fix(vad): throttle speech activity events while speaking

[perfloop-agent]

Summary

VADController now routes speaking-state activity through _maybe_speech_activity(), so the configured speech_activity_period is honored while audio continues to arrive. This removes redundant UserSpeakingFrame broadcasts through the downstream pipeline while preserving raw audio forwarding and speech start/stop transitions.

The focused regression tests cover the controller's default activity period and the VADProcessor activity, start, stop, and forwarding paths. The changelog records the user-visible fix.

Testing

• uv run --locked --group dev pytest tests/test_vad_controller.py tests/test_vad_processor.py
• uv run --locked --group dev ruff check src/pipecat/audio/vad/vad_controller.py tests/test_vad_controller.py tests/test_vad_processor.py
• uv run --locked --group dev ruff format --check src/pipecat/audio/vad/vad_controller.py tests/test_vad_controller.py tests/test_vad_processor.py
• uv run --locked --group dev towncrier build --draft --version Unreleased
• CI Unit Tests configuration: 5,653 passed and 127 skipped.
• Repository-wide Pyright: 0 errors, with two pre-existing pyaudio warnings.

Workload: PipelineWorker VADProcessor with SileroVADAnalyzer replaying the repository's bundled 16 kHz speech trace five times as 20 ms frames with the default 200 ms speech-activity period

Metric	Before	After	Change
VADProcessor process CPU nanoseconds for five speech-trace replays	820000000	690000000	15.9% lower
VADProcessor elapsed nanoseconds for five speech-trace replays	842079916	730178996	13.3% lower
Checks: 6 passed.

---

Generated by Perfloop. Human sponsor: Tomás Senart. Measurements and checks.

[tsenart]

Hi, I'm Tomás, I build Perfloop. I approved this contribution and will handle review feedback here.

We've had performance PRs merged in ClickHouse, Prometheus, Ruff/ty, parquet-go, RoaringBitmap, and more.

Before approving, we verified the change against main: the controller's docstring documents on_speech_activity as periodic with a 0.2 s default, but _maybe_speech_activity() has never been called since VADController landed in January, so every audio frame fires the event and the aggregator broadcasts a UserSpeakingFrame per frame. The consumers of that event, the pipeline idle timeout and the wake-phrase turn strategy, work on second-scale timers, and speech start and stop remain separate events. The tests ran exactly as your CI "Unit Tests" job runs them (5,653 passed, 127 skipped) plus repository-wide Pyright; the workflows here only need an approve-and-run.

[aconchillo]

Good catch. LGTM!

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/pipecat/audio/vad/vad_controller.py	97.56% <100.00%> (+4.87%)	⬆️

... and 13 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.