refactor party tonight

src/dailyai/services/ai_services.py

@@ -2,6 +2,8 @@ import asyncio
 import logging
 import re
 
+from httpx import request
+
 from dailyai.queue_frame import QueueFrame, FrameType
 
 from abc import abstractmethod
@@ -13,9 +15,7 @@ from collections.abc import Iterable, AsyncIterable
 
 class AIService:
 
-    def __init__(
+    def __init__(self):
-        self
-    ):
         self.logger = logging.getLogger("dailyai")
 
     def stop(self):
@@ -27,30 +27,60 @@ class AIService:
     def possible_output_frame_types(self) -> set[FrameType]:
         return set()
 
+    async def run_to_queue(self, queue: asyncio.Queue, frames, add_end_of_stream=False) -> None:
+        async for frame in self.run(frames):
+            print("got frame", frame.frame_type)
+            await queue.put(frame)
+
+        if add_end_of_stream:
+            await queue.put(QueueFrame(FrameType.END_STREAM, None))
+
     async def run(
-            self,
+        self,
-            requested_frame_types:set[FrameType],
+        frames: Iterable[QueueFrame]
-            frames:Iterable[QueueFrame] | AsyncIterable[QueueFrame]
+        | AsyncIterable[QueueFrame]
-            ) -> AsyncGenerator[QueueFrame, None]:
+        | asyncio.Queue[QueueFrame],
-        if self.possible_output_frame_types().intersection(requested_frame_types) == set():
+        requested_frame_types: set[FrameType] | None=None,
+    ) -> AsyncGenerator[QueueFrame, None]:
+        if requested_frame_types and self.possible_output_frame_types().intersection(requested_frame_types) == set():
             raise Exception(f"Requested frame types {requested_frame_types} are not supported by this service.")
 
+        if not requested_frame_types:
+            requested_frame_types = self.possible_output_frame_types()
+
+        print("running", self.__class__.__name__, "with frame types", requested_frame_types)
+
         if isinstance(frames, AsyncIterable):
             async for frame in frames:
-                output_frame: QueueFrame | None = await self.process_frame(requested_frame_types, frame)
+                async for output_frame in self.process_frame(requested_frame_types, frame):
-                if output_frame:
+                    print(
+                        "yielding frame", self.__class__.__name__, output_frame.frame_type
+                    )
                     yield output_frame
         elif isinstance(frames, Iterable):
             for frame in frames:
-                output_frame = await self.process_frame(requested_frame_types, frame)
+                async for output_frame in self.process_frame(requested_frame_types, frame):
-                if output_frame:
+                    print(
+                        "yielding frame", self.__class__.__name__, output_frame.frame_type
+                    )
                     yield output_frame
+        elif isinstance(frames, asyncio.Queue):
+            while True:
+                frame = await frames.get()
+                async for output_frame in self.process_frame(requested_frame_types, frame):
+                    print(
+                        "yielding frame", self.__class__.__name__, output_frame.frame_type
+                    )
+                    yield output_frame
+                if frame.frame_type == FrameType.END_STREAM:
+                    break
         else:
             raise Exception("Frames must be an iterable or async iterable")
 
     @abstractmethod
-    async def process_frame(self, requested_frame_types:set[FrameType], frame:QueueFrame) -> QueueFrame | None:
+    async def process_frame(self, requested_frame_types:set[FrameType], frame:QueueFrame) -> AsyncGenerator[QueueFrame, None]:
-        pass
+        # Yield something so the linter can deduce what should happen here.
+        yield QueueFrame(FrameType.END_STREAM, None)
 
 class SentenceAggregator(AIService):
     def __init__(self, **kwargs):
@@ -63,29 +93,26 @@ class SentenceAggregator(AIService):
     def possible_output_frame_types(self) -> set[FrameType]:
         return set([FrameType.SENTENCE])
 
