feat: dynamic endpointing delay #4720
Conversation
| from ..utils import is_given | ||
|
|
||
|
|
||
| class ExponentialMovingAverage: |
There was a problem hiding this comment.
We have one here:
chenghao-mou
force-pushed
the
feat/AGT-2485-dynamic-endpointing-delay
branch
from
fa7cffb to
ff0c7b1
Compare
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| if is_given(min_delay): | ||
| self._min_delay = min_delay | ||
| self._utterance_pause.reset(initial=self._min_delay, min_val=self._min_delay) | ||
|
|
||
| if is_given(max_delay): | ||
| self._max_delay = max_delay | ||
| self._turn_pause.reset(initial=self._max_delay, max_val=self._max_delay) |
There was a problem hiding this comment.
🟡 update_options leaves cross-filter clamp bounds stale
When update_options changes min_delay or max_delay, it only updates the bounds on one of the two ExpFilter instances, leaving the other filter's bounds stale. Both _utterance_pause and _turn_pause are initialized with min_val=min_delay, max_val=max_delay, but update_options only resets one filter per parameter.
Root Cause and Impact
At initialization (endpointing.py:39-44), both filters share the same bounds:
self._utterance_pause = ExpFilter(alpha=alpha, initial=min_delay, min_val=min_delay, max_val=max_delay)
self._turn_pause = ExpFilter(alpha=alpha, initial=max_delay, min_val=min_delay, max_val=max_delay)But update_options only partially updates:
update_options(max_delay=2.0)resets_turn_pausewithmax_val=2.0, but_utterance_pause._max_valstays at the oldmax_delay(e.g.1.0). This meansmin_delay(read from_utterance_pause.value) can never exceed the oldmax_delay, even though the configuredmax_delayhas increased.update_options(min_delay=0.5)resets_utterance_pausewithmin_val=0.5, but_turn_pause._min_valstays at the oldmin_delay(e.g.0.3). This meansmax_delay(read from_turn_pause.value) could drop below the newmin_delay, violating the invariantmax_delay >= min_delay.
Impact: After dynamically updating endpointing delay bounds, the EMA-filtered values can be clamped to incorrect/stale limits, potentially making the dynamic adjustment ineffective or violating the min ≤ max invariant.
| if is_given(min_delay): | |
| self._min_delay = min_delay | |
| self._utterance_pause.reset(initial=self._min_delay, min_val=self._min_delay) | |
| if is_given(max_delay): | |
| self._max_delay = max_delay | |
| self._turn_pause.reset(initial=self._max_delay, max_val=self._max_delay) | |
| if is_given(min_delay): | |
| self._min_delay = min_delay | |
| self._utterance_pause.reset(initial=self._min_delay, min_val=self._min_delay) | |
| self._turn_pause.reset(min_val=self._min_delay) | |
| if is_given(max_delay): | |
| self._max_delay = max_delay | |
| self._turn_pause.reset(initial=self._max_delay, max_val=self._max_delay) | |
| self._utterance_pause.reset(max_val=self._max_delay) | |
Was this helpful? React with 👍 or 👎 to provide feedback.
| if is_given(alpha): | ||
| self._alpha = alpha | ||
| self._filtered = -1.0 | ||
| self._filtered = initial |
There was a problem hiding this comment.
🟡 ExpFilter.reset unconditionally clears filtered value even when only updating bounds
The reset method always executes self._filtered = initial (line 30), even when initial is not provided (NOT_GIVEN). This means calling reset(min_val=0.5) to only update the min bound will also discard the accumulated EMA state.
Root Cause and Impact
At exp_filter.py:30, self._filtered = initial runs unconditionally. Since initial defaults to NOT_GIVEN, any call to reset() that doesn't explicitly pass initial will wipe the filtered value:
def reset(self, alpha=NOT_GIVEN, initial=NOT_GIVEN, min_val=NOT_GIVEN, max_val=NOT_GIVEN):
if is_given(alpha):
self._alpha = alpha
self._filtered = initial # always runs — clears EMA state if initial not providedThis is particularly relevant because the suggested fix for BUG-0001 would call self._turn_pause.reset(min_val=self._min_delay) (without initial), which would unintentionally clear the _turn_pause EMA state. Without fixing this, any caller that wants to update only min_val or max_val via reset() will lose the accumulated filter state.
Impact: Accumulated EMA state is lost when reset is called to update only bounds, causing the filter to restart from scratch.
| self._filtered = initial | |
| if is_given(initial): | |
| self._filtered = initial | |
Was this helpful? React with 👍 or 👎 to provide feedback.
2 participants
Adjust the endpointing delays based on the pause statistics of a running session:
EMA for both
with values capped to the original min and max.