Pipecat-ai

Added

- Added `OpenAITTSService`.

- Allow passing `output_format` and `model_id` to `CartesiaTTSService` to change
audio sample format and the model to use.

- Added `DailyRESTHelper` which helps you create Daily rooms and tokens in an
easy way.

- `PipelineTask` now has a `has_finished()` method to indicate if the task has
completed. If a task is never ran `has_finished()` will return False.

- `PipelineRunner` now supports SIGTERM. If received, the runner will be
cancelled.

Fixed

- Fixed an issue where `BaseInputTransport` and `BaseOutputTransport` where
stopping push tasks before pushing `EndFrame` frames could cause the bots to
get stuck.

- Fixed an error closing local audio transports.

- Fixed an issue with Deepgram TTS that was introduced in the previous release.

- Fixed `AnthropicLLMService` interruptions. If an interruption occurred, a
`user` message could be appended after the previous `user` message. Anthropic
does not allow that because it requires alternate `user` and `assistant`
messages.

Performance

- The `BaseInputTransport` does not pull audio frames from sub-classes any
more. Instead, sub-classes now push audio frames into a queue in the base
class. Also, `DailyInputTransport` now pushes audio frames every 20ms instead
of 10ms.

- Remove redundant camera input thread from `DailyInputTransport`. This should
improve performance a little bit when processing participant videos.

- Load Cartesia voice on startup.

Added

- Added WebsocketServerTransport. This will create a websocket server and will
read messages coming from a client. The messages are serialized/deserialized
with protobufs. See `examples/websocket-server` for a detailed example.

- Added function calling (LLMService.register_function()). This will allow the
LLM to call functions you have registered when needed. For example, if you
register a function to get the weather in Los Angeles and ask the LLM about
the weather in Los Angeles, the LLM will call your function.
See https://platform.openai.com/docs/guides/function-calling

- Added new `LangchainProcessor`.

- Added Cartesia TTS support (https://cartesia.ai/)

Fixed

- Fixed SileroVAD frame processor.

- Fixed an issue where `camera_out_enabled` would cause the highg CPU usage if
no image was provided.

Performance

- Removed unnecessary audio input tasks.

Added

- Exposed `on_dialin_ready` for Daily transport SIP endpoint handling. This
notifies when the Daily room SIP endpoints are ready. This allows integrating
with third-party services like Twilio.

- Exposed Daily transport `on_app_message` event.

- Added Daily transport `on_call_state_updated` event.

- Added Daily transport `start_recording()`, `stop_recording` and
`stop_dialout`.

Changed

- Added `PipelineParams`. This replaces the `allow_interruptions` argument in
`PipelineTask` and will allow future parameters in the future.

- Fixed Deepgram Aura TTS base_url and added ErrorFrame reporting.

- GoogleLLMService `api_key` argument is now mandatory.

Fixed

- Daily tranport `dialin-ready` doesn't not block anymore and it now handles
timeouts.

- Fixed AzureLLMService.

Fixed

- Fixed an issue handling Daily transport `dialin-ready` event.

Added

- Added Daily transport `start_dialout()` to be able to make phone or SIP calls.
See https://reference-python.daily.co/api_reference.html#daily.CallClient.start_dialout

- Added Daily transport support for dial-in use cases.

- Added Daily transport events: `on_dialout_connected`, `on_dialout_stopped`,
`on_dialout_error` and `on_dialout_warning`. See
https://reference-python.daily.co/api_reference.html#daily.EventHandler

Added

- Added vision support to Anthropic service.

- Added `WakeCheckFilter` which allows you to pass information downstream only
if you say a certain phrase/word.

Changed

- `FrameSerializer.serialize()` and `FrameSerializer.deserialize()` are now
`async`.

- `Filter` has been renamed to `FrameFilter` and it's now under
`processors/filters`.

Fixed

- Fixed Anthropic service to use new frame types.

- Fixed an issue in `LLMUserResponseAggregator` and `UserResponseAggregator`
that would cause frames after a brief pause to not be pushed to the LLM.

- Clear the audio output buffer if we are interrupted.

- Re-add exponential smoothing after volume calculation. This makes sure the
volume value being used doesn't fluctuate so much.