model=self._opts.model,
        voice=self._opts.voice,  # type: ignore
        response_format="mp3",
        speed=self._opts.speed,
        timeout=httpx.Timeout(30, connect=self._conn_options.timeout),
    )
    decoder = utils.codecs.Mp3StreamDecoder()
    request_id = utils.shortuuid()
    audio_bstream = utils.audio.AudioByteStream(
        sample_rate=OPENAI_TTS_SAMPLE_RATE,
        num_channels=OPENAI_TTS_CHANNELS,
    )

    try:
        async with oai_stream as stream:
            async for data in stream.iter_bytes():
                for frame in decoder.decode_chunk(data):
                    for frame in audio_bstream.write(frame.data.tobytes()):
                        self._event_ch.send_nowait(
                            tts.SynthesizedAudio(
                                frame=frame,
                                request_id=request_id,
                            )
                        )

            for frame in audio_bstream.flush():
                self._event_ch.send_nowait(
                    tts.SynthesizedAudio(
                        frame=frame,
                        request_id=request_id,
                    )
                )

    except openai.APITimeoutError:
        raise APITimeoutError()
    except openai.APIStatusError as e:
        raise APIStatusError(
            e.message,
            status_code=e.status_code,
            request_id=e.request_id,
            body=e.body,
        )
    except Exception as e:
        raise APIConnectionError() from e


ChunkedStream._run = _run