-    async def process_frame(self, requested_frame_types: set[FrameType], frame: QueueFrame) -> QueueFrame | None:
+    async def process_frame(self, requested_frame_types: set[FrameType], frame: QueueFrame) -> AsyncGenerator[QueueFrame, None]:
         if not FrameType.SENTENCE in requested_frame_types:
-            return None
+            return
 
         if frame.frame_type == FrameType.TEXT_CHUNK:
             if type(frame.frame_data) != str:
-                raise Exception("Sentence aggregator requires a string for the data field")
+                raise Exception(
+                    "Sentence aggregator requires a string for the data field"
+                )
 
             self.current_sentence += frame.frame_data
             if self.current_sentence.endswith((".", "?", "!")):
                 sentence = self.current_sentence
                 self.current_sentence = ""
-                return QueueFrame(FrameType.SENTENCE, sentence)
+                yield QueueFrame(FrameType.SENTENCE, sentence)
-            return None
         elif frame.frame_type == FrameType.END_STREAM:
             if self.current_sentence:
-                return QueueFrame(FrameType.SENTENCE, self.current_sentence)
+                yield QueueFrame(FrameType.SENTENCE, self.current_sentence)
-            else:
-                return None
         elif frame.frame_type == FrameType.SENTENCE:
-            return frame
+            yield frame
-        else:
-            return None
 
 
 class LLMService(AIService):
@@ -93,30 +120,29 @@ class LLMService(AIService):
         return set([FrameType.LLM_MESSAGE, FrameType.SENTENCE, FrameType.TRANSCRIPTION])
 
     def allowed_output_frame_types(self) -> set[FrameType]:
-        return set([FrameType.SENTENCE, FrameType.SENTENCE, FrameType.TEXT_CHUNK])
+        return set([FrameType.SENTENCE, FrameType.TEXT_CHUNK])
 
-    async def run_llm_async_sentences(self, messages) -> AsyncGenerator[str, None]:
+    @abstractmethod
-        current_text = ""
+    async def run_llm_async(self, messages) -> AsyncGenerator[str, None]:
-        async for text in self.run_llm_async(messages):
+        yield ""
-            current_text += text
-            if re.match(r"^.*[.!?]$", text):
-                yield current_text
-                current_text = ""
 
-        if current_text:
+    @abstractmethod
-            yield current_text
+    async def run_llm(self, messages) -> str:
-
+        pass
-    async def process_frame(self, frame:QueueFrame) -> QueueFrame | None:
-        if not self.output_queue:
-            raise Exception("Output queue must be set before using the run method.")
 
+    async def process_frame(self, requested_frame_types: set[FrameType], frame: QueueFrame) -> AsyncGenerator[QueueFrame, None]:
         if frame.frame_type == FrameType.LLM_MESSAGE:
             if type(frame.frame_data) != list:
                 raise Exception("LLM service requires a dict for the data field")
 
             messages: list[dict[str, str]] = frame.frame_data
-            async for message in self.run_llm_async_sentences(messages):
+            if FrameType.SENTENCE in requested_frame_types:
-                await self.output_queue.put(QueueFrame(FrameType.SENTENCE, message))
+                yield QueueFrame(FrameType.SENTENCE, await self.run_llm(messages))
+            else:
+                async for text_chunk in self.run_llm_async(messages):
+                    yield QueueFrame(FrameType.TEXT_CHUNK, text_chunk)
+
+        # TODO: handle other frame types! Need to aggregate into messages
 
 
 class TTSService(AIService):
@@ -124,6 +150,12 @@ class TTSService(AIService):
     def get_mic_sample_rate(self):
         return 16000
 
+    def allowed_input_frame_types(self) -> set[FrameType]:
+        return set([FrameType.SENTENCE, FrameType.TRANSCRIPTION, FrameType.TEXT_CHUNK])
+
+    def possible_output_frame_types(self) -> set[FrameType]:
+        return set([FrameType.AUDIO])
+
     # Converts the sentence to audio. Yields a list of audio frames that can
     # be sent to the microphone device
     @abstractmethod
@@ -131,25 +163,48 @@ class TTSService(AIService):
         # yield empty bytes here, so linting can infer what this method does
         yield bytes()
 
-    async def process_frame(self, frame:QueueFrame) -> QueueFrame | None:
+    async def process_frame(self, requested_frame_types: set[FrameType], frame: QueueFrame) -> AsyncGenerator[QueueFrame, None]:
-        if not self.output_queue:
+        if not FrameType.AUDIO in requested_frame_types:
-            raise Exception("Output queue must be set before using the run method.")
+            return
 
-        if frame.frame_type == FrameType.SENTENCE:
+        if type(frame.frame_data) != str:
-            if type(frame.frame_data) != str:
+            raise Exception("TTS service requires a string for the data field")
-                raise Exception("TTS service requires a string for the data field")
 
