The Whisper-based ONNX model expects 16 kHz audio, but the
_predict_endpoint method had five hardcoded references to 16000 without
checking the actual pipeline sample rate. When running at 8 kHz (e.g.
Twilio telephony), audio was fed to the feature extractor at the wrong
rate, causing the model to perceive speech at 2x speed with shifted
formant frequencies and produce incorrect end-of-turn predictions.
Add automatic resampling via numpy interpolation before feature
extraction and replace all hardcoded sample rate values with a
_MODEL_SAMPLE_RATE constant. Also fix the WAV debug logger to write
files with the correct sample rate header.
Fixes#3844
When filter_incomplete_user_turns is enabled and an LLMMessagesUpdateFrame
replaces the context via set_messages(), the turn completion instructions
system message was lost. This caused the LLM to stop emitting turn
completion markers. Re-inject the instructions after set_messages() to
fix this.
Allow pushing frames upstream through the pipeline by passing
FrameDirection.UPSTREAM. Downstream frames use the existing push queue,
while upstream frames are pushed directly from the pipeline sink.
The ServiceSettings refactor (PR #3714) changed self._settings from
dicts to dataclass subclasses, but tracing code still used .items(),
in containment, and subscript access, causing AttributeError on
every traced call. Use given_fields() for iteration and attribute
access for named fields.
Replace _rtvi_external instance variable with a local prepend_rtvi flag
since it is only used during __init__ to decide whether to prepend the
RTVIProcessor to the pipeline.
When the user places an RTVIProcessor inside their pipeline and provides
a custom RTVIObserver subclass in observers, PipelineTask correctly
detects both and logs "skipping default ones." However it then
unconditionally prepends self._rtvi to the pipeline, causing the
processor to appear twice in the frame chain.
Track whether the RTVIProcessor was found externally (inside the user
pipeline) vs created internally. Only prepend it when created internally.
Fixes#3867
Even when summarization_timeout is explicitly set to None, use a
DEFAULT_SUMMARIZATION_TIMEOUT (120s) fallback so the LLM call can
never hang indefinitely. Applied in both LLMService and the dedicated
LLM path in LLMContextSummarizer.
The dedicated LLM logic lived in LLMAssistantAggregator, creating two
code paths and requiring the aggregator to call a private LLMService
method. Move it into the summarizer which already owns the config and
summarization lifecycle, keeping the aggregator handler as a single-line
upstream push.
Adds a configurable summarization_timeout (default 120s) that cancels
summary generation if the LLM hangs. On timeout, an error result is
returned so _summarization_in_progress resets and future
summarizations are unblocked.
Adds an field to LLMContextSummarizationConfig that allows
routing summarization to a separate LLM service (e.g., Gemini Flash)
instead of the pipeline's primary model. This avoids paying for
expensive inference when compressing context in long-running sessions.
Allows applications to customize how the summary is wrapped when
injected into context (e.g., XML tags, custom delimiters) so system
prompts can distinguish summaries from live conversation.
Add deprecation warnings to start_processing_metrics() and
stop_processing_metrics() on FrameProcessorMetrics and FrameProcessor.
Mark ProcessingMetricsData as deprecated in docstring. All existing
behavior is preserved — the warnings inform users that these will be
removed in a future version.
- Add InterruptionFrame handling with stop_all_metrics()
- Add processing metrics (start/stop) at response boundaries
- Fix agent transcript handling for voice and text modalities:
- Voice mode: push LLMTextFrame (append_to_context=False) and
TTSTextFrame for deltas, skip duplicated final text
- Text mode: push LLMTextFrame with proper response lifecycle,
no TTSTextFrame (downstream TTS handles audio)
- Add output_medium parameter to AgentInputParams and OneShotInputParams
- Improve TTFB measurement using VAD speech end time
- Update example with user turn strategies and transcript events
- Add text-only output example (50a-ultravox-realtime-text.py)
Move the sentence vs token aggregation concern into text aggregators
so all text flows through them regardless of mode. This enables
pattern detection and tag handling to work in TOKEN mode.
- Add TextAggregationMode enum (SENTENCE, TOKEN) as the user-facing
TTS setting, separate from the internal AggregationType
- Add TOKEN mode support to Simple, SkipTags, and PatternPair aggregators
- Add text_aggregation_mode parameter to TTSService and all TTS subclasses
- Deprecate aggregate_sentences in favor of text_aggregation_mode
- Merge TTSService._process_text_frame() into a single codepath
Add TextAggregationMetricsData measuring the time from the first LLM
token to the first complete sentence, representing the latency cost of
sentence aggregation in the TTS pipeline.
