import asyncio from typing import AsyncGenerator, Tuple from pyflowery.exceptions import ResponseError from pyflowery.models import FloweryAPIConfig, Language, Voice from pyflowery.rest_adapter import RestAdapter class FloweryAPI: """Main class for interacting with the Flowery API Attributes: config (FloweryAPIConfig): Configuration object for the API adapter (RestAdapter): Adapter for making HTTP requests """ def __init__(self, config: FloweryAPIConfig) -> None: self.config = config self.adapter = RestAdapter(config=config) self._voices_cache: Tuple[Voice, ...] = () try: loop = asyncio.get_running_loop() except RuntimeError: loop = None if loop and loop.is_running(): self.config.logger.info("Async event loop is already running. Adding `_populate_voices_cache()` to the event loop.") asyncio.create_task(self._populate_voices_cache()) else: asyncio.run(self._populate_voices_cache()) async def _populate_voices_cache(self) -> None: """Populate the voices cache. This method is called automatically when the FloweryAPI object is created, and should not be called directly.""" self._voices_cache = tuple([voice async for voice in self.fetch_voices()]) # pylint: disable=consider-using-generator if not self._voices_cache: raise ValueError("Failed to populate voices cache! Please report this issue at https://www.coastalcommits.com/cswimr/PyFlowery/issues.") self.config.logger.info("Voices cache populated!") def get_voices(self, voice_id: str | None = None, name: str | None = None) -> Tuple[Voice, ...] | None: """Get a set of voices from the cache. Args: voice_id (str): The ID of the voice name (str): The name of the voice Returns: Tuple[Voice] | None: A tuple of Voice objects if found, otherwise None """ if voice_id: voice = next((voice for voice in self._voices_cache if voice.id == voice_id), None) return (voice,) if voice else None if name: voices = [] for voice in self._voices_cache: if voice.name == name: voices.append(voice) return tuple(voices) or None return self._voices_cache or None async def fetch_voice(self, voice_id: str) -> Voice: """Fetch a voice from the Flowery API. This method bypasses the cache and directly queries the Flowery API. You should usually use `get_voices()` instead. Args: voice_id (str): The ID of the voice Raises: ValueError: Raised when the voice is not found TooManyRequests: Raised when the Flowery API returns a 429 status code ClientError: Raised when the Flowery API returns a 4xx status code InternalServerError: Raised when the Flowery API returns a 5xx status code RetryLimitExceeded: Raised when the retry limit defined in the `FloweryAPIConfig` class (default 3) is exceeded Returns: Voice: The voice """ async for voice in self.fetch_voices(): if voice.id == voice_id: return voice raise ValueError(f"Voice with ID {voice_id} not found.") async def fetch_voices(self) -> AsyncGenerator[Voice, None]: """Fetch a list of voices from the Flowery API Raises: TooManyRequests: Raised when the Flowery API returns a 429 status code ClientError: Raised when the Flowery API returns a 4xx status code InternalServerError: Raised when the Flowery API returns a 5xx status code ResponseError: Raised when the Flowery API returns an empty response or a response with an unexpected format RetryLimitExceeded: Raised when the retry limit defined in the `FloweryAPIConfig` class (default 3) is exceeded Returns: AsyncGenerator[Voice, None]: A generator of Voices """ request = await self.adapter.get(endpoint="/tts/voices") if request is not None: if isinstance(request.data, dict): for voice in request.data.get("voices", []): yield Voice( id=voice["id"], name=voice["name"], gender=voice["gender"], source=voice["source"], language=Language(**voice["language"]), ) else: raise ResponseError(f"Invalid response from Flowery API: {request.data!r}") else: raise ResponseError("Invalid response from Flowery API: Empty Response!}") async def fetch_tts( self, text: str, voice: Voice | str | None = None, translate: bool = False, silence: int = 0, audio_format: str = "mp3", speed: float = 1.0 ) -> bytes: """Fetch a TTS audio file from the Flowery API Args: text (str): The text to convert to speech voice (Voice | str): The voice to use for the speech translate (bool): Whether to translate the text silence (int): Number of seconds of silence to add to the end of the audio audio_format (str): The audio format to return speed (float): The speed of the speech Raises: ValueError: Raised when the provided text is too long and `allow_truncation` in the `FloweryAPIConfig` class is set to `False` (default). TooManyRequests: Raised when the Flowery API returns a 429 status code ClientError: Raised when the Flowery API returns a 4xx status code InternalServerError: Raised when the Flowery API returns a 5xx status code ResponseError: Raised when the Flowery API returns an empty response or a response with an unexpected format RetryLimitExceeded: Raised when the retry limit defined in the `FloweryAPIConfig` class (default 3) is exceeded Returns: bytes: The audio file in bytes """ if len(text) > 2048: if not self.config.allow_truncation: raise ValueError("Text must be less than or equal to 2048 characters") self.config.logger.warning("Text is too long, will be truncated to 2048 characters by the API") params = { "text": text, "translate": str(translate).lower(), "silence": silence, "audio_format": audio_format, "speed": speed, } if voice: params["voice"] = voice.id if isinstance(voice, Voice) else voice request = await self.adapter.get(endpoint="/tts", params=params, timeout=180) if request is not None: if isinstance(request.data, bytes): return request.data raise ResponseError(f"Invalid response from Flowery API: {request.data!r}") raise ResponseError("Invalid response from Flowery API: Empty Response!}")