-            text = frame.frame_data
+        async for audio_chunk in self.run_tts(frame.frame_data):
-            async for audio in self.run_tts(text):
+            yield QueueFrame(FrameType.AUDIO, audio_chunk)
-                await self.output_queue.put(QueueFrame(FrameType.AUDIO, audio))
+
+    # Convenience function to send the audio for a sentence to the given queue
+    async def say(self, sentence, queue: asyncio.Queue):
+        async for audio_chunk in self.run_tts(sentence):
+            await queue.put(QueueFrame(FrameType.AUDIO, audio_chunk))
 
 
 class ImageGenService(AIService):
+    def __init__(self, image_size, **kwargs):
+        super().__init__(**kwargs)
+        self.image_size = image_size
+
+    def allowed_input_frame_types(self) -> set[FrameType]:
+        return set([FrameType.SENTENCE, FrameType.TRANSCRIPTION, FrameType.TEXT_CHUNK, FrameType.IMAGE_DESCRIPTION])
+
+    def possible_output_frame_types(self) -> set[FrameType]:
+        return set([FrameType.IMAGE])
+
     # Renders the image. Returns an Image object.
     @abstractmethod
-    async def run_image_gen(self, sentence, size) -> tuple[str, bytes]:
+    async def run_image_gen(self, sentence) -> tuple[str, bytes]:
         pass
 
+    async def process_frame(self, requested_frame_types: set[FrameType], frame: QueueFrame) -> AsyncGenerator[QueueFrame, None]:
+        if not FrameType.IMAGE in requested_frame_types:
+            return
+
+        if type(frame.frame_data) != str:
+            raise Exception("Image service requires a string for the data field")
+
+        (_, image_data) = await self.run_image_gen(frame.frame_data)
+        yield QueueFrame(FrameType.IMAGE, image_data)
+
 
 @dataclass
 class AIServiceConfig:

src/dailyai/services/azure_ai_services.py

@@ -16,8 +16,8 @@ from PIL import Image
 from azure.cognitiveservices.speech import SpeechSynthesizer, SpeechConfig, ResultReason, CancellationReason
 
 class AzureTTSService(TTSService):
-    def __init__(self, input_queue=None, output_queue=None, speech_key=None, speech_region=None):
+    def __init__(self, speech_key=None, speech_region=None):
-        super().__init__(input_queue, output_queue)
+        super().__init__()
 
         speech_key = speech_key or os.getenv("AZURE_SPEECH_SERVICE_KEY")
         speech_region = speech_region or os.getenv("AZURE_SPEECH_SERVICE_REGION")
@@ -35,7 +35,10 @@ class AzureTTSService(TTSService):
            "<prosody rate='1.05'>" \
            f"{sentence}" \
            "</prosody></mstts:express-as></voice></speak> "
-        result = await asyncio.to_thread(self.speech_synthesizer.speak_ssml, (ssml))
+        try:
+            result = await asyncio.to_thread(self.speech_synthesizer.speak_ssml, (ssml))
+        except Exception as e:
+            self.logger.error("Error in azure tts", e)
         self.logger.info("Got azure tts result")
         if result.reason == ResultReason.SynthesizingAudioCompleted:
             self.logger.info("Returning result")
@@ -48,8 +51,8 @@ class AzureTTSService(TTSService):
                 self.logger.info("Error details: {}".format(cancellation_details.error_details))
 
 class AzureLLMService(LLMService):
-    def __init__(self, input_queue=None, output_queue=None, api_key=None, azure_endpoint=None, api_version=None, model=None):
+    def __init__(self, api_key=None, azure_endpoint=None, api_version=None, model=None):
-        super().__init__(input_queue, output_queue)
+        super().__init__()
         api_key = api_key or os.getenv("AZURE_CHATGPT_KEY")
 
         azure_endpoint = azure_endpoint or os.getenv("AZURE_CHATGPT_ENDPOINT")
@@ -92,14 +95,14 @@ class AzureLLMService(LLMService):
 
 class AzureImageGenServiceREST(ImageGenService):
 
