tts.say() doesn't execute

[agilebean]

Description

Bug

Environment

• pipecat-ai version: 0.0.47
• python version: 3.12
• OS: MacOS 14.2

Issue description

The tts.say() doesn't work, i.e. doesn't push the audio.

Repro steps

tts = PlayHTTTSService(
        user_id=os.getenv("PLAYHT_USER_ID"),
        api_key=os.getenv("PLAYHT_API_KEY"),
        voice_url=voice_url
    )
tts.say("This is a text message to test the TTS say method.")

Potential root cause

the say method flushes the audio as last statement:

async def say(self, text: str):
    aggregate_sentences = self._aggregate_sentences
    self._aggregate_sentences = False
    await self.process_frame(TextFrame(text=text), FrameDirection.DOWNSTREAM)
    self._aggregate_sentences = aggregate_sentences
    await self.flush_audio()

However, this is an abstract method defined in TTSService:

@abstractmethod
async def flush_audio(self):
    pass

In several TTS classes, this method is not implemented, e.g. PlayHTTTSService

Therefore, the await self.flush_audio() call doesn't perform any operation. Consequently, the process_frame() might not end because it doesn't process the TTSStoppedFrame here:

async def run_tts(self, text: str) -> AsyncGenerator[Frame, None]:
    try:
        yield TTSStartedFrame()
        # ... process audio chunks ...
        yield TTSAudioRawFrame(chunk, self._settings["sample_rate"], 1)
        # ... more processing ...
        yield TTSStoppedFrame()  # This frame might never be processed

The above is just a speculation because I lack the understanding of the underlying mechanisms of processing frames.

Solution

1. Is there a generic solution for flush_audio which works for all TTS services?
2. If not, what would be a workaround?
   A workaround would be very helpful also to use with other methods like push_frame(TTSSpeakFrame()) or task.queue_frames to ensure completion, as currently, they all seem to queue the frames.