- Implements TurnAwareTranscriptProcessor that combines user and assistant transcript tracking with turn boundary detection
- Correctly handles interruptions by capturing only what was actually spoken
- Emits on_turn_started and on_turn_ended events with accumulated transcripts
- Handles async frame processing with strategic delays to ensure proper text accumulation
- Adds comprehensive tests covering basic flow, interruptions, and multiple turns
- Includes documentation and usage examples
Introduced `LLMTextProcessor`: A new processor meant to allow customization for how
LLMTextFrames should be aggregated and considered. It's purpose is to turn
`LLMTextFrame`s into `AggregatedTextFrame`s. By default, a TTSService will still
aggregate `LLMTextFrame`s by sentence for the service to consume. However, if you
wish to override how the llm text is aggregated, you should no longer override the
TTS's internal text_aggregator, but instead, insert this processor between your LLM
and TTS in the pipeline.
This frame introduces an `aggregated_by` field to describe the type of text included
in the frame and allows unspoken groupings of text to be pushed through the pipeline
and treated similar to TTSTextFrames.
Modified the BaseTextAggregator type so that when text gets aggregated, metadata can
be associated with it. Currently, that just means a `type`, so that the aggregation
can be classified or described. Changes made to support this:
- **IMPORTANT**: Aggregators are now expected to strip leading/trailing white space
characters before returning their aggregation from `aggregation()` or `.text`. This
way all aggregators have a consistent contract allowing downstream use to know how
to stitch aggregations back together
- Introduced a new `Aggregation` dataclass to represent both the aggregated `text` and
a string identifying the `type` of aggregation (ex. "sentence", "word", "my custom
aggregation")
- **BREAKING**: `BaseTextAggregator.text` now returns an `Aggregation` (instead of `str`).
To update: `aggregated_text = myAggregator.text` -> `aggregated_text = myAggregator.text.text`
- **BREAKING**: `BaseTextAggregator.aggregate()` now returns `Optional[Aggregation]`
(instead of `Optional[str]`). To update:
```
aggregation = myAggregator.aggregate(text)
if (aggregation):
print(f"successfully aggregated text: {aggregation.text}") // instead of {aggregation}
```
- `SimpleTextAggregator`, `SkipTagsAggregator`, `PatternPairAggregator` updated to
produce/consume `Aggregation` objects.
- All uses of the above Aggregators have been updated accordingly.
- Usage in classes that are already deprecated
- Usage related to realtime LLMs, which don't yet support `LLMContext`
- Usage in (soon-to-be-deprecated) code paths related to `OpenAILLMContext` itself and associated machinery
Immediate is the "default", i.e. has the more obvious name (e.g. `ManuallySwitchServiceFrame` v `ManuallySwitchServiceControlFrame`), since that's *probably* what users will want to reach for. Also, the immediate frames are more likely to behave like what we had before the last few commits, where the service switch would always "jump the queue" by having an immediate effect once it hit the `ServiceSwitcher` in the pipeline, jumping ahead of frames in front of it destined for the service.
Watchdog timers have been removed. They were introduced in 0.0.72 to help
diagnose pipeline freezes. Unfortunately, they proved ineffective since they
required developers to use Pipecat-specific queues, iterators, and events to
correctly reset the timer, which limited their usefulness and added friction.