-    def __init__(self, api_key=None, azure_endpoint=None, api_version=None, model=None):
+    def __init__(self, image_size:str, api_key=None, azure_endpoint=None, api_version=None, model=None):
-        super().__init__()
+        super().__init__(image_size=image_size)
         self.api_key = api_key or os.getenv("AZURE_DALLE_KEY")
         self.azure_endpoint = azure_endpoint or os.getenv("AZURE_DALLE_ENDPOINT")
         self.api_version = api_version or "2023-06-01-preview"
         self.model = model or os.getenv("AZURE_DALLE_DEPLOYMENT_ID")
 
-    async def run_image_gen(self, sentence, size) -> tuple[str, bytes]:
+    async def run_image_gen(self, sentence) -> tuple[str, bytes]:
         # TODO hoist the session to app-level
         async with aiohttp.ClientSession() as session:
             url = f"{self.azure_endpoint}openai/images/generations:submit?api-version={self.api_version}"
@@ -107,7 +110,7 @@ class AzureImageGenServiceREST(ImageGenService):
             body = {
                 # Enter your prompt text here
                 "prompt": sentence,
-                "size": size,
+                "size": self.image_size,
                 "n": 1,
             }
             async with session.post(url, headers=headers, json=body) as submission:
@@ -153,14 +156,14 @@ class AzureImageGenService(ImageGenService):
             api_version=api_version,
         )
 
-    async def run_image_gen(self, sentence, size) -> tuple[str, bytes]:
+    async def run_image_gen(self, sentence) -> tuple[str, bytes]:
         self.logger.info("Generating azure image", sentence)
 
         image = self.client.images.generate(
             model=self.model,
             prompt=sentence,
             n=1,
-            size=size,
+            size=self.image_size,
         )
 
         url = image["data"][0]["url"]

src/dailyai/services/daily_transport_service.py

@@ -206,6 +206,9 @@ class DailyTransportService(EventHandler):
             if frame.frame_type == FrameType.END_STREAM:
                 break
 
+    def wait_for_send_queue_to_empty(self):
+        self.threadsafe_send_queue.join()
+
     async def run(self) -> None:
         self.configure_daily()
 

src/dailyai/services/elevenlabs_ai_service.py

@@ -9,8 +9,8 @@ from dailyai.services.ai_services import TTSService
 
 
 class ElevenLabsTTSService(TTSService):
-    def __init__(self, input_queue=None, output_queue=None, api_key=None, voice_id=None):
+    def __init__(self, api_key=None, voice_id=None):
-        super().__init__(input_queue, output_queue)
+        super().__init__()
 
         self.api_key = api_key or os.getenv("ELEVENLABS_API_KEY")
         self.voice_id = voice_id or os.getenv("ELEVENLABS_VOICE_ID")

src/dailyai/services/open_ai_services.py

@@ -50,20 +50,20 @@ class OpenAILLMService(LLMService):
             return None
 
 class OpenAIImageGenService(ImageGenService):
-    def __init__(self, api_key=None, model=None):
+    def __init__(self, image_size:str, api_key=None, model=None):
-        super().__init__()
+        super().__init__(image_size=image_size)
         api_key = api_key or os.getenv("OPEN_AI_KEY")
         self.model = model or os.getenv("OPEN_AI_IMAGE_MODEL") or "dall-e-3"
         self.client = AsyncOpenAI(api_key=api_key)
 
-    async def run_image_gen(self, sentence, size) -> tuple[str, bytes]:
+    async def run_image_gen(self, sentence) -> tuple[str, bytes]:
         self.logger.info("Generating OpenAI image", sentence)
 
         image = await self.client.images.generate(
             prompt=sentence,
             model=self.model,
             n=1,
-            size=size
+            size=self.image_size
         )
         image_url = image.data[0].url
         if not image_url:

src/samples/theoretical-to-real/01-say-one-thing.py

@@ -27,21 +27,16 @@ async def main(room_url):
     # similarly, create a tts service
     tts = AzureTTSService()
 
-    # Get the generator for the audio. This will start running in the background,
-    # and when we ask the generator for its items, we'll get what it's generated.
-    audio_generator: AsyncGenerator[bytes, None] = tts.run_tts("hello world")
-
     # Register an event handler so we can play the audio when the participant joins.
     @transport.event_handler("on_participant_joined")
     async def on_participant_joined(transport, participant):
         if participant["info"]["isLocal"]:
             return
 
