wx44wx/pipecat
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Improved the accuracy of the UserBotLatencyObserver and UserBotLatencyLogObserver

Browse Source
This commit is contained in:
Mark Backman
2026-02-06 19:36:08 -05:00
parent 34b068d657
commit 5e66702cf5
3 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
changelog/3637.changed.5.md Normal file
View File

@@ -0,0 +1 @@
- Improved the accuracy of the `UserBotLatencyObserver` and `UserBotLatencyLogObserver` by measuring from the time when the user actually starts speaking.

2
src/pipecat/observers/loggers/user_bot_latency_log_observer.py
View File

@@ -79,7 +79,7 @@ class UserBotLatencyLogObserver(BaseObserver):
if isinstance(data.frame, VADUserStartedSpeakingFrame): if isinstance(data.frame, VADUserStartedSpeakingFrame):
self._user_stopped_time = 0 self._user_stopped_time = 0
elif isinstance(data.frame, VADUserStoppedSpeakingFrame): elif isinstance(data.frame, VADUserStoppedSpeakingFrame):
self._user_stopped_time = data.frame.timestamp self._user_stopped_time = data.frame.timestamp - data.frame.stop_secs
elif isinstance(data.frame, (EndFrame, CancelFrame)): elif isinstance(data.frame, (EndFrame, CancelFrame)):
self._log_summary() self._log_summary()
elif isinstance(data.frame, BotStartedSpeakingFrame) and self._user_stopped_time: elif isinstance(data.frame, BotStartedSpeakingFrame) and self._user_stopped_time:

6
src/pipecat/observers/user_bot_latency_observer.py
View File

@@ -72,8 +72,10 @@ class UserBotLatencyObserver(BaseObserver):
# Reset when user starts speaking # Reset when user starts speaking
self._user_stopped_time = None self._user_stopped_time = None
elif isinstance(data.frame, VADUserStoppedSpeakingFrame): elif isinstance(data.frame, VADUserStoppedSpeakingFrame):
# Record timestamp when user stops speaking # Record the actual time the user stopped speaking, which is
self._user_stopped_time = data.frame.timestamp # the VAD determination time minus the stop_secs silence duration
# that had to elapse before the VAD confirmed speech ended.
self._user_stopped_time = data.frame.timestamp - data.frame.stop_secs
elif isinstance(data.frame, BotStartedSpeakingFrame) and self._user_stopped_time: elif isinstance(data.frame, BotStartedSpeakingFrame) and self._user_stopped_time:
# Calculate and emit latency # Calculate and emit latency
latency = time.time() - self._user_stopped_time latency = time.time() - self._user_stopped_time