-        async for audio in audio_generator:
+        await tts.say("Hello there, " + participant["info"]["userName"] + "!", transport.send_queue)
-            transport.output_queue.put(QueueFrame(FrameType.AUDIO, audio))
 
         # wait for the output queue to be empty, then leave the meeting
-        transport.output_queue.join()
+        transport.wait_for_send_queue_to_empty()
         transport.stop()
 
     await transport.run()

src/samples/theoretical-to-real/02-llm-say-one-thing.py

@@ -4,6 +4,7 @@ from typing import AsyncGenerator
 
 from dailyai.queue_frame import QueueFrame, FrameType
 from dailyai.services.daily_transport_service import DailyTransportService
+from dailyai.services.ai_services import SentenceAggregator
 from dailyai.services.azure_ai_services import AzureLLMService
 from dailyai.services.elevenlabs_ai_service import ElevenLabsTTSService
 
@@ -17,29 +18,27 @@ async def main(room_url):
     )
     transport.mic_enabled = True
 
-    text_to_llm_queue = asyncio.Queue()
+    tts = ElevenLabsTTSService(voice_id="29vD33N1CtxCmqQRPOHJ")
-    llm_to_tts_queue = asyncio.Queue()
+    llm = AzureLLMService()
-
-    tts = ElevenLabsTTSService(
-        llm_to_tts_queue, transport.get_async_send_queue(), voice_id="29vD33N1CtxCmqQRPOHJ"
-    )
-    llm = AzureLLMService(text_to_llm_queue, llm_to_tts_queue)
 
     messages = [{
         "role": "system",
         "content": "You are an LLM in a WebRTC session, and this is a 'hello world' demo. Say hello to the world."
     }]
-    await text_to_llm_queue.put(QueueFrame(FrameType.LLM_MESSAGE, messages))
+    tts_task = asyncio.create_task(
-    await text_to_llm_queue.put(QueueFrame(FrameType.END_STREAM, None))
+        tts.run_to_queue(
-
+            transport.send_queue,
-    llm_task = asyncio.create_task(llm.run())
+            SentenceAggregator().run(
+                llm.run([QueueFrame(FrameType.LLM_MESSAGE, messages)])
+            )
+        )
+    )
 
     @transport.event_handler("on_first_other_participant_joined")
     async def on_first_other_participant_joined(transport):
-        await asyncio.gather(llm_task, tts.run())
+        await tts_task
 
-        # wait for the output queue to be empty, then leave the meeting
+        transport.wait_for_send_queue_to_empty()
-        transport.output_queue.join()
         transport.stop()
 
     await transport.run()

src/samples/theoretical-to-real/03-still-frame.py

@@ -21,13 +21,14 @@ async def main(room_url):
     transport.camera_width = 1024
     transport.camera_height = 1024
 
-    imagegen = OpenAIImageGenService()
+    imagegen = OpenAIImageGenService(image_size="1024x1024")
-    image_task = asyncio.create_task(imagegen.run_image_gen("a cat in the style of picasso", "1024x1024"))
+    image_task = asyncio.create_task(
+        imagegen.run_to_queue(transport.send_queue, [QueueFrame(FrameType.IMAGE_DESCRIPTION, "a cat in the style of picasso")])
+    )
 
     @transport.event_handler("on_participant_joined")
     async def on_participant_joined(transport, participant):
-        (_, image_bytes) = await image_task
+        await image_task
-        transport.output_queue.put(QueueFrame(FrameType.IMAGE, image_bytes))
 
     await transport.run()
 
@@ -38,6 +39,6 @@ if __name__ == "__main__":
         "-u", "--url", type=str, required=True, help="URL of the Daily room to join"
     )
 
-    args: argparse.Namespace = parser.parse_args()
+    args, unknown = parser.parse_known_args()
 
     asyncio.run(main(args.url))

src/samples/theoretical-to-real/04-utterance-and-speech.py

@@ -2,9 +2,11 @@ import argparse
 import asyncio
 import re
 
+from dailyai.services.ai_services import SentenceAggregator
 from dailyai.services.daily_transport_service import DailyTransportService
 from dailyai.services.azure_ai_services import AzureLLMService, AzureTTSService
 from dailyai.queue_frame import QueueFrame, FrameType
+from dailyai.services.elevenlabs_ai_service import ElevenLabsTTSService
 
 async def main(room_url:str):
     global transport
@@ -22,34 +24,46 @@ async def main(room_url:str):
     transport.camera_enabled = False
 
     llm = AzureLLMService()
-    tts = AzureTTSService()
+    azure_tts = AzureTTSService()
+    elevenlabs_tts = ElevenLabsTTSService(voice_id="ErXwobaYiN019PkySvjV")
+
+    messages = [{"role": "system", "content": "tell the user a joke about llamas"}]
+
+    # Start a task to run the LLM to create a joke, and convert the LLM output to audio frames. This task
+    # will run in parallel with generating and speaking the audio for static text, so there's no delay to
+    # speak the LLM response.
+    buffer_queue = asyncio.Queue()
+    llm_response_task = asyncio.create_task(
+        elevenlabs_tts.run_to_queue(
+            buffer_queue,
+            SentenceAggregator().run(
+                llm.run([QueueFrame(FrameType.LLM_MESSAGE, messages)])
+            ),
+            True,
+        )
+    )
 
     @transport.event_handler("on_participant_joined")
     async def on_joined(transport, participant):
         if participant["id"] == transport.my_participant_id:
             return
 
-        # queue two pieces of speech: one specified as a text literal,
+        await azure_tts.run_to_queue(
-        # and one generated by an llm. We'll kick off the llm first, and let
+            transport.send_queue,
-        # it generate a response while we're speaking the literal string.
+            [QueueFrame(FrameType.SENTENCE, "My friend the LLM is now going to tell a joke about llamas.")]
-        #
+        )
-        # Note that in this case, we don't use `run_llm_async` because we're
-        # taking advantage of the time spent speaking the first phrase to generate
-        # the entire LLM response, and this happens asynchronously in a task.
-        llm_response_task = asyncio.create_task(llm.run_llm(
-            [{"role": "system", "content": "tell the user a joke about llamas"}]
-        ))
 
-        async for audio_chunk in tts.run_tts("My friend the LLM is now going to tell a joke about llamas."):
+        async def buffer_to_send_queue():
-            transport.output_queue.put(QueueFrame(FrameType.AUDIO, audio_chunk))
+            while True:
+                frame = await buffer_queue.get()
+                await transport.send_queue.put(frame)
+                buffer_queue.task_done()
+                if frame.frame_type == FrameType.END_STREAM:
+                    break
 
-        llm_response = await llm_response_task
+        await asyncio.gather(llm_response_task, buffer_to_send_queue())
-        async for audio_chunk in tts.run_tts(llm_response):
-            transport.output_queue.put(QueueFrame(FrameType.AUDIO, audio_chunk))
 
-
+        transport.wait_for_send_queue_to_empty()
-        # wait for the output queue to be empty, then leave the meeting
-        transport.output_queue.join()
         transport.stop()
 
     await transport.run()
@@ -61,6 +75,6 @@ if __name__ == "__main__":
         "-u", "--url", type=str, required=True, help="URL of the Daily room to join"
     )
 
-    args: argparse.Namespace = parser.parse_args()
+    args, unknown = parser.parse_known_args()
 
     asyncio.run(main(args.url))

src/samples/theoretical-to-real/05-sync-speech-and-text.py

@@ -26,10 +26,9 @@ async def main(room_url):
     transport.camera_height = 1024
 
     llm = AzureLLMService()
-    #tts = ElevenLabsTTSService(voice_id="ErXwobaYiN019PkySvjV")
-    tts = ElevenLabsTTSService()
     dalle = FalImageGenService()
-    # dalle = OpenAIImageGenService() 
+    tts = ElevenLabsTTSService(voice_id="ErXwobaYiN019PkySvjV")
+    #dalle = OpenAIImageGenService(image_size="1024x1024")
 
     # Get a complete audio chunk from the given text. Splitting this into its own
     # coroutine lets us ensure proper ordering of the audio chunks on the output queue.
@@ -61,7 +60,7 @@ async def main(room_url):
 
             tts_tasks.append(get_all_audio(sentence))
 
-        tts_tasks.insert(0, dalle.run_image_gen(image_text, "1024x1024"))
+        tts_tasks.insert(0, dalle.run_image_gen(image_text))
 
         print(f"waiting for tasks to finish for {month}")
         data = await asyncio.